isotree 0.2.0 → 0.3.0
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/CHANGELOG.md +16 -1
- data/LICENSE.txt +2 -2
- data/README.md +41 -23
- data/ext/isotree/ext.cpp +144 -31
- data/ext/isotree/extconf.rb +7 -7
- data/lib/isotree/dataset.rb +0 -1
- data/lib/isotree/isolation_forest.rb +114 -30
- data/lib/isotree/version.rb +1 -1
- data/vendor/isotree/LICENSE +1 -1
- data/vendor/isotree/README.md +165 -27
- data/vendor/isotree/include/isotree.hpp +2111 -0
- data/vendor/isotree/include/isotree_oop.hpp +394 -0
- data/vendor/isotree/inst/COPYRIGHTS +62 -0
- data/vendor/isotree/src/RcppExports.cpp +525 -52
- data/vendor/isotree/src/Rwrapper.cpp +1931 -268
- data/vendor/isotree/src/c_interface.cpp +953 -0
- data/vendor/isotree/src/crit.hpp +4232 -0
- data/vendor/isotree/src/dist.hpp +1886 -0
- data/vendor/isotree/src/exp_depth_table.hpp +134 -0
- data/vendor/isotree/src/extended.hpp +1444 -0
- data/vendor/isotree/src/external_facing_generic.hpp +399 -0
- data/vendor/isotree/src/fit_model.hpp +2401 -0
- data/vendor/isotree/src/{dealloc.cpp → headers_joined.hpp} +38 -22
- data/vendor/isotree/src/helpers_iforest.hpp +813 -0
- data/vendor/isotree/src/{impute.cpp → impute.hpp} +353 -122
- data/vendor/isotree/src/indexer.cpp +515 -0
- data/vendor/isotree/src/instantiate_template_headers.cpp +118 -0
- data/vendor/isotree/src/instantiate_template_headers.hpp +240 -0
- data/vendor/isotree/src/isoforest.hpp +1659 -0
- data/vendor/isotree/src/isotree.hpp +1804 -392
- data/vendor/isotree/src/isotree_exportable.hpp +99 -0
- data/vendor/isotree/src/merge_models.cpp +159 -16
- data/vendor/isotree/src/mult.hpp +1321 -0
- data/vendor/isotree/src/oop_interface.cpp +842 -0
- data/vendor/isotree/src/oop_interface.hpp +278 -0
- data/vendor/isotree/src/other_helpers.hpp +219 -0
- data/vendor/isotree/src/predict.hpp +1932 -0
- data/vendor/isotree/src/python_helpers.hpp +134 -0
- data/vendor/isotree/src/ref_indexer.hpp +154 -0
- data/vendor/isotree/src/robinmap/LICENSE +21 -0
- data/vendor/isotree/src/robinmap/README.md +483 -0
- data/vendor/isotree/src/robinmap/include/tsl/robin_growth_policy.h +406 -0
- data/vendor/isotree/src/robinmap/include/tsl/robin_hash.h +1620 -0
- data/vendor/isotree/src/robinmap/include/tsl/robin_map.h +807 -0
- data/vendor/isotree/src/robinmap/include/tsl/robin_set.h +660 -0
- data/vendor/isotree/src/serialize.cpp +4300 -139
- data/vendor/isotree/src/sql.cpp +141 -59
- data/vendor/isotree/src/subset_models.cpp +174 -0
- data/vendor/isotree/src/utils.hpp +3808 -0
- data/vendor/isotree/src/xoshiro.hpp +467 -0
- data/vendor/isotree/src/ziggurat.hpp +405 -0
- metadata +40 -106
- data/vendor/cereal/LICENSE +0 -24
- data/vendor/cereal/README.md +0 -85
- data/vendor/cereal/include/cereal/access.hpp +0 -351
- data/vendor/cereal/include/cereal/archives/adapters.hpp +0 -163
- data/vendor/cereal/include/cereal/archives/binary.hpp +0 -169
- data/vendor/cereal/include/cereal/archives/json.hpp +0 -1019
- data/vendor/cereal/include/cereal/archives/portable_binary.hpp +0 -334
- data/vendor/cereal/include/cereal/archives/xml.hpp +0 -956
- data/vendor/cereal/include/cereal/cereal.hpp +0 -1089
- data/vendor/cereal/include/cereal/details/helpers.hpp +0 -422
- data/vendor/cereal/include/cereal/details/polymorphic_impl.hpp +0 -796
- data/vendor/cereal/include/cereal/details/polymorphic_impl_fwd.hpp +0 -65
- data/vendor/cereal/include/cereal/details/static_object.hpp +0 -127
- data/vendor/cereal/include/cereal/details/traits.hpp +0 -1411
- data/vendor/cereal/include/cereal/details/util.hpp +0 -84
- data/vendor/cereal/include/cereal/external/base64.hpp +0 -134
- data/vendor/cereal/include/cereal/external/rapidjson/allocators.h +0 -284
- data/vendor/cereal/include/cereal/external/rapidjson/cursorstreamwrapper.h +0 -78
- data/vendor/cereal/include/cereal/external/rapidjson/document.h +0 -2652
- data/vendor/cereal/include/cereal/external/rapidjson/encodedstream.h +0 -299
- data/vendor/cereal/include/cereal/external/rapidjson/encodings.h +0 -716
- data/vendor/cereal/include/cereal/external/rapidjson/error/en.h +0 -74
- data/vendor/cereal/include/cereal/external/rapidjson/error/error.h +0 -161
- data/vendor/cereal/include/cereal/external/rapidjson/filereadstream.h +0 -99
- data/vendor/cereal/include/cereal/external/rapidjson/filewritestream.h +0 -104
- data/vendor/cereal/include/cereal/external/rapidjson/fwd.h +0 -151
- data/vendor/cereal/include/cereal/external/rapidjson/internal/biginteger.h +0 -290
- data/vendor/cereal/include/cereal/external/rapidjson/internal/diyfp.h +0 -271
- data/vendor/cereal/include/cereal/external/rapidjson/internal/dtoa.h +0 -245
- data/vendor/cereal/include/cereal/external/rapidjson/internal/ieee754.h +0 -78
- data/vendor/cereal/include/cereal/external/rapidjson/internal/itoa.h +0 -308
- data/vendor/cereal/include/cereal/external/rapidjson/internal/meta.h +0 -186
- data/vendor/cereal/include/cereal/external/rapidjson/internal/pow10.h +0 -55
- data/vendor/cereal/include/cereal/external/rapidjson/internal/regex.h +0 -740
- data/vendor/cereal/include/cereal/external/rapidjson/internal/stack.h +0 -232
- data/vendor/cereal/include/cereal/external/rapidjson/internal/strfunc.h +0 -69
- data/vendor/cereal/include/cereal/external/rapidjson/internal/strtod.h +0 -290
- data/vendor/cereal/include/cereal/external/rapidjson/internal/swap.h +0 -46
- data/vendor/cereal/include/cereal/external/rapidjson/istreamwrapper.h +0 -128
- data/vendor/cereal/include/cereal/external/rapidjson/memorybuffer.h +0 -70
- data/vendor/cereal/include/cereal/external/rapidjson/memorystream.h +0 -71
- data/vendor/cereal/include/cereal/external/rapidjson/msinttypes/inttypes.h +0 -316
- data/vendor/cereal/include/cereal/external/rapidjson/msinttypes/stdint.h +0 -300
- data/vendor/cereal/include/cereal/external/rapidjson/ostreamwrapper.h +0 -81
- data/vendor/cereal/include/cereal/external/rapidjson/pointer.h +0 -1414
- data/vendor/cereal/include/cereal/external/rapidjson/prettywriter.h +0 -277
- data/vendor/cereal/include/cereal/external/rapidjson/rapidjson.h +0 -656
- data/vendor/cereal/include/cereal/external/rapidjson/reader.h +0 -2230
- data/vendor/cereal/include/cereal/external/rapidjson/schema.h +0 -2497
- data/vendor/cereal/include/cereal/external/rapidjson/stream.h +0 -223
- data/vendor/cereal/include/cereal/external/rapidjson/stringbuffer.h +0 -121
- data/vendor/cereal/include/cereal/external/rapidjson/writer.h +0 -709
- data/vendor/cereal/include/cereal/external/rapidxml/license.txt +0 -52
- data/vendor/cereal/include/cereal/external/rapidxml/manual.html +0 -406
- data/vendor/cereal/include/cereal/external/rapidxml/rapidxml.hpp +0 -2624
- data/vendor/cereal/include/cereal/external/rapidxml/rapidxml_iterators.hpp +0 -175
- data/vendor/cereal/include/cereal/external/rapidxml/rapidxml_print.hpp +0 -428
- data/vendor/cereal/include/cereal/external/rapidxml/rapidxml_utils.hpp +0 -123
- data/vendor/cereal/include/cereal/macros.hpp +0 -154
- data/vendor/cereal/include/cereal/specialize.hpp +0 -139
- data/vendor/cereal/include/cereal/types/array.hpp +0 -79
- data/vendor/cereal/include/cereal/types/atomic.hpp +0 -55
- data/vendor/cereal/include/cereal/types/base_class.hpp +0 -203
- data/vendor/cereal/include/cereal/types/bitset.hpp +0 -176
- data/vendor/cereal/include/cereal/types/boost_variant.hpp +0 -164
- data/vendor/cereal/include/cereal/types/chrono.hpp +0 -72
- data/vendor/cereal/include/cereal/types/common.hpp +0 -129
- data/vendor/cereal/include/cereal/types/complex.hpp +0 -56
- data/vendor/cereal/include/cereal/types/concepts/pair_associative_container.hpp +0 -73
- data/vendor/cereal/include/cereal/types/deque.hpp +0 -62
- data/vendor/cereal/include/cereal/types/forward_list.hpp +0 -68
- data/vendor/cereal/include/cereal/types/functional.hpp +0 -43
- data/vendor/cereal/include/cereal/types/list.hpp +0 -62
- data/vendor/cereal/include/cereal/types/map.hpp +0 -36
- data/vendor/cereal/include/cereal/types/memory.hpp +0 -425
- data/vendor/cereal/include/cereal/types/optional.hpp +0 -66
- data/vendor/cereal/include/cereal/types/polymorphic.hpp +0 -483
- data/vendor/cereal/include/cereal/types/queue.hpp +0 -132
- data/vendor/cereal/include/cereal/types/set.hpp +0 -103
- data/vendor/cereal/include/cereal/types/stack.hpp +0 -76
- data/vendor/cereal/include/cereal/types/string.hpp +0 -61
- data/vendor/cereal/include/cereal/types/tuple.hpp +0 -123
- data/vendor/cereal/include/cereal/types/unordered_map.hpp +0 -36
- data/vendor/cereal/include/cereal/types/unordered_set.hpp +0 -99
- data/vendor/cereal/include/cereal/types/utility.hpp +0 -47
- data/vendor/cereal/include/cereal/types/valarray.hpp +0 -89
- data/vendor/cereal/include/cereal/types/variant.hpp +0 -109
- data/vendor/cereal/include/cereal/types/vector.hpp +0 -112
- data/vendor/cereal/include/cereal/version.hpp +0 -52
- data/vendor/isotree/src/Makevars +0 -4
- data/vendor/isotree/src/crit.cpp +0 -912
- data/vendor/isotree/src/dist.cpp +0 -749
- data/vendor/isotree/src/extended.cpp +0 -790
- data/vendor/isotree/src/fit_model.cpp +0 -1090
- data/vendor/isotree/src/helpers_iforest.cpp +0 -324
- data/vendor/isotree/src/isoforest.cpp +0 -771
- data/vendor/isotree/src/mult.cpp +0 -607
- data/vendor/isotree/src/predict.cpp +0 -853
- data/vendor/isotree/src/utils.cpp +0 -1566
@@ -0,0 +1,4232 @@
|
|
1
|
+
/* Isolation forests and variations thereof, with adjustments for incorporation
|
2
|
+
* of categorical variables and missing values.
|
3
|
+
* Writen for C++11 standard and aimed at being used in R and Python.
|
4
|
+
*
|
5
|
+
* This library is based on the following works:
|
6
|
+
* [1] Liu, Fei Tony, Kai Ming Ting, and Zhi-Hua Zhou.
|
7
|
+
* "Isolation forest."
|
8
|
+
* 2008 Eighth IEEE International Conference on Data Mining. IEEE, 2008.
|
9
|
+
* [2] Liu, Fei Tony, Kai Ming Ting, and Zhi-Hua Zhou.
|
10
|
+
* "Isolation-based anomaly detection."
|
11
|
+
* ACM Transactions on Knowledge Discovery from Data (TKDD) 6.1 (2012): 3.
|
12
|
+
* [3] Hariri, Sahand, Matias Carrasco Kind, and Robert J. Brunner.
|
13
|
+
* "Extended Isolation Forest."
|
14
|
+
* arXiv preprint arXiv:1811.02141 (2018).
|
15
|
+
* [4] Liu, Fei Tony, Kai Ming Ting, and Zhi-Hua Zhou.
|
16
|
+
* "On detecting clustered anomalies using SCiForest."
|
17
|
+
* Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, Berlin, Heidelberg, 2010.
|
18
|
+
* [5] https://sourceforge.net/projects/iforest/
|
19
|
+
* [6] https://math.stackexchange.com/questions/3388518/expected-number-of-paths-required-to-separate-elements-in-a-binary-tree
|
20
|
+
* [7] Quinlan, J. Ross. C4. 5: programs for machine learning. Elsevier, 2014.
|
21
|
+
* [8] Cortes, David.
|
22
|
+
* "Distance approximation using Isolation Forests."
|
23
|
+
* arXiv preprint arXiv:1910.12362 (2019).
|
24
|
+
* [9] Cortes, David.
|
25
|
+
* "Imputing missing values with unsupervised random trees."
|
26
|
+
* arXiv preprint arXiv:1911.06646 (2019).
|
27
|
+
* [10] https://math.stackexchange.com/questions/3333220/expected-average-depth-in-random-binary-tree-constructed-top-to-bottom
|
28
|
+
* [11] Cortes, David.
|
29
|
+
* "Revisiting randomized choices in isolation forests."
|
30
|
+
* arXiv preprint arXiv:2110.13402 (2021).
|
31
|
+
* [12] Guha, Sudipto, et al.
|
32
|
+
* "Robust random cut forest based anomaly detection on streams."
|
33
|
+
* International conference on machine learning. PMLR, 2016.
|
34
|
+
* [13] Cortes, David.
|
35
|
+
* "Isolation forests: looking beyond tree depth."
|
36
|
+
* arXiv preprint arXiv:2111.11639 (2021).
|
37
|
+
* [14] Ting, Kai Ming, Yue Zhu, and Zhi-Hua Zhou.
|
38
|
+
* "Isolation kernel and its effect on SVM"
|
39
|
+
* Proceedings of the 24th ACM SIGKDD
|
40
|
+
* International Conference on Knowledge Discovery & Data Mining. 2018.
|
41
|
+
*
|
42
|
+
* BSD 2-Clause License
|
43
|
+
* Copyright (c) 2019-2022, David Cortes
|
44
|
+
* All rights reserved.
|
45
|
+
* Redistribution and use in source and binary forms, with or without
|
46
|
+
* modification, are permitted provided that the following conditions are met:
|
47
|
+
* * Redistributions of source code must retain the above copyright notice, this
|
48
|
+
* list of conditions and the following disclaimer.
|
49
|
+
* * Redistributions in binary form must reproduce the above copyright notice,
|
50
|
+
* this list of conditions and the following disclaimer in the documentation
|
51
|
+
* and/or other materials provided with the distribution.
|
52
|
+
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
53
|
+
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
54
|
+
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
55
|
+
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
56
|
+
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
57
|
+
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
58
|
+
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
59
|
+
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
60
|
+
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
61
|
+
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
62
|
+
*/
|
63
|
+
#include "isotree.hpp"
|
64
|
+
|
65
|
+
/* TODO: should the kurtosis calculation impute values when using ndim=1 + missing_action=Impute?
|
66
|
+
It should be the theoretically correct approach, but will cause the kurtosis to increase
|
67
|
+
significantly if there is a large number of missing values, which would lead to prefer
|
68
|
+
splitting on columns with mostly missing values. */
|
69
|
+
|
70
|
+
/* TODO: this kurtosis caps the minimum values to zero, but the minimum achievable value is 1,
|
71
|
+
see how are imprecise results used outside of the function in the different kind of calculations
|
72
|
+
that use kurtosis and maybe change the logic. */
|
73
|
+
|
74
|
+
#define pw1(x) ((x))
|
75
|
+
#define pw2(x) ((x) * (x))
|
76
|
+
#define pw3(x) ((x) * (x) * (x))
|
77
|
+
#define pw4(x) ((x) * (x) * (x) * (x))
|
78
|
+
|
79
|
+
/* https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Higher-order_statistics */
|
80
|
+
template <class real_t, class ldouble_safe>
|
81
|
+
double calc_kurtosis(size_t ix_arr[], size_t st, size_t end, real_t x[], MissingAction missing_action)
|
82
|
+
{
|
83
|
+
ldouble_safe m = 0;
|
84
|
+
ldouble_safe M2 = 0, M3 = 0, M4 = 0;
|
85
|
+
ldouble_safe delta, delta_s, delta_div;
|
86
|
+
ldouble_safe diff, n;
|
87
|
+
ldouble_safe out;
|
88
|
+
|
89
|
+
if (missing_action == Fail)
|
90
|
+
{
|
91
|
+
for (size_t row = st; row <= end; row++)
|
92
|
+
{
|
93
|
+
n = (ldouble_safe)(row - st + 1);
|
94
|
+
|
95
|
+
delta = x[ix_arr[row]] - m;
|
96
|
+
delta_div = delta / n;
|
97
|
+
delta_s = delta_div * delta_div;
|
98
|
+
diff = delta * (delta_div * (ldouble_safe)(row - st));
|
99
|
+
|
100
|
+
m += delta_div;
|
101
|
+
M4 += diff * delta_s * (n * n - 3 * n + 3) + 6 * delta_s * M2 - 4 * delta_div * M3;
|
102
|
+
M3 += diff * delta_div * (n - 2) - 3 * delta_div * M2;
|
103
|
+
M2 += diff;
|
104
|
+
}
|
105
|
+
|
106
|
+
if (unlikely(!is_na_or_inf(M2) && M2 <= 0))
|
107
|
+
{
|
108
|
+
if (!check_more_than_two_unique_values(ix_arr, st, end, x, missing_action))
|
109
|
+
return -HUGE_VAL;
|
110
|
+
}
|
111
|
+
|
112
|
+
out = ( M4 / M2 ) * ( (ldouble_safe)(end - st + 1) / M2 );
|
113
|
+
return (!is_na_or_inf(out))? std::fmax((double)out, 0.) : (-HUGE_VAL);
|
114
|
+
}
|
115
|
+
|
116
|
+
else
|
117
|
+
{
|
118
|
+
size_t cnt = 0;
|
119
|
+
for (size_t row = st; row <= end; row++)
|
120
|
+
{
|
121
|
+
if (likely(!is_na_or_inf(x[ix_arr[row]])))
|
122
|
+
{
|
123
|
+
cnt++;
|
124
|
+
n = (ldouble_safe) cnt;
|
125
|
+
|
126
|
+
delta = x[ix_arr[row]] - m;
|
127
|
+
delta_div = delta / n;
|
128
|
+
delta_s = delta_div * delta_div;
|
129
|
+
diff = delta * (delta_div * (ldouble_safe)(cnt - 1));
|
130
|
+
|
131
|
+
m += delta_div;
|
132
|
+
M4 += diff * delta_s * (n * n - 3 * n + 3) + 6 * delta_s * M2 - 4 * delta_div * M3;
|
133
|
+
M3 += diff * delta_div * (n - 2) - 3 * delta_div * M2;
|
134
|
+
M2 += diff;
|
135
|
+
}
|
136
|
+
}
|
137
|
+
|
138
|
+
if (unlikely(cnt == 0)) return -HUGE_VAL;
|
139
|
+
if (unlikely(!is_na_or_inf(M2) && M2 <= 0))
|
140
|
+
{
|
141
|
+
if (!check_more_than_two_unique_values(ix_arr, st, end, x, missing_action))
|
142
|
+
return -HUGE_VAL;
|
143
|
+
}
|
144
|
+
|
145
|
+
out = ( M4 / M2 ) * ( (ldouble_safe)cnt / M2 );
|
146
|
+
return (!is_na_or_inf(out))? std::fmax((double)out, 0.) : (-HUGE_VAL);
|
147
|
+
}
|
148
|
+
}
|
149
|
+
|
150
|
+
template <class real_t, class ldouble_safe>
|
151
|
+
double calc_kurtosis(real_t x[], size_t n, MissingAction missing_action)
|
152
|
+
{
|
153
|
+
ldouble_safe m = 0;
|
154
|
+
ldouble_safe M2 = 0, M3 = 0, M4 = 0;
|
155
|
+
ldouble_safe delta, delta_s, delta_div;
|
156
|
+
ldouble_safe diff, n_;
|
157
|
+
ldouble_safe out;
|
158
|
+
|
159
|
+
if (missing_action == Fail)
|
160
|
+
{
|
161
|
+
for (size_t row = 0; row < n; row++)
|
162
|
+
{
|
163
|
+
n_ = (ldouble_safe)(row + 1);
|
164
|
+
|
165
|
+
delta = x[row] - m;
|
166
|
+
delta_div = delta / n_;
|
167
|
+
delta_s = delta_div * delta_div;
|
168
|
+
diff = delta * (delta_div * (ldouble_safe)row);
|
169
|
+
|
170
|
+
m += delta_div;
|
171
|
+
M4 += diff * delta_s * (n_ * n_ - 3 * n_ + 3) + 6 * delta_s * M2 - 4 * delta_div * M3;
|
172
|
+
M3 += diff * delta_div * (n_ - 2) - 3 * delta_div * M2;
|
173
|
+
M2 += diff;
|
174
|
+
}
|
175
|
+
|
176
|
+
out = ( M4 / M2 ) * ( (ldouble_safe)n / M2 );
|
177
|
+
return (!is_na_or_inf(out))? std::fmax((double)out, 0.) : (-HUGE_VAL);
|
178
|
+
}
|
179
|
+
|
180
|
+
else
|
181
|
+
{
|
182
|
+
size_t cnt = 0;
|
183
|
+
for (size_t row = 0; row < n; row++)
|
184
|
+
{
|
185
|
+
if (likely(!is_na_or_inf(x[row])))
|
186
|
+
{
|
187
|
+
cnt++;
|
188
|
+
n_ = (ldouble_safe) cnt;
|
189
|
+
|
190
|
+
delta = x[row] - m;
|
191
|
+
delta_div = delta / n_;
|
192
|
+
delta_s = delta_div * delta_div;
|
193
|
+
diff = delta * (delta_div * (ldouble_safe)(cnt - 1));
|
194
|
+
|
195
|
+
m += delta_div;
|
196
|
+
M4 += diff * delta_s * (n_ * n_ - 3 * n_ + 3) + 6 * delta_s * M2 - 4 * delta_div * M3;
|
197
|
+
M3 += diff * delta_div * (n_ - 2) - 3 * delta_div * M2;
|
198
|
+
M2 += diff;
|
199
|
+
}
|
200
|
+
}
|
201
|
+
|
202
|
+
if (unlikely(cnt == 0)) return -HUGE_VAL;
|
203
|
+
|
204
|
+
out = ( M4 / M2 ) * ( (ldouble_safe)cnt / M2 );
|
205
|
+
return (!is_na_or_inf(out))? std::fmax((double)out, 0.) : (-HUGE_VAL);
|
206
|
+
}
|
207
|
+
}
|
208
|
+
|
209
|
+
/* TODO: is this algorithm correct? */
|
210
|
+
template <class real_t, class mapping, class ldouble_safe>
|
211
|
+
double calc_kurtosis_weighted(size_t ix_arr[], size_t st, size_t end, real_t x[],
|
212
|
+
MissingAction missing_action, mapping &restrict w)
|
213
|
+
{
|
214
|
+
ldouble_safe m = 0;
|
215
|
+
ldouble_safe M2 = 0, M3 = 0, M4 = 0;
|
216
|
+
ldouble_safe delta, delta_s, delta_div;
|
217
|
+
ldouble_safe diff;
|
218
|
+
ldouble_safe n = 0;
|
219
|
+
ldouble_safe out;
|
220
|
+
ldouble_safe n_prev = 0.;
|
221
|
+
ldouble_safe w_this;
|
222
|
+
|
223
|
+
for (size_t row = st; row <= end; row++)
|
224
|
+
{
|
225
|
+
if (likely(!is_na_or_inf(x[ix_arr[row]])))
|
226
|
+
{
|
227
|
+
w_this = w[ix_arr[row]];
|
228
|
+
n += w_this;
|
229
|
+
|
230
|
+
delta = x[ix_arr[row]] - m;
|
231
|
+
delta_div = delta / n;
|
232
|
+
delta_s = delta_div * delta_div;
|
233
|
+
diff = delta * (delta_div * n_prev);
|
234
|
+
n_prev = n;
|
235
|
+
|
236
|
+
m += w_this * (delta_div);
|
237
|
+
M4 += w_this * (diff * delta_s * (n * n - 3 * n + 3) + 6 * delta_s * M2 - 4 * delta_div * M3);
|
238
|
+
M3 += w_this * (diff * delta_div * (n - 2) - 3 * delta_div * M2);
|
239
|
+
M2 += w_this * (diff);
|
240
|
+
}
|
241
|
+
}
|
242
|
+
|
243
|
+
if (unlikely(n <= 0)) return -HUGE_VAL;
|
244
|
+
if (unlikely(!is_na_or_inf(M2) && M2 <= std::numeric_limits<double>::epsilon()))
|
245
|
+
{
|
246
|
+
if (!check_more_than_two_unique_values(ix_arr, st, end, x, missing_action))
|
247
|
+
return -HUGE_VAL;
|
248
|
+
}
|
249
|
+
|
250
|
+
out = ( M4 / M2 ) * ( n / M2 );
|
251
|
+
return (!is_na_or_inf(out))? std::fmax((double)out, 0.) : (-HUGE_VAL);
|
252
|
+
}
|
253
|
+
|
254
|
+
template <class real_t, class ldouble_safe>
|
255
|
+
double calc_kurtosis_weighted(real_t *restrict x, size_t n_, MissingAction missing_action, real_t *restrict w)
|
256
|
+
{
|
257
|
+
ldouble_safe m = 0;
|
258
|
+
ldouble_safe M2 = 0, M3 = 0, M4 = 0;
|
259
|
+
ldouble_safe delta, delta_s, delta_div;
|
260
|
+
ldouble_safe diff;
|
261
|
+
ldouble_safe n = 0;
|
262
|
+
ldouble_safe out;
|
263
|
+
ldouble_safe n_prev = 0.;
|
264
|
+
ldouble_safe w_this;
|
265
|
+
|
266
|
+
for (size_t row = 0; row < n_; row++)
|
267
|
+
{
|
268
|
+
if (likely(!is_na_or_inf(x[row])))
|
269
|
+
{
|
270
|
+
w_this = w[row];
|
271
|
+
n += w_this;
|
272
|
+
|
273
|
+
delta = x[row] - m;
|
274
|
+
delta_div = delta / n;
|
275
|
+
delta_s = delta_div * delta_div;
|
276
|
+
diff = delta * (delta_div * n_prev);
|
277
|
+
n_prev = n;
|
278
|
+
|
279
|
+
m += w_this * (delta_div);
|
280
|
+
M4 += w_this * (diff * delta_s * (n * n - 3 * n + 3) + 6 * delta_s * M2 - 4 * delta_div * M3);
|
281
|
+
M3 += w_this * (diff * delta_div * (n - 2) - 3 * delta_div * M2);
|
282
|
+
M2 += w_this * (diff);
|
283
|
+
}
|
284
|
+
}
|
285
|
+
|
286
|
+
if (unlikely(n <= 0)) return -HUGE_VAL;
|
287
|
+
|
288
|
+
out = ( M4 / M2 ) * ( n / M2 );
|
289
|
+
return (!is_na_or_inf(out))? std::fmax((double)out, 0.) : (-HUGE_VAL);
|
290
|
+
}
|
291
|
+
|
292
|
+
|
293
|
+
/* TODO: make these compensated sums */
|
294
|
+
/* TODO: can this use the same algorithm as above but with a correction at the end,
|
295
|
+
like it was done for the variance? */
|
296
|
+
template <class real_t, class sparse_ix, class ldouble_safe>
|
297
|
+
double calc_kurtosis(size_t *restrict ix_arr, size_t st, size_t end, size_t col_num,
|
298
|
+
real_t Xc[], sparse_ix *restrict Xc_ind, sparse_ix *restrict Xc_indptr,
|
299
|
+
MissingAction missing_action)
|
300
|
+
{
|
301
|
+
/* ix_arr must be already sorted beforehand */
|
302
|
+
if (Xc_indptr[col_num] == Xc_indptr[col_num + 1])
|
303
|
+
return -HUGE_VAL;
|
304
|
+
|
305
|
+
ldouble_safe s1 = 0;
|
306
|
+
ldouble_safe s2 = 0;
|
307
|
+
ldouble_safe s3 = 0;
|
308
|
+
ldouble_safe s4 = 0;
|
309
|
+
ldouble_safe x_sq;
|
310
|
+
size_t cnt = end - st + 1;
|
311
|
+
|
312
|
+
if (unlikely(cnt <= 1)) return -HUGE_VAL;
|
313
|
+
|
314
|
+
size_t st_col = Xc_indptr[col_num];
|
315
|
+
size_t end_col = Xc_indptr[col_num + 1] - 1;
|
316
|
+
size_t curr_pos = st_col;
|
317
|
+
size_t ind_end_col = Xc_ind[end_col];
|
318
|
+
size_t *ptr_st = std::lower_bound(ix_arr + st, ix_arr + end + 1, Xc_ind[st_col]);
|
319
|
+
|
320
|
+
ldouble_safe xval;
|
321
|
+
|
322
|
+
if (missing_action != Fail)
|
323
|
+
{
|
324
|
+
for (size_t *row = ptr_st;
|
325
|
+
row != ix_arr + end + 1 && curr_pos != end_col + 1 && ind_end_col >= *row;
|
326
|
+
)
|
327
|
+
{
|
328
|
+
if (Xc_ind[curr_pos] == (sparse_ix)(*row))
|
329
|
+
{
|
330
|
+
xval = Xc[curr_pos];
|
331
|
+
if (unlikely(is_na_or_inf(xval)))
|
332
|
+
{
|
333
|
+
cnt--;
|
334
|
+
}
|
335
|
+
|
336
|
+
else
|
337
|
+
{
|
338
|
+
/* TODO: is it safe to use FMA here? some calculations rely on assuming that
|
339
|
+
some of these 's' are larger than the others. Would this procedure be guaranteed
|
340
|
+
to preserve such differences if done with a mixture of sums and FMAs? */
|
341
|
+
x_sq = square(xval);
|
342
|
+
s1 += xval;
|
343
|
+
s2 = std::fma(xval, xval, s2);
|
344
|
+
s3 = std::fma(x_sq, xval, s3);
|
345
|
+
s4 = std::fma(x_sq, x_sq, s4);
|
346
|
+
// s1 += pw1(xval);
|
347
|
+
// s2 += pw2(xval);
|
348
|
+
// s3 += pw3(xval);
|
349
|
+
// s4 += pw4(xval);
|
350
|
+
}
|
351
|
+
|
352
|
+
if (row == ix_arr + end || curr_pos == end_col) break;
|
353
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
354
|
+
}
|
355
|
+
|
356
|
+
else
|
357
|
+
{
|
358
|
+
if (Xc_ind[curr_pos] > (sparse_ix)(*row))
|
359
|
+
row = std::lower_bound(row + 1, ix_arr + end + 1, Xc_ind[curr_pos]);
|
360
|
+
else
|
361
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *row) - Xc_ind;
|
362
|
+
}
|
363
|
+
}
|
364
|
+
|
365
|
+
if (unlikely(cnt <= (end - st + 1) - (Xc_indptr[col_num+1] - Xc_indptr[col_num]))) return -HUGE_VAL;
|
366
|
+
}
|
367
|
+
|
368
|
+
else
|
369
|
+
{
|
370
|
+
for (size_t *row = ptr_st;
|
371
|
+
row != ix_arr + end + 1 && curr_pos != end_col + 1 && ind_end_col >= *row;
|
372
|
+
)
|
373
|
+
{
|
374
|
+
if (Xc_ind[curr_pos] == (sparse_ix)(*row))
|
375
|
+
{
|
376
|
+
xval = Xc[curr_pos];
|
377
|
+
x_sq = square(xval);
|
378
|
+
s1 += xval;
|
379
|
+
s2 = std::fma(xval, xval, s2);
|
380
|
+
s3 = std::fma(x_sq, xval, s3);
|
381
|
+
s4 = std::fma(x_sq, x_sq, s4);
|
382
|
+
// s1 += pw1(xval);
|
383
|
+
// s2 += pw2(xval);
|
384
|
+
// s3 += pw3(xval);
|
385
|
+
// s4 += pw4(xval);
|
386
|
+
|
387
|
+
if (row == ix_arr + end || curr_pos == end_col) break;
|
388
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
389
|
+
}
|
390
|
+
|
391
|
+
else
|
392
|
+
{
|
393
|
+
if (Xc_ind[curr_pos] > (sparse_ix)(*row))
|
394
|
+
row = std::lower_bound(row + 1, ix_arr + end + 1, Xc_ind[curr_pos]);
|
395
|
+
else
|
396
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *row) - Xc_ind;
|
397
|
+
}
|
398
|
+
}
|
399
|
+
}
|
400
|
+
|
401
|
+
if (unlikely(cnt <= 1 || s2 == 0 || s2 == pw2(s1))) return -HUGE_VAL;
|
402
|
+
ldouble_safe cnt_l = (ldouble_safe) cnt;
|
403
|
+
ldouble_safe sn = s1 / cnt_l;
|
404
|
+
ldouble_safe v = s2 / cnt_l - pw2(sn);
|
405
|
+
if (unlikely(std::isnan(v))) return -HUGE_VAL;
|
406
|
+
if (
|
407
|
+
v <= std::numeric_limits<double>::epsilon() &&
|
408
|
+
!check_more_than_two_unique_values(ix_arr, st, end, col_num,
|
409
|
+
Xc_indptr, Xc_ind, Xc,
|
410
|
+
missing_action)
|
411
|
+
)
|
412
|
+
return -HUGE_VAL;
|
413
|
+
if (unlikely(v <= 0)) return 0.;
|
414
|
+
ldouble_safe out = (s4 - 4 * s3 * sn + 6 * s2 * pw2(sn) - 4 * s1 * pw3(sn) + cnt_l * pw4(sn)) / (cnt_l * pw2(v));
|
415
|
+
return (!is_na_or_inf(out))? std::fmax((double)out, 0.) : (-HUGE_VAL);
|
416
|
+
}
|
417
|
+
|
418
|
+
template <class real_t, class sparse_ix, class ldouble_safe>
|
419
|
+
double calc_kurtosis(size_t col_num, size_t nrows,
|
420
|
+
real_t Xc[], sparse_ix *restrict Xc_ind, sparse_ix *restrict Xc_indptr,
|
421
|
+
MissingAction missing_action)
|
422
|
+
{
|
423
|
+
if (Xc_indptr[col_num] == Xc_indptr[col_num + 1])
|
424
|
+
return -HUGE_VAL;
|
425
|
+
|
426
|
+
ldouble_safe s1 = 0;
|
427
|
+
ldouble_safe s2 = 0;
|
428
|
+
ldouble_safe s3 = 0;
|
429
|
+
ldouble_safe s4 = 0;
|
430
|
+
ldouble_safe x_sq;
|
431
|
+
size_t cnt = nrows;
|
432
|
+
|
433
|
+
if (unlikely(cnt <= 1)) return -HUGE_VAL;
|
434
|
+
|
435
|
+
ldouble_safe xval;
|
436
|
+
|
437
|
+
if (missing_action != Fail)
|
438
|
+
{
|
439
|
+
for (auto ix = Xc_indptr[col_num]; ix < Xc_indptr[col_num+1]; ix++)
|
440
|
+
{
|
441
|
+
xval = Xc[ix];
|
442
|
+
if (unlikely(is_na_or_inf(xval)))
|
443
|
+
{
|
444
|
+
cnt--;
|
445
|
+
}
|
446
|
+
|
447
|
+
else
|
448
|
+
{
|
449
|
+
x_sq = square(xval);
|
450
|
+
s1 += xval;
|
451
|
+
s2 = std::fma(xval, xval, s2);
|
452
|
+
s3 = std::fma(x_sq, xval, s3);
|
453
|
+
s4 = std::fma(x_sq, x_sq, s4);
|
454
|
+
// s1 += pw1(xval);
|
455
|
+
// s2 += pw2(xval);
|
456
|
+
// s3 += pw3(xval);
|
457
|
+
// s4 += pw4(xval);
|
458
|
+
}
|
459
|
+
}
|
460
|
+
|
461
|
+
if (cnt <= (nrows) - (Xc_indptr[col_num+1] - Xc_indptr[col_num])) return -HUGE_VAL;
|
462
|
+
}
|
463
|
+
|
464
|
+
else
|
465
|
+
{
|
466
|
+
for (auto ix = Xc_indptr[col_num]; ix < Xc_indptr[col_num+1]; ix++)
|
467
|
+
{
|
468
|
+
xval = Xc[ix];
|
469
|
+
x_sq = square(xval);
|
470
|
+
s1 += xval;
|
471
|
+
s2 = std::fma(xval, xval, s2);
|
472
|
+
s3 = std::fma(x_sq, xval, s3);
|
473
|
+
s4 = std::fma(x_sq, x_sq, s4);
|
474
|
+
// s1 += pw1(xval);
|
475
|
+
// s2 += pw2(xval);
|
476
|
+
// s3 += pw3(xval);
|
477
|
+
// s4 += pw4(xval);
|
478
|
+
}
|
479
|
+
}
|
480
|
+
|
481
|
+
if (unlikely(cnt <= 1 || s2 == 0 || s2 == pw2(s1))) return -HUGE_VAL;
|
482
|
+
ldouble_safe cnt_l = (ldouble_safe) cnt;
|
483
|
+
ldouble_safe sn = s1 / cnt_l;
|
484
|
+
ldouble_safe v = s2 / cnt_l - pw2(sn);
|
485
|
+
if (unlikely(std::isnan(v))) return -HUGE_VAL;
|
486
|
+
if (
|
487
|
+
v <= std::numeric_limits<double>::epsilon() &&
|
488
|
+
!check_more_than_two_unique_values(nrows, col_num,
|
489
|
+
Xc_indptr, Xc_ind, Xc,
|
490
|
+
missing_action)
|
491
|
+
)
|
492
|
+
return -HUGE_VAL;
|
493
|
+
if (unlikely(v <= 0)) return 0.;
|
494
|
+
ldouble_safe out = (s4 - 4 * s3 * sn + 6 * s2 * pw2(sn) - 4 * s1 * pw3(sn) + cnt_l * pw4(sn)) / (cnt_l * pw2(v));
|
495
|
+
return (!is_na_or_inf(out))? std::fmax((double)out, 0.) : (-HUGE_VAL);
|
496
|
+
}
|
497
|
+
|
498
|
+
|
499
|
+
template <class real_t, class sparse_ix, class mapping, class ldouble_safe>
|
500
|
+
double calc_kurtosis_weighted(size_t *restrict ix_arr, size_t st, size_t end, size_t col_num,
|
501
|
+
real_t Xc[], sparse_ix *restrict Xc_ind, sparse_ix *restrict Xc_indptr,
|
502
|
+
MissingAction missing_action, mapping &restrict w)
|
503
|
+
{
|
504
|
+
/* ix_arr must be already sorted beforehand */
|
505
|
+
if (Xc_indptr[col_num] == Xc_indptr[col_num + 1])
|
506
|
+
return -HUGE_VAL;
|
507
|
+
|
508
|
+
ldouble_safe s1 = 0;
|
509
|
+
ldouble_safe s2 = 0;
|
510
|
+
ldouble_safe s3 = 0;
|
511
|
+
ldouble_safe s4 = 0;
|
512
|
+
ldouble_safe x_sq;
|
513
|
+
ldouble_safe w_this;
|
514
|
+
ldouble_safe cnt = 0;
|
515
|
+
for (size_t row = st; row <= end; row++)
|
516
|
+
cnt += w[ix_arr[row]];
|
517
|
+
|
518
|
+
if (unlikely(cnt <= 0)) return -HUGE_VAL;
|
519
|
+
|
520
|
+
size_t st_col = Xc_indptr[col_num];
|
521
|
+
size_t end_col = Xc_indptr[col_num + 1] - 1;
|
522
|
+
size_t curr_pos = st_col;
|
523
|
+
size_t ind_end_col = Xc_ind[end_col];
|
524
|
+
size_t *ptr_st = std::lower_bound(ix_arr + st, ix_arr + end + 1, Xc_ind[st_col]);
|
525
|
+
|
526
|
+
ldouble_safe xval;
|
527
|
+
|
528
|
+
if (missing_action != Fail)
|
529
|
+
{
|
530
|
+
for (size_t *row = ptr_st;
|
531
|
+
row != ix_arr + end + 1 && curr_pos != end_col + 1 && ind_end_col >= *row;
|
532
|
+
)
|
533
|
+
{
|
534
|
+
if (Xc_ind[curr_pos] == (sparse_ix)(*row))
|
535
|
+
{
|
536
|
+
w_this = w[*row];
|
537
|
+
xval = Xc[curr_pos];
|
538
|
+
|
539
|
+
if (unlikely(is_na_or_inf(xval)))
|
540
|
+
{
|
541
|
+
cnt -= w_this;
|
542
|
+
}
|
543
|
+
|
544
|
+
else
|
545
|
+
{
|
546
|
+
x_sq = xval * xval;
|
547
|
+
s1 = std::fma(w_this, xval, s1);
|
548
|
+
s2 = std::fma(w_this, x_sq, s2);
|
549
|
+
s3 = std::fma(w_this, x_sq*xval, s3);
|
550
|
+
s4 = std::fma(w_this, x_sq*x_sq, s4);
|
551
|
+
// s1 += w_this * pw1(xval);
|
552
|
+
// s2 += w_this * pw2(xval);
|
553
|
+
// s3 += w_this * pw3(xval);
|
554
|
+
// s4 += w_this * pw4(xval);
|
555
|
+
}
|
556
|
+
|
557
|
+
if (row == ix_arr + end || curr_pos == end_col) break;
|
558
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
559
|
+
}
|
560
|
+
|
561
|
+
else
|
562
|
+
{
|
563
|
+
if (Xc_ind[curr_pos] > (sparse_ix)(*row))
|
564
|
+
row = std::lower_bound(row + 1, ix_arr + end + 1, Xc_ind[curr_pos]);
|
565
|
+
else
|
566
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *row) - Xc_ind;
|
567
|
+
}
|
568
|
+
}
|
569
|
+
|
570
|
+
if (unlikely(cnt <= 0)) return -HUGE_VAL;
|
571
|
+
}
|
572
|
+
|
573
|
+
else
|
574
|
+
{
|
575
|
+
for (size_t *row = ptr_st;
|
576
|
+
row != ix_arr + end + 1 && curr_pos != end_col + 1 && ind_end_col >= *row;
|
577
|
+
)
|
578
|
+
{
|
579
|
+
if (Xc_ind[curr_pos] == (sparse_ix)(*row))
|
580
|
+
{
|
581
|
+
w_this = w[*row];
|
582
|
+
xval = Xc[curr_pos];
|
583
|
+
|
584
|
+
x_sq = xval * xval;
|
585
|
+
s1 = std::fma(w_this, xval, s1);
|
586
|
+
s2 = std::fma(w_this, x_sq, s2);
|
587
|
+
s3 = std::fma(w_this, x_sq*xval, s3);
|
588
|
+
s4 = std::fma(w_this, x_sq*x_sq, s4);
|
589
|
+
// s1 += w_this * pw1(xval);
|
590
|
+
// s2 += w_this * pw2(xval);
|
591
|
+
// s3 += w_this * pw3(xval);
|
592
|
+
// s4 += w_this * pw4(xval);
|
593
|
+
|
594
|
+
if (row == ix_arr + end || curr_pos == end_col) break;
|
595
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
596
|
+
}
|
597
|
+
|
598
|
+
else
|
599
|
+
{
|
600
|
+
if (Xc_ind[curr_pos] > (sparse_ix)(*row))
|
601
|
+
row = std::lower_bound(row + 1, ix_arr + end + 1, Xc_ind[curr_pos]);
|
602
|
+
else
|
603
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *row) - Xc_ind;
|
604
|
+
}
|
605
|
+
}
|
606
|
+
}
|
607
|
+
|
608
|
+
if (unlikely(cnt <= 1 || s2 == 0 || s2 == pw2(s1))) return -HUGE_VAL;
|
609
|
+
ldouble_safe sn = s1 / cnt;
|
610
|
+
ldouble_safe v = s2 / cnt - pw2(sn);
|
611
|
+
if (unlikely(std::isnan(v))) return -HUGE_VAL;
|
612
|
+
if (
|
613
|
+
v <= std::numeric_limits<double>::epsilon() &&
|
614
|
+
!check_more_than_two_unique_values(ix_arr, st, end, col_num,
|
615
|
+
Xc_indptr, Xc_ind, Xc,
|
616
|
+
missing_action)
|
617
|
+
)
|
618
|
+
return -HUGE_VAL;
|
619
|
+
if (v <= 0) return 0.;
|
620
|
+
ldouble_safe out = (s4 - 4 * s3 * sn + 6 * s2 * pw2(sn) - 4 * s1 * pw3(sn) + cnt * pw4(sn)) / (cnt * pw2(v));
|
621
|
+
return (!is_na_or_inf(out))? std::fmax((double)out, 0.) : (-HUGE_VAL);
|
622
|
+
}
|
623
|
+
|
624
|
+
template <class real_t, class sparse_ix, class ldouble_safe>
|
625
|
+
double calc_kurtosis_weighted(size_t col_num, size_t nrows,
|
626
|
+
real_t *restrict Xc, sparse_ix *restrict Xc_ind, sparse_ix *restrict Xc_indptr,
|
627
|
+
MissingAction missing_action, real_t *restrict w)
|
628
|
+
{
|
629
|
+
if (Xc_indptr[col_num] == Xc_indptr[col_num + 1])
|
630
|
+
return -HUGE_VAL;
|
631
|
+
|
632
|
+
ldouble_safe s1 = 0;
|
633
|
+
ldouble_safe s2 = 0;
|
634
|
+
ldouble_safe s3 = 0;
|
635
|
+
ldouble_safe s4 = 0;
|
636
|
+
ldouble_safe x_sq;
|
637
|
+
ldouble_safe w_this;
|
638
|
+
ldouble_safe cnt = nrows - (Xc_indptr[col_num + 1] - Xc_indptr[col_num]);
|
639
|
+
for (auto ix = Xc_indptr[col_num]; ix < Xc_indptr[col_num + 1]; ix++)
|
640
|
+
cnt += w[Xc_ind[ix]];
|
641
|
+
|
642
|
+
if (unlikely(cnt <= 0)) return -HUGE_VAL;
|
643
|
+
|
644
|
+
ldouble_safe xval;
|
645
|
+
|
646
|
+
if (missing_action != Fail)
|
647
|
+
{
|
648
|
+
for (auto ix = Xc_indptr[col_num]; ix < Xc_indptr[col_num + 1]; ix++)
|
649
|
+
{
|
650
|
+
w_this = w[Xc_ind[ix]];
|
651
|
+
xval = Xc[ix];
|
652
|
+
|
653
|
+
if (unlikely(is_na_or_inf(xval)))
|
654
|
+
{
|
655
|
+
cnt -= w_this;
|
656
|
+
}
|
657
|
+
|
658
|
+
else
|
659
|
+
{
|
660
|
+
x_sq = xval * xval;
|
661
|
+
s1 = std::fma(w_this, xval, s1);
|
662
|
+
s2 = std::fma(w_this, x_sq, s2);
|
663
|
+
s3 = std::fma(w_this, x_sq*xval, s3);
|
664
|
+
s4 = std::fma(w_this, x_sq*x_sq, s4);
|
665
|
+
// s1 += w_this * pw1(xval);
|
666
|
+
// s2 += w_this * pw2(xval);
|
667
|
+
// s3 += w_this * pw3(xval);
|
668
|
+
// s4 += w_this * pw4(xval);
|
669
|
+
}
|
670
|
+
}
|
671
|
+
|
672
|
+
if (cnt <= 0) return -HUGE_VAL;
|
673
|
+
}
|
674
|
+
|
675
|
+
else
|
676
|
+
{
|
677
|
+
for (auto ix = Xc_indptr[col_num]; ix < Xc_indptr[col_num + 1]; ix++)
|
678
|
+
{
|
679
|
+
w_this = w[Xc_ind[ix]];
|
680
|
+
xval = Xc[ix];
|
681
|
+
|
682
|
+
x_sq = xval * xval;
|
683
|
+
s1 = std::fma(w_this, xval, s1);
|
684
|
+
s2 = std::fma(w_this, x_sq, s2);
|
685
|
+
s3 = std::fma(w_this, x_sq*xval, s3);
|
686
|
+
s4 = std::fma(w_this, x_sq*x_sq, s4);
|
687
|
+
// s1 += w_this * pw1(xval);
|
688
|
+
// s2 += w_this * pw2(xval);
|
689
|
+
// s3 += w_this * pw3(xval);
|
690
|
+
// s4 += w_this * pw4(xval);
|
691
|
+
}
|
692
|
+
}
|
693
|
+
|
694
|
+
if (unlikely(cnt <= 1 || s2 == 0 || s2 == pw2(s1))) return -HUGE_VAL;
|
695
|
+
ldouble_safe sn = s1 / cnt;
|
696
|
+
ldouble_safe v = s2 / cnt - pw2(sn);
|
697
|
+
if (unlikely(std::isnan(v))) return -HUGE_VAL;
|
698
|
+
if (
|
699
|
+
v <= std::numeric_limits<double>::epsilon() &&
|
700
|
+
!check_more_than_two_unique_values(nrows, col_num,
|
701
|
+
Xc_indptr, Xc_ind, Xc,
|
702
|
+
missing_action)
|
703
|
+
)
|
704
|
+
return -HUGE_VAL;
|
705
|
+
if (unlikely(v <= 0)) return -HUGE_VAL;
|
706
|
+
ldouble_safe out = (s4 - 4 * s3 * sn + 6 * s2 * pw2(sn) - 4 * s1 * pw3(sn) + cnt * pw4(sn)) / (cnt * pw2(v));
|
707
|
+
return (!is_na_or_inf(out))? std::fmax((double)out, 0.) : (-HUGE_VAL);
|
708
|
+
}
|
709
|
+
|
710
|
+
|
711
|
+
template <class ldouble_safe>
|
712
|
+
double calc_kurtosis_internal(size_t cnt, int x[], int ncat, size_t buffer_cnt[], double buffer_prob[],
|
713
|
+
MissingAction missing_action, CategSplit cat_split_type, RNG_engine &rnd_generator)
|
714
|
+
{
|
715
|
+
/* This calculation proceeds as follows:
|
716
|
+
- If splitting by subsets, it will assign a random weight ~Unif(0,1) to
|
717
|
+
each category, and approximate kurtosis by sampling from such distribution
|
718
|
+
with the same probabilities as given by the current counts.
|
719
|
+
- If splitting by isolating one category, will binarize at each categorical level,
|
720
|
+
assume the values are zero or one, and output the average assuming each categorical
|
721
|
+
level has equal probability of being picked.
|
722
|
+
(Note that both are misleading heuristics, but might be better than random)
|
723
|
+
*/
|
724
|
+
double sum_kurt = 0;
|
725
|
+
|
726
|
+
cnt -= buffer_cnt[ncat];
|
727
|
+
if (cnt <= 1) return -HUGE_VAL;
|
728
|
+
ldouble_safe cnt_l = (ldouble_safe) cnt;
|
729
|
+
for (int cat = 0; cat < ncat; cat++)
|
730
|
+
buffer_prob[cat] = buffer_cnt[cat] / cnt_l;
|
731
|
+
|
732
|
+
switch (cat_split_type)
|
733
|
+
{
|
734
|
+
case SubSet:
|
735
|
+
{
|
736
|
+
ldouble_safe temp_v;
|
737
|
+
ldouble_safe s1, s2, s3, s4;
|
738
|
+
ldouble_safe coef;
|
739
|
+
ldouble_safe coef2;
|
740
|
+
ldouble_safe w_this;
|
741
|
+
UniformUnitInterval runif(0, 1);
|
742
|
+
size_t ntry = 50;
|
743
|
+
for (size_t iternum = 0; iternum < 50; iternum++)
|
744
|
+
{
|
745
|
+
s1 = 0; s2 = 0; s3 = 0; s4 = 0;
|
746
|
+
for (int cat = 0; cat < ncat; cat++)
|
747
|
+
{
|
748
|
+
coef = runif(rnd_generator);
|
749
|
+
coef2 = coef * coef;
|
750
|
+
w_this = buffer_prob[cat];
|
751
|
+
s1 = std::fma(w_this, coef, s1);
|
752
|
+
s2 = std::fma(w_this, coef2, s2);
|
753
|
+
s3 = std::fma(w_this, coef2*coef, s3);
|
754
|
+
s4 = std::fma(w_this, coef2*coef2, s4);
|
755
|
+
// s1 += buffer_prob[cat] * pw1(coef);
|
756
|
+
// s2 += buffer_prob[cat] * pw2(coef);
|
757
|
+
// s3 += buffer_prob[cat] * pw3(coef);
|
758
|
+
// s4 += buffer_prob[cat] * pw4(coef);
|
759
|
+
}
|
760
|
+
temp_v = s2 - pw2(s1);
|
761
|
+
if (temp_v <= 0)
|
762
|
+
ntry--;
|
763
|
+
else
|
764
|
+
sum_kurt += (s4 - 4 * s3 * pw1(s1) + 6 * s2 * pw2(s1) - 4 * s1 * pw3(s1) + pw4(s1)) / pw2(temp_v);
|
765
|
+
}
|
766
|
+
if (unlikely(!ntry))
|
767
|
+
return -HUGE_VAL;
|
768
|
+
else if (unlikely(is_na_or_inf(sum_kurt)))
|
769
|
+
return -HUGE_VAL;
|
770
|
+
else
|
771
|
+
return std::fmax(sum_kurt, 0.) / (double)ntry;
|
772
|
+
}
|
773
|
+
|
774
|
+
case SingleCateg:
|
775
|
+
{
|
776
|
+
double p;
|
777
|
+
int ncat_present = ncat;
|
778
|
+
for (int cat = 0; cat < ncat; cat++)
|
779
|
+
{
|
780
|
+
p = buffer_prob[cat];
|
781
|
+
if (p == 0)
|
782
|
+
ncat_present--;
|
783
|
+
else
|
784
|
+
sum_kurt += (p - 4 * p * pw1(p) + 6 * p * pw2(p) - 4 * p * pw3(p) + pw4(p)) / pw2(p - pw2(p));
|
785
|
+
}
|
786
|
+
if (ncat_present <= 1)
|
787
|
+
return -HUGE_VAL;
|
788
|
+
else if (unlikely(is_na_or_inf(sum_kurt)))
|
789
|
+
return -HUGE_VAL;
|
790
|
+
else
|
791
|
+
return std::fmax(sum_kurt, 0.) / (double)ncat_present;
|
792
|
+
}
|
793
|
+
}
|
794
|
+
|
795
|
+
return -1; /* this will never be reached, but CRAN complains otherwise */
|
796
|
+
}
|
797
|
+
|
798
|
+
template <class ldouble_safe>
|
799
|
+
double calc_kurtosis(size_t *restrict ix_arr, size_t st, size_t end, int x[], int ncat, size_t *restrict buffer_cnt, double buffer_prob[],
|
800
|
+
MissingAction missing_action, CategSplit cat_split_type, RNG_engine &rnd_generator)
|
801
|
+
{
|
802
|
+
/* This calculation proceeds as follows:
|
803
|
+
- If splitting by subsets, it will assign a random weight ~Unif(0,1) to
|
804
|
+
each category, and approximate kurtosis by sampling from such distribution
|
805
|
+
with the same probabilities as given by the current counts.
|
806
|
+
- If splitting by isolating one category, will binarize at each categorical level,
|
807
|
+
assume the values are zero or one, and output the average assuming each categorical
|
808
|
+
level has equal probability of being picked.
|
809
|
+
(Note that both are misleading heuristics, but might be better than random)
|
810
|
+
*/
|
811
|
+
size_t cnt = end - st + 1;
|
812
|
+
std::fill(buffer_cnt, buffer_cnt + ncat + 1, (size_t)0);
|
813
|
+
|
814
|
+
if (missing_action == Fail)
|
815
|
+
{
|
816
|
+
for (size_t row = st; row <= end; row++)
|
817
|
+
buffer_cnt[x[ix_arr[row]]]++;
|
818
|
+
}
|
819
|
+
|
820
|
+
else
|
821
|
+
{
|
822
|
+
for (size_t row = st; row <= end; row++)
|
823
|
+
{
|
824
|
+
if (likely(x[ix_arr[row]] >= 0))
|
825
|
+
buffer_cnt[x[ix_arr[row]]]++;
|
826
|
+
else
|
827
|
+
buffer_cnt[ncat]++;
|
828
|
+
}
|
829
|
+
}
|
830
|
+
|
831
|
+
return calc_kurtosis_internal<ldouble_safe>(
|
832
|
+
cnt, x, ncat, buffer_cnt, buffer_prob,
|
833
|
+
missing_action, cat_split_type, rnd_generator);
|
834
|
+
}
|
835
|
+
|
836
|
+
template <class ldouble_safe>
|
837
|
+
double calc_kurtosis(size_t nrows, int x[], int ncat, size_t buffer_cnt[], double buffer_prob[],
|
838
|
+
MissingAction missing_action, CategSplit cat_split_type, RNG_engine &rnd_generator)
|
839
|
+
{
|
840
|
+
size_t cnt = nrows;
|
841
|
+
std::fill(buffer_cnt, buffer_cnt + ncat + 1, (size_t)0);
|
842
|
+
|
843
|
+
if (missing_action == Fail)
|
844
|
+
{
|
845
|
+
for (size_t row = 0; row < nrows; row++)
|
846
|
+
buffer_cnt[x[row]]++;
|
847
|
+
}
|
848
|
+
|
849
|
+
else
|
850
|
+
{
|
851
|
+
for (size_t row = 0; row < nrows; row++)
|
852
|
+
{
|
853
|
+
if (likely(x[row] >= 0))
|
854
|
+
buffer_cnt[x[row]]++;
|
855
|
+
else
|
856
|
+
buffer_cnt[ncat]++;
|
857
|
+
}
|
858
|
+
}
|
859
|
+
|
860
|
+
return calc_kurtosis_internal<ldouble_safe>(
|
861
|
+
cnt, x, ncat, buffer_cnt, buffer_prob,
|
862
|
+
missing_action, cat_split_type, rnd_generator);
|
863
|
+
}
|
864
|
+
|
865
|
+
|
866
|
+
/* TODO: this one should get a buffer preallocated from outside */
|
867
|
+
template <class mapping, class ldouble_safe>
|
868
|
+
double calc_kurtosis_weighted_internal(std::vector<ldouble_safe> &buffer_cnt, int x[], int ncat,
|
869
|
+
double buffer_prob[], MissingAction missing_action, CategSplit cat_split_type,
|
870
|
+
RNG_engine &rnd_generator, mapping &restrict w)
|
871
|
+
{
|
872
|
+
double sum_kurt = 0;
|
873
|
+
|
874
|
+
ldouble_safe cnt = std::accumulate(buffer_cnt.begin(), buffer_cnt.end(), (ldouble_safe)0);
|
875
|
+
|
876
|
+
cnt -= buffer_cnt[ncat];
|
877
|
+
if (unlikely(cnt <= 1)) return -HUGE_VAL;
|
878
|
+
for (int cat = 0; cat < ncat; cat++)
|
879
|
+
buffer_prob[cat] = buffer_cnt[cat] / cnt;
|
880
|
+
|
881
|
+
switch (cat_split_type)
|
882
|
+
{
|
883
|
+
case SubSet:
|
884
|
+
{
|
885
|
+
ldouble_safe temp_v;
|
886
|
+
ldouble_safe s1, s2, s3, s4;
|
887
|
+
ldouble_safe coef, coef2;
|
888
|
+
ldouble_safe w_this;
|
889
|
+
UniformUnitInterval runif(0, 1);
|
890
|
+
size_t ntry = 50;
|
891
|
+
for (size_t iternum = 0; iternum < 50; iternum++)
|
892
|
+
{
|
893
|
+
s1 = 0; s2 = 0; s3 = 0; s4 = 0;
|
894
|
+
for (int cat = 0; cat < ncat; cat++)
|
895
|
+
{
|
896
|
+
coef = runif(rnd_generator);
|
897
|
+
coef2 = coef * coef;
|
898
|
+
w_this = buffer_prob[cat];
|
899
|
+
s1 = std::fma(w_this, coef, s1);
|
900
|
+
s2 = std::fma(w_this, coef2, s2);
|
901
|
+
s3 = std::fma(w_this, coef2*coef, s3);
|
902
|
+
s4 = std::fma(w_this, coef2*coef2, s4);
|
903
|
+
// s1 += buffer_prob[cat] * pw1(coef);
|
904
|
+
// s2 += buffer_prob[cat] * pw2(coef);
|
905
|
+
// s3 += buffer_prob[cat] * pw3(coef);
|
906
|
+
// s4 += buffer_prob[cat] * pw4(coef);
|
907
|
+
}
|
908
|
+
temp_v = s2 - pw2(s1);
|
909
|
+
if (unlikely(temp_v <= 0))
|
910
|
+
ntry--;
|
911
|
+
else
|
912
|
+
sum_kurt += (s4 - 4 * s3 * pw1(s1) + 6 * s2 * pw2(s1) - 4 * s1 * pw3(s1) + pw4(s1)) / pw2(temp_v);
|
913
|
+
}
|
914
|
+
if (unlikely(!ntry))
|
915
|
+
return -HUGE_VAL;
|
916
|
+
else if (unlikely(is_na_or_inf(sum_kurt)))
|
917
|
+
return -HUGE_VAL;
|
918
|
+
else
|
919
|
+
return std::fmax(sum_kurt, 0.) / (double)ntry;
|
920
|
+
}
|
921
|
+
|
922
|
+
case SingleCateg:
|
923
|
+
{
|
924
|
+
double p;
|
925
|
+
int ncat_present = ncat;
|
926
|
+
for (int cat = 0; cat < ncat; cat++)
|
927
|
+
{
|
928
|
+
p = buffer_prob[cat];
|
929
|
+
if (p == 0)
|
930
|
+
ncat_present--;
|
931
|
+
else
|
932
|
+
sum_kurt += (p - 4 * p * pw1(p) + 6 * p * pw2(p) - 4 * p * pw3(p) + pw4(p)) / pw2(p - pw2(p));
|
933
|
+
}
|
934
|
+
if (ncat_present <= 1)
|
935
|
+
return -HUGE_VAL;
|
936
|
+
else if (unlikely(is_na_or_inf(sum_kurt)))
|
937
|
+
return -HUGE_VAL;
|
938
|
+
else
|
939
|
+
return std::fmax(sum_kurt, 0.) / (double)ncat_present;
|
940
|
+
}
|
941
|
+
}
|
942
|
+
|
943
|
+
return -1; /* this will never be reached, but CRAN complains otherwise */
|
944
|
+
}
|
945
|
+
|
946
|
+
template <class mapping, class ldouble_safe>
|
947
|
+
double calc_kurtosis_weighted(size_t ix_arr[], size_t st, size_t end, int x[], int ncat, double buffer_prob[],
|
948
|
+
MissingAction missing_action, CategSplit cat_split_type, RNG_engine &rnd_generator,
|
949
|
+
mapping &restrict w)
|
950
|
+
{
|
951
|
+
std::vector<ldouble_safe> buffer_cnt(ncat+1, 0.);
|
952
|
+
ldouble_safe w_this;
|
953
|
+
|
954
|
+
for (size_t row = st; row <= end; row++)
|
955
|
+
{
|
956
|
+
w_this = w[ix_arr[row]];
|
957
|
+
if (likely(x[ix_arr[row]] >= 0))
|
958
|
+
buffer_cnt[x[ix_arr[row]]] += w_this;
|
959
|
+
else
|
960
|
+
buffer_cnt[ncat] += w_this;
|
961
|
+
}
|
962
|
+
|
963
|
+
return calc_kurtosis_weighted_internal<mapping, ldouble_safe>(
|
964
|
+
buffer_cnt, x, ncat,
|
965
|
+
buffer_prob, missing_action, cat_split_type,
|
966
|
+
rnd_generator, w);
|
967
|
+
}
|
968
|
+
|
969
|
+
template <class real_t, class ldouble_safe>
|
970
|
+
double calc_kurtosis_weighted(size_t nrows, int x[], int ncat, double *restrict buffer_prob,
|
971
|
+
MissingAction missing_action, CategSplit cat_split_type,
|
972
|
+
RNG_engine &rnd_generator, real_t *restrict w)
|
973
|
+
{
|
974
|
+
std::vector<ldouble_safe> buffer_cnt(ncat+1, 0.);
|
975
|
+
ldouble_safe w_this;
|
976
|
+
|
977
|
+
for (size_t row = 0; row < nrows; row++)
|
978
|
+
{
|
979
|
+
w_this = w[row];
|
980
|
+
if (likely(x[row] >= 0))
|
981
|
+
buffer_cnt[x[row]] += w_this;
|
982
|
+
else
|
983
|
+
buffer_cnt[ncat] += w_this;
|
984
|
+
}
|
985
|
+
|
986
|
+
return calc_kurtosis_weighted_internal<real_t *restrict, ldouble_safe>(
|
987
|
+
buffer_cnt, x, ncat,
|
988
|
+
buffer_prob, missing_action, cat_split_type,
|
989
|
+
rnd_generator, w);
|
990
|
+
}
|
991
|
+
|
992
|
+
template <class int_t, class ldouble_safe>
|
993
|
+
double expected_sd_cat(double p[], size_t n, int_t pos[])
|
994
|
+
{
|
995
|
+
if (n <= 1) return 0;
|
996
|
+
|
997
|
+
ldouble_safe cum_var = -square(p[pos[0]]) / 3.0 - p[pos[0]] * p[pos[1]] / 2.0 + p[pos[0]] / 3.0 - square(p[pos[1]]) / 3.0 + p[pos[1]] / 3.0;
|
998
|
+
for (size_t cat1 = 2; cat1 < n; cat1++)
|
999
|
+
{
|
1000
|
+
cum_var += p[pos[cat1]] / 3.0 - square(p[pos[cat1]]) / 3.0;
|
1001
|
+
for (size_t cat2 = 0; cat2 < cat1; cat2++)
|
1002
|
+
cum_var -= p[pos[cat1]] * p[pos[cat2]] / 2.0;
|
1003
|
+
}
|
1004
|
+
return std::sqrt(std::fmax(cum_var, (ldouble_safe)0));
|
1005
|
+
}
|
1006
|
+
|
1007
|
+
template <class number, class int_t, class ldouble_safe>
|
1008
|
+
double expected_sd_cat(number *restrict counts, double *restrict p, size_t n, int_t *restrict pos)
|
1009
|
+
{
|
1010
|
+
if (n <= 1) return 0;
|
1011
|
+
|
1012
|
+
number tot = std::accumulate(pos, pos + n, (number)0, [&counts](number tot, const size_t ix){return tot + counts[ix];});
|
1013
|
+
ldouble_safe cnt_div = (ldouble_safe) tot;
|
1014
|
+
for (size_t cat = 0; cat < n; cat++)
|
1015
|
+
p[pos[cat]] = (ldouble_safe)counts[pos[cat]] / cnt_div;
|
1016
|
+
|
1017
|
+
return expected_sd_cat<int_t, ldouble_safe>(p, n, pos);
|
1018
|
+
}
|
1019
|
+
|
1020
|
+
template <class number, class int_t, class ldouble_safe>
|
1021
|
+
double expected_sd_cat_single(number *restrict counts, double *restrict p, size_t n, int_t *restrict pos, size_t cat_exclude, number cnt)
|
1022
|
+
{
|
1023
|
+
if (cat_exclude == 0)
|
1024
|
+
return expected_sd_cat<number, int_t, ldouble_safe>(counts, p, n-1, pos + 1);
|
1025
|
+
|
1026
|
+
else if (cat_exclude == (n-1))
|
1027
|
+
return expected_sd_cat<number, int_t, ldouble_safe>(counts, p, n-1, pos);
|
1028
|
+
|
1029
|
+
size_t ix_exclude = pos[cat_exclude];
|
1030
|
+
|
1031
|
+
ldouble_safe cnt_div = (ldouble_safe) (cnt - counts[ix_exclude]);
|
1032
|
+
for (size_t cat = 0; cat < n; cat++)
|
1033
|
+
p[pos[cat]] = (ldouble_safe)counts[pos[cat]] / cnt_div;
|
1034
|
+
|
1035
|
+
ldouble_safe cum_var;
|
1036
|
+
if (cat_exclude != 1)
|
1037
|
+
cum_var = -square(p[pos[0]]) / 3.0 - p[pos[0]] * p[pos[1]] / 2.0 + p[pos[0]] / 3.0 - square(p[pos[1]]) / 3.0 + p[pos[1]] / 3.0;
|
1038
|
+
else
|
1039
|
+
cum_var = -square(p[pos[0]]) / 3.0 - p[pos[0]] * p[pos[2]] / 2.0 + p[pos[0]] / 3.0 - square(p[pos[2]]) / 3.0 + p[pos[2]] / 3.0;
|
1040
|
+
for (size_t cat1 = (cat_exclude == 1)? 3 : 2; cat1 < n; cat1++)
|
1041
|
+
{
|
1042
|
+
if (pos[cat1] == ix_exclude) continue;
|
1043
|
+
cum_var += p[pos[cat1]] / 3.0 - square(p[pos[cat1]]) / 3.0;
|
1044
|
+
for (size_t cat2 = 0; cat2 < cat1; cat2++)
|
1045
|
+
{
|
1046
|
+
if (pos[cat2] == ix_exclude) continue;
|
1047
|
+
cum_var -= p[pos[cat1]] * p[pos[cat2]] / 2.0;
|
1048
|
+
}
|
1049
|
+
|
1050
|
+
}
|
1051
|
+
return std::sqrt(std::fmax(cum_var, (ldouble_safe)0));
|
1052
|
+
}
|
1053
|
+
|
1054
|
+
template <class number, class int_t, class ldouble_safe>
|
1055
|
+
double expected_sd_cat_internal(int ncat, number *restrict buffer_cnt, ldouble_safe cnt_l,
|
1056
|
+
int_t *restrict buffer_pos, double *restrict buffer_prob)
|
1057
|
+
{
|
1058
|
+
/* move zero-valued to the beginning */
|
1059
|
+
std::iota(buffer_pos, buffer_pos + ncat, (int_t)0);
|
1060
|
+
int_t st_pos = 0;
|
1061
|
+
int ncat_present = 0;
|
1062
|
+
int_t temp;
|
1063
|
+
for (int cat = 0; cat < ncat; cat++)
|
1064
|
+
{
|
1065
|
+
if (buffer_cnt[cat])
|
1066
|
+
{
|
1067
|
+
ncat_present++;
|
1068
|
+
buffer_prob[cat] = (ldouble_safe) buffer_cnt[cat] / cnt_l;
|
1069
|
+
}
|
1070
|
+
|
1071
|
+
else
|
1072
|
+
{
|
1073
|
+
temp = buffer_pos[st_pos];
|
1074
|
+
buffer_pos[st_pos] = buffer_pos[cat];
|
1075
|
+
buffer_pos[cat] = temp;
|
1076
|
+
st_pos++;
|
1077
|
+
}
|
1078
|
+
}
|
1079
|
+
|
1080
|
+
if (ncat_present <= 1) return 0;
|
1081
|
+
return expected_sd_cat<int_t, ldouble_safe>(buffer_prob, ncat_present, buffer_pos + st_pos);
|
1082
|
+
}
|
1083
|
+
|
1084
|
+
|
1085
|
+
template <class int_t, class ldouble_safe>
|
1086
|
+
double expected_sd_cat(size_t *restrict ix_arr, size_t st, size_t end, int x[], int ncat,
|
1087
|
+
MissingAction missing_action,
|
1088
|
+
size_t *restrict buffer_cnt, int_t *restrict buffer_pos, double buffer_prob[])
|
1089
|
+
{
|
1090
|
+
/* generate counts */
|
1091
|
+
std::fill(buffer_cnt, buffer_cnt + ncat + 1, (size_t)0);
|
1092
|
+
size_t cnt = end - st + 1;
|
1093
|
+
|
1094
|
+
if (missing_action != Fail)
|
1095
|
+
{
|
1096
|
+
int xval;
|
1097
|
+
for (size_t row = st; row <= end; row++)
|
1098
|
+
{
|
1099
|
+
xval = x[ix_arr[row]];
|
1100
|
+
if (unlikely(xval < 0))
|
1101
|
+
buffer_cnt[ncat]++;
|
1102
|
+
else
|
1103
|
+
buffer_cnt[xval]++;
|
1104
|
+
}
|
1105
|
+
cnt -= buffer_cnt[ncat];
|
1106
|
+
if (cnt == 0) return 0;
|
1107
|
+
}
|
1108
|
+
|
1109
|
+
else
|
1110
|
+
{
|
1111
|
+
for (size_t row = st; row <= end; row++)
|
1112
|
+
{
|
1113
|
+
if (likely(x[ix_arr[row]] >= 0)) buffer_cnt[x[ix_arr[row]]]++;
|
1114
|
+
}
|
1115
|
+
}
|
1116
|
+
|
1117
|
+
return expected_sd_cat_internal<size_t, int_t, ldouble_safe>(ncat, buffer_cnt, cnt, buffer_pos, buffer_prob);
|
1118
|
+
}
|
1119
|
+
|
1120
|
+
template <class mapping, class int_t, class ldouble_safe>
|
1121
|
+
double expected_sd_cat_weighted(size_t *restrict ix_arr, size_t st, size_t end, int x[], int ncat,
|
1122
|
+
MissingAction missing_action, mapping &restrict w,
|
1123
|
+
double *restrict buffer_cnt, int_t *restrict buffer_pos, double *restrict buffer_prob)
|
1124
|
+
{
|
1125
|
+
/* generate counts */
|
1126
|
+
std::fill(buffer_cnt, buffer_cnt + ncat + 1, 0.);
|
1127
|
+
ldouble_safe cnt = 0;
|
1128
|
+
|
1129
|
+
if (missing_action != Fail)
|
1130
|
+
{
|
1131
|
+
int xval;
|
1132
|
+
double w_this;
|
1133
|
+
for (size_t row = st; row <= end; row++)
|
1134
|
+
{
|
1135
|
+
xval = x[ix_arr[row]];
|
1136
|
+
w_this = w[ix_arr[row]];
|
1137
|
+
|
1138
|
+
if (unlikely(xval < 0)) {
|
1139
|
+
buffer_cnt[ncat] += w_this;
|
1140
|
+
}
|
1141
|
+
else {
|
1142
|
+
buffer_cnt[xval] += w_this;
|
1143
|
+
cnt += w_this;
|
1144
|
+
}
|
1145
|
+
}
|
1146
|
+
if (cnt == 0) return 0;
|
1147
|
+
}
|
1148
|
+
|
1149
|
+
else
|
1150
|
+
{
|
1151
|
+
for (size_t row = st; row <= end; row++)
|
1152
|
+
{
|
1153
|
+
if (likely(x[ix_arr[row]] >= 0))
|
1154
|
+
{
|
1155
|
+
buffer_cnt[x[ix_arr[row]]] += w[ix_arr[row]];
|
1156
|
+
}
|
1157
|
+
}
|
1158
|
+
for (int cat = 0; cat < ncat; cat++)
|
1159
|
+
cnt += buffer_cnt[cat];
|
1160
|
+
if (unlikely(cnt == 0)) return 0;
|
1161
|
+
}
|
1162
|
+
|
1163
|
+
return expected_sd_cat_internal<double, int_t, ldouble_safe>(ncat, buffer_cnt, cnt, buffer_pos, buffer_prob);
|
1164
|
+
}
|
1165
|
+
|
1166
|
+
/* Note: this isn't exactly comparable to the pooled gain from numeric variables,
|
1167
|
+
but among all the possible options, this is what happens to end up in the most
|
1168
|
+
similar scale when considering standardized gain. */
|
1169
|
+
template <class number, class ldouble_safe>
|
1170
|
+
double categ_gain(number cnt_left, number cnt_right,
|
1171
|
+
ldouble_safe s_left, ldouble_safe s_right,
|
1172
|
+
ldouble_safe base_info, ldouble_safe cnt)
|
1173
|
+
{
|
1174
|
+
return (
|
1175
|
+
base_info -
|
1176
|
+
(((cnt_left <= 1)? 0 : ((ldouble_safe)cnt_left * std::log((ldouble_safe)cnt_left))) - s_left) -
|
1177
|
+
(((cnt_right <= 1)? 0 : ((ldouble_safe)cnt_right * std::log((ldouble_safe)cnt_right))) - s_right)
|
1178
|
+
) / cnt;
|
1179
|
+
}
|
1180
|
+
|
1181
|
+
|
1182
|
+
/* A couple notes about gain calculation:
|
1183
|
+
|
1184
|
+
Here one wants to find the best split point, maximizing either:
|
1185
|
+
(1/sigma) * (sigma - (1/n)*(n_left*sigma_left + n_right*sigma_right))
|
1186
|
+
or:
|
1187
|
+
(1/sigma) * (sigma - (1/2)*(sigma_left + sigma_right))
|
1188
|
+
|
1189
|
+
All the algorithms here use the sorted-indices approach, which is
|
1190
|
+
an exact method (note that there's still room for optimization by adding the
|
1191
|
+
unsorted approach for small sample sizes and for sparse matrices).
|
1192
|
+
|
1193
|
+
A naive approach would move observations one at a time from right
|
1194
|
+
to left using this formula:
|
1195
|
+
sigma = (ssq - s^2/n) / n
|
1196
|
+
ssq = sum(x^2)
|
1197
|
+
s = sum(x)
|
1198
|
+
But such approach has poor numerical precision, and this library is
|
1199
|
+
aimed precisely at cases in which there are outliers in the data.
|
1200
|
+
It's possible to improve the numerical precision by standardizing the
|
1201
|
+
data beforehand, but this library uses instead a more exact two-pass
|
1202
|
+
sigma calculation observation-by-observation (from left to right and
|
1203
|
+
from right to left, keeping the calculations of the first pass in an
|
1204
|
+
array and calculating gain in the second pass), but there's
|
1205
|
+
other methods too.
|
1206
|
+
|
1207
|
+
If one is aiming at maximizing the pooled gain, it's possible to
|
1208
|
+
simplify either the gain or the increase in gain without involving
|
1209
|
+
'ssq'. Assuming one already has 'ssq' and 's' calculated for the left and
|
1210
|
+
right partitions, and one wants to move one ovservation from right to left,
|
1211
|
+
the following hold:
|
1212
|
+
s_right = s - s_left
|
1213
|
+
ssq_right = ssq - ssq_left
|
1214
|
+
n_right = n - n_left
|
1215
|
+
If one then moves observation x, these are updated as follows:
|
1216
|
+
s_left_new = s_left + x
|
1217
|
+
s_right_new = s - s_left - x
|
1218
|
+
ssq_left_new = ssq_left + x^2
|
1219
|
+
ssq_right_new = ssq - ssq_left - x^2
|
1220
|
+
n_left_new = n_left + 1
|
1221
|
+
n_right_new = n - n_left - 1
|
1222
|
+
Gain is then:
|
1223
|
+
(1/sigma) * (sigma - (1/n)*({ssq_left_new - s_left_new^2/n_left_new} + {ssq_right_new - s_right_new^2/n_right_new}))
|
1224
|
+
Which simplifies to:
|
1225
|
+
1 - (1/(sigma*n))(ssq - ( (s_left + x)^2/(n_left+1) + (s - (s_left + x))^2/(n - (n_left+1)) ))
|
1226
|
+
Since 'sigma', n', and 'ssq' are constant, they can be ignored when determining the
|
1227
|
+
maximum gain - that is, one is interest in finding the point that maximizes:
|
1228
|
+
(s_left+x)^2/(n_left+1) + (s-(s_left+x))^2/(n-(n_left+1))
|
1229
|
+
And this calculation will be robust-enough when dealing with numbers that were
|
1230
|
+
already standardized beforehand, as the extended model does at each step.
|
1231
|
+
Note however that, when fitting this model, one is usually interested in evaluating
|
1232
|
+
the actual gain, standardized by the standard deviation, as it will try different
|
1233
|
+
linear combinations which will give different standard deviations, so this simpler
|
1234
|
+
formula cannot be applied unless only one linear combination is probed.
|
1235
|
+
|
1236
|
+
One can also look at:
|
1237
|
+
diff_gain = (1/sigma) * (gain_new - gain)
|
1238
|
+
Which can be simplified to something that doesn't include sums of squares:
|
1239
|
+
(1/(sigma*n))*( -s_left^2/n_left - (s-s_left)^2/(n-n_left) + (s_left+x)^2/(n_left+1) + (s-(s_left+x))^2/(n-(n_left+1)) )
|
1240
|
+
And this calculation would in theory allow getting the actual standardized gain.
|
1241
|
+
In practice however, this calculation can have poor numerical precision when the
|
1242
|
+
sample size is large, so the functions here do not even attempt at calculating it,
|
1243
|
+
and this is the reason why the two-pass approach is preferred.
|
1244
|
+
|
1245
|
+
The averaged SD formula unfortunately doesn't reduce to something that would involve
|
1246
|
+
only sums.
|
1247
|
+
*/
|
1248
|
+
|
1249
|
+
/* TODO: maybe it's not a good idea to use the two-pass approach with un-standardized
|
1250
|
+
variables at large sample sizes (ndim=1), considering that they come in sorted order.
|
1251
|
+
Maybe it should instead use sums of centered squares: sigma = sqrt((x-mean(x))^2/n)
|
1252
|
+
The sums of centered squares method is also likely to be more precise. */
|
1253
|
+
|
1254
|
+
template <class real_t>
|
1255
|
+
double midpoint(real_t x, real_t y)
|
1256
|
+
{
|
1257
|
+
real_t m = x + (y-x)/(real_t)2;
|
1258
|
+
if (likely((double)m < (double)y))
|
1259
|
+
return m;
|
1260
|
+
else {
|
1261
|
+
m = std::nextafter(m, y);
|
1262
|
+
if (m > x && m < y)
|
1263
|
+
return m;
|
1264
|
+
else
|
1265
|
+
return x;
|
1266
|
+
}
|
1267
|
+
}
|
1268
|
+
|
1269
|
+
template <class real_t>
|
1270
|
+
double midpoint_with_reorder(real_t x, real_t y)
|
1271
|
+
{
|
1272
|
+
if (x < y)
|
1273
|
+
return midpoint(x, y);
|
1274
|
+
else
|
1275
|
+
return midpoint(y, x);
|
1276
|
+
}
|
1277
|
+
|
1278
|
+
#define sd_gain(sd, sd_left, sd_right) (1. - ((sd_left) + (sd_right)) / (2. * (sd)))
|
1279
|
+
#define pooled_gain(sd, cnt, sd_left, sd_right, cnt_left, cnt_right) \
|
1280
|
+
(1. - (1./(sd))*( ( ((real_t)(cnt_left))/(cnt) )*(sd_left) + ( ((real_t)(cnt_right)/(cnt)) )*(sd_right) ))
|
1281
|
+
|
1282
|
+
|
1283
|
+
/* TODO: make this a compensated sum */
|
1284
|
+
template <class real_t, class real_t_>
|
1285
|
+
double find_split_rel_gain_t(real_t_ *restrict x, size_t n, double &restrict split_point)
|
1286
|
+
{
|
1287
|
+
real_t this_gain;
|
1288
|
+
real_t best_gain = -HUGE_VAL;
|
1289
|
+
real_t x1 = 0, x2 = 0;
|
1290
|
+
real_t sum_left = 0, sum_right = 0, sum_tot = 0;
|
1291
|
+
for (size_t row = 0; row < n; row++)
|
1292
|
+
sum_tot += x[row];
|
1293
|
+
for (size_t row = 0; row < n-1; row++)
|
1294
|
+
{
|
1295
|
+
sum_left += x[row];
|
1296
|
+
if (x[row] == x[row+1])
|
1297
|
+
continue;
|
1298
|
+
|
1299
|
+
sum_right = sum_tot - sum_left;
|
1300
|
+
this_gain = sum_left * (sum_left / (real_t)(row+1))
|
1301
|
+
+ sum_right * (sum_right / (real_t)(n-row-1));
|
1302
|
+
if (this_gain > best_gain)
|
1303
|
+
{
|
1304
|
+
best_gain = this_gain;
|
1305
|
+
x1 = x[row]; x2 = x[row+1];
|
1306
|
+
}
|
1307
|
+
}
|
1308
|
+
|
1309
|
+
if (best_gain <= -HUGE_VAL)
|
1310
|
+
return best_gain;
|
1311
|
+
split_point = midpoint(x1, x2);
|
1312
|
+
return std::fmax((double)best_gain, std::numeric_limits<double>::epsilon());
|
1313
|
+
}
|
1314
|
+
|
1315
|
+
template <class real_t_, class ldouble_safe>
|
1316
|
+
double find_split_rel_gain(real_t_ *restrict x, size_t n, double &restrict split_point)
|
1317
|
+
{
|
1318
|
+
if (n < THRESHOLD_LONG_DOUBLE)
|
1319
|
+
return find_split_rel_gain_t<double, real_t_>(x, n, split_point);
|
1320
|
+
else
|
1321
|
+
return find_split_rel_gain_t<ldouble_safe, real_t_>(x, n, split_point);
|
1322
|
+
}
|
1323
|
+
|
1324
|
+
/* Note: there is no 'weighted' version of 'find_split_rel_gain' with unindexed 'x', because calling it would
|
1325
|
+
imply having to argsort the 'x' values in order to sort the weights, which is less efficient. */
|
1326
|
+
|
1327
|
+
template <class real_t, class real_t_>
|
1328
|
+
double find_split_rel_gain_t(real_t_ *restrict x, real_t_ xmean, size_t *restrict ix_arr, size_t st, size_t end, double &split_point, size_t &restrict split_ix)
|
1329
|
+
{
|
1330
|
+
real_t this_gain;
|
1331
|
+
real_t best_gain = -HUGE_VAL;
|
1332
|
+
split_ix = 0; /* <- avoid out-of-bounds at the end */
|
1333
|
+
real_t sum_left = 0, sum_right = 0, sum_tot = 0;
|
1334
|
+
for (size_t row = st; row <= end; row++)
|
1335
|
+
sum_tot += x[ix_arr[row]] - xmean;
|
1336
|
+
for (size_t row = st; row < end; row++)
|
1337
|
+
{
|
1338
|
+
sum_left += x[ix_arr[row]] - xmean;
|
1339
|
+
if (x[ix_arr[row]] == x[ix_arr[row+1]])
|
1340
|
+
continue;
|
1341
|
+
|
1342
|
+
sum_right = sum_tot - sum_left;
|
1343
|
+
this_gain = sum_left * (sum_left / (real_t)(row - st + 1))
|
1344
|
+
+ sum_right * (sum_right / (real_t)(end - row));
|
1345
|
+
if (this_gain > best_gain)
|
1346
|
+
{
|
1347
|
+
best_gain = this_gain;
|
1348
|
+
split_ix = row;
|
1349
|
+
}
|
1350
|
+
}
|
1351
|
+
|
1352
|
+
if (best_gain <= -HUGE_VAL)
|
1353
|
+
return best_gain;
|
1354
|
+
split_point = midpoint(x[ix_arr[split_ix]], x[ix_arr[split_ix+1]]);
|
1355
|
+
return std::fmax((double)best_gain, std::numeric_limits<double>::epsilon());
|
1356
|
+
}
|
1357
|
+
|
1358
|
+
template <class real_t_, class ldouble_safe>
|
1359
|
+
double find_split_rel_gain(real_t_ *restrict x, real_t_ xmean, size_t *restrict ix_arr, size_t st, size_t end, double &restrict split_point, size_t &restrict split_ix)
|
1360
|
+
{
|
1361
|
+
if ((end-st+1) < THRESHOLD_LONG_DOUBLE)
|
1362
|
+
return find_split_rel_gain_t<double, real_t_>(x, xmean, ix_arr, st, end, split_point, split_ix);
|
1363
|
+
else
|
1364
|
+
return find_split_rel_gain_t<ldouble_safe, real_t_>(x, xmean, ix_arr, st, end, split_point, split_ix);
|
1365
|
+
}
|
1366
|
+
|
1367
|
+
template <class real_t, class real_t_, class mapping>
|
1368
|
+
double find_split_rel_gain_weighted_t(real_t_ *restrict x, real_t_ xmean, size_t *restrict ix_arr, size_t st, size_t end, double &split_point, size_t &restrict split_ix, mapping &restrict w)
|
1369
|
+
{
|
1370
|
+
real_t this_gain;
|
1371
|
+
real_t best_gain = -HUGE_VAL;
|
1372
|
+
split_ix = 0; /* <- avoid out-of-bounds at the end */
|
1373
|
+
real_t sum_left = 0, sum_right = 0, sum_tot = 0, sumw = 0, sumw_tot = 0;
|
1374
|
+
|
1375
|
+
for (size_t row = st; row <= end; row++)
|
1376
|
+
sumw_tot += w[ix_arr[row]];
|
1377
|
+
for (size_t row = st; row <= end; row++)
|
1378
|
+
sum_tot += x[ix_arr[row]] - xmean;
|
1379
|
+
for (size_t row = st; row < end; row++)
|
1380
|
+
{
|
1381
|
+
sumw += w[ix_arr[row]];
|
1382
|
+
sum_left += x[ix_arr[row]] - xmean;
|
1383
|
+
if (x[ix_arr[row]] == x[ix_arr[row+1]])
|
1384
|
+
continue;
|
1385
|
+
|
1386
|
+
sum_right = sum_tot - sum_left;
|
1387
|
+
this_gain = sum_left * (sum_left / sumw)
|
1388
|
+
+ sum_right * (sum_right / (sumw_tot - sumw));
|
1389
|
+
if (this_gain > best_gain)
|
1390
|
+
{
|
1391
|
+
best_gain = this_gain;
|
1392
|
+
split_ix = row;
|
1393
|
+
}
|
1394
|
+
}
|
1395
|
+
|
1396
|
+
if (best_gain <= -HUGE_VAL)
|
1397
|
+
return best_gain;
|
1398
|
+
split_point = midpoint(x[ix_arr[split_ix]], x[ix_arr[split_ix+1]]);
|
1399
|
+
return std::fmax((double)best_gain, std::numeric_limits<double>::epsilon());
|
1400
|
+
}
|
1401
|
+
|
1402
|
+
template <class real_t_, class mapping, class ldouble_safe>
|
1403
|
+
double find_split_rel_gain_weighted(real_t_ *restrict x, real_t_ xmean, size_t *restrict ix_arr, size_t st, size_t end, double &restrict split_point, size_t &restrict split_ix, mapping &restrict w)
|
1404
|
+
{
|
1405
|
+
if ((end-st+1) < THRESHOLD_LONG_DOUBLE)
|
1406
|
+
return find_split_rel_gain_weighted_t<double, real_t_, mapping>(x, xmean, ix_arr, st, end, split_point, split_ix, w);
|
1407
|
+
else
|
1408
|
+
return find_split_rel_gain_weighted_t<ldouble_safe, real_t_, mapping>(x, xmean, ix_arr, st, end, split_point, split_ix, w);
|
1409
|
+
}
|
1410
|
+
|
1411
|
+
|
1412
|
+
template <class real_t, class real_t_>
|
1413
|
+
real_t calc_sd_right_to_left(real_t_ *restrict x, size_t n, double *restrict sd_arr)
|
1414
|
+
{
|
1415
|
+
real_t running_mean = 0;
|
1416
|
+
real_t running_ssq = 0;
|
1417
|
+
real_t mean_prev = x[n-1];
|
1418
|
+
for (size_t row = 0; row < n-1; row++)
|
1419
|
+
{
|
1420
|
+
running_mean += (x[n-row-1] - running_mean) / (real_t)(row+1);
|
1421
|
+
running_ssq += (x[n-row-1] - running_mean) * (x[n-row-1] - mean_prev);
|
1422
|
+
mean_prev = running_mean;
|
1423
|
+
sd_arr[n-row-1] = (row == 0)? 0. : std::sqrt(running_ssq / (real_t)(row+1));
|
1424
|
+
}
|
1425
|
+
running_mean += (x[0] - running_mean) / (real_t)n;
|
1426
|
+
running_ssq += (x[0] - running_mean) * (x[0] - mean_prev);
|
1427
|
+
return std::sqrt(running_ssq / (real_t)n);
|
1428
|
+
}
|
1429
|
+
|
1430
|
+
template <class real_t_, class ldouble_safe>
|
1431
|
+
ldouble_safe calc_sd_right_to_left_weighted(real_t_ *restrict x, size_t n, double *restrict sd_arr,
|
1432
|
+
double *restrict w, ldouble_safe &cumw, size_t *restrict sorted_ix)
|
1433
|
+
{
|
1434
|
+
ldouble_safe running_mean = 0;
|
1435
|
+
ldouble_safe running_ssq = 0;
|
1436
|
+
ldouble_safe mean_prev = x[sorted_ix[n-1]];
|
1437
|
+
ldouble_safe cnt = 0;
|
1438
|
+
double w_this;
|
1439
|
+
for (size_t row = 0; row < n-1; row++)
|
1440
|
+
{
|
1441
|
+
w_this = w[sorted_ix[n-row-1]];
|
1442
|
+
cnt += w_this;
|
1443
|
+
running_mean += w_this * (x[sorted_ix[n-row-1]] - running_mean) / cnt;
|
1444
|
+
running_ssq += w_this * ((x[sorted_ix[n-row-1]] - running_mean) * (x[sorted_ix[n-row-1]] - mean_prev));
|
1445
|
+
mean_prev = running_mean;
|
1446
|
+
sd_arr[n-row-1] = (row == 0)? 0. : std::sqrt(running_ssq / cnt);
|
1447
|
+
}
|
1448
|
+
w_this = w[sorted_ix[0]];
|
1449
|
+
cnt += w_this;
|
1450
|
+
running_mean += (x[sorted_ix[0]] - running_mean) / cnt;
|
1451
|
+
running_ssq += w_this * ((x[sorted_ix[0]] - running_mean) * (x[sorted_ix[0]] - mean_prev));
|
1452
|
+
cumw = cnt;
|
1453
|
+
return std::sqrt(running_ssq / cnt);
|
1454
|
+
}
|
1455
|
+
|
1456
|
+
template <class real_t, class real_t_>
|
1457
|
+
real_t calc_sd_right_to_left(real_t_ *restrict x, real_t_ xmean, size_t ix_arr[], size_t st, size_t end, double *restrict sd_arr)
|
1458
|
+
{
|
1459
|
+
real_t running_mean = 0;
|
1460
|
+
real_t running_ssq = 0;
|
1461
|
+
real_t mean_prev = x[ix_arr[end]] - xmean;
|
1462
|
+
size_t n = end - st + 1;
|
1463
|
+
for (size_t row = 0; row < n-1; row++)
|
1464
|
+
{
|
1465
|
+
running_mean += ((x[ix_arr[end-row]] - xmean) - running_mean) / (real_t)(row+1);
|
1466
|
+
running_ssq += ((x[ix_arr[end-row]] - xmean) - running_mean) * ((x[ix_arr[end-row]] - xmean) - mean_prev);
|
1467
|
+
mean_prev = running_mean;
|
1468
|
+
sd_arr[n-row-1] = (row == 0)? 0. : std::sqrt(running_ssq / (real_t)(row+1));
|
1469
|
+
}
|
1470
|
+
running_mean += ((x[ix_arr[st]] - xmean) - running_mean) / (real_t)n;
|
1471
|
+
running_ssq += ((x[ix_arr[st]] - xmean) - running_mean) * ((x[ix_arr[st]] - xmean) - mean_prev);
|
1472
|
+
return std::sqrt(running_ssq / (real_t)n);
|
1473
|
+
}
|
1474
|
+
|
1475
|
+
template <class real_t_, class mapping, class ldouble_safe>
|
1476
|
+
ldouble_safe calc_sd_right_to_left_weighted(real_t_ *restrict x, real_t_ xmean, size_t ix_arr[], size_t st, size_t end,
|
1477
|
+
double *restrict sd_arr, mapping &restrict w, ldouble_safe &cumw)
|
1478
|
+
{
|
1479
|
+
ldouble_safe running_mean = 0;
|
1480
|
+
ldouble_safe running_ssq = 0;
|
1481
|
+
real_t_ mean_prev = x[ix_arr[end]] - xmean;
|
1482
|
+
size_t n = end - st + 1;
|
1483
|
+
ldouble_safe cnt = 0;
|
1484
|
+
double w_this;
|
1485
|
+
for (size_t row = 0; row < n-1; row++)
|
1486
|
+
{
|
1487
|
+
w_this = w[ix_arr[end-row]];
|
1488
|
+
cnt += w_this;
|
1489
|
+
running_mean += w_this * ((x[ix_arr[end-row]] - xmean) - running_mean) / cnt;
|
1490
|
+
running_ssq += w_this * (((x[ix_arr[end-row]] - xmean) - running_mean) * ((x[ix_arr[end-row]] - xmean) - mean_prev));
|
1491
|
+
mean_prev = running_mean;
|
1492
|
+
sd_arr[n-row-1] = (row == 0)? 0. : std::sqrt(running_ssq / cnt);
|
1493
|
+
}
|
1494
|
+
w_this = w[ix_arr[st]];
|
1495
|
+
cnt += w_this;
|
1496
|
+
running_mean += ((x[ix_arr[st]] - xmean) - running_mean) / cnt;
|
1497
|
+
running_ssq += w_this * (((x[ix_arr[st]] - xmean) - running_mean) * ((x[ix_arr[st]] - xmean) - mean_prev));
|
1498
|
+
cumw = cnt;
|
1499
|
+
return std::sqrt(running_ssq / cnt);
|
1500
|
+
}
|
1501
|
+
|
1502
|
+
template <class real_t, class real_t_>
|
1503
|
+
double find_split_std_gain_t(real_t_ *restrict x, size_t n, double *restrict sd_arr,
|
1504
|
+
GainCriterion criterion, double min_gain, double &restrict split_point)
|
1505
|
+
{
|
1506
|
+
real_t full_sd = calc_sd_right_to_left<real_t>(x, n, sd_arr);
|
1507
|
+
real_t running_mean = 0;
|
1508
|
+
real_t running_ssq = 0;
|
1509
|
+
real_t mean_prev = x[0];
|
1510
|
+
real_t best_gain = -HUGE_VAL;
|
1511
|
+
real_t this_sd, this_gain;
|
1512
|
+
real_t n_ = (real_t)n;
|
1513
|
+
size_t best_ix = 0;
|
1514
|
+
for (size_t row = 0; row < n-1; row++)
|
1515
|
+
{
|
1516
|
+
running_mean += (x[row] - running_mean) / (real_t)(row+1);
|
1517
|
+
running_ssq += (x[row] - running_mean) * (x[row] - mean_prev);
|
1518
|
+
mean_prev = running_mean;
|
1519
|
+
if (x[row] == x[row+1])
|
1520
|
+
continue;
|
1521
|
+
|
1522
|
+
this_sd = (row == 0)? 0. : std::sqrt(running_ssq / (real_t)(row+1));
|
1523
|
+
this_gain = (criterion == Pooled)?
|
1524
|
+
pooled_gain(full_sd, n_, this_sd, sd_arr[row+1], row+1, n-row-1)
|
1525
|
+
:
|
1526
|
+
sd_gain(full_sd, this_sd, sd_arr[row+1]);
|
1527
|
+
if (this_gain > best_gain && this_gain > min_gain)
|
1528
|
+
{
|
1529
|
+
best_gain = this_gain;
|
1530
|
+
best_ix = row;
|
1531
|
+
}
|
1532
|
+
}
|
1533
|
+
|
1534
|
+
if (best_gain > -HUGE_VAL)
|
1535
|
+
split_point = midpoint(x[best_ix], x[best_ix+1]);
|
1536
|
+
|
1537
|
+
return best_gain;
|
1538
|
+
}
|
1539
|
+
|
1540
|
+
template <class real_t_, class ldouble_safe>
|
1541
|
+
double find_split_std_gain(real_t_ *restrict x, size_t n, double *restrict sd_arr,
|
1542
|
+
GainCriterion criterion, double min_gain, double &restrict split_point)
|
1543
|
+
{
|
1544
|
+
if (n < THRESHOLD_LONG_DOUBLE)
|
1545
|
+
return find_split_std_gain_t<double, real_t_>(x, n, sd_arr, criterion, min_gain, split_point);
|
1546
|
+
else
|
1547
|
+
return find_split_std_gain_t<ldouble_safe, real_t_>(x, n, sd_arr, criterion, min_gain, split_point);
|
1548
|
+
}
|
1549
|
+
|
1550
|
+
template <class real_t, class ldouble_safe>
|
1551
|
+
double find_split_std_gain_weighted(real_t *restrict x, size_t n, double *restrict sd_arr,
|
1552
|
+
GainCriterion criterion, double min_gain, double &restrict split_point,
|
1553
|
+
double *restrict w, size_t *restrict sorted_ix)
|
1554
|
+
{
|
1555
|
+
ldouble_safe cumw;
|
1556
|
+
double full_sd = calc_sd_right_to_left_weighted(x, n, sd_arr, w, cumw, sorted_ix);
|
1557
|
+
ldouble_safe running_mean = 0;
|
1558
|
+
ldouble_safe running_ssq = 0;
|
1559
|
+
ldouble_safe mean_prev = x[sorted_ix[0]];
|
1560
|
+
double best_gain = -HUGE_VAL;
|
1561
|
+
double this_sd, this_gain;
|
1562
|
+
double w_this;
|
1563
|
+
ldouble_safe currw = 0;
|
1564
|
+
size_t best_ix = 0;
|
1565
|
+
|
1566
|
+
for (size_t row = 0; row < n-1; row++)
|
1567
|
+
{
|
1568
|
+
w_this = w[sorted_ix[row]];
|
1569
|
+
currw += w_this;
|
1570
|
+
running_mean += w_this * (x[sorted_ix[row]] - running_mean) / currw;
|
1571
|
+
running_ssq += w_this * ((x[sorted_ix[row]] - running_mean) * (x[sorted_ix[row]] - mean_prev));
|
1572
|
+
mean_prev = running_mean;
|
1573
|
+
if (x[sorted_ix[row]] == x[sorted_ix[row+1]])
|
1574
|
+
continue;
|
1575
|
+
|
1576
|
+
this_sd = (row == 0)? 0. : std::sqrt(running_ssq / currw);
|
1577
|
+
this_gain = (criterion == Pooled)?
|
1578
|
+
pooled_gain(full_sd, cumw, this_sd, sd_arr[row+1], currw, cumw-currw)
|
1579
|
+
:
|
1580
|
+
sd_gain(full_sd, this_sd, sd_arr[row+1]);
|
1581
|
+
if (this_gain > best_gain && this_gain > min_gain)
|
1582
|
+
{
|
1583
|
+
best_gain = this_gain;
|
1584
|
+
best_ix = row;
|
1585
|
+
}
|
1586
|
+
}
|
1587
|
+
|
1588
|
+
if (best_gain > -HUGE_VAL)
|
1589
|
+
split_point = midpoint(x[sorted_ix[best_ix]], x[sorted_ix[best_ix+1]]);
|
1590
|
+
|
1591
|
+
return best_gain;
|
1592
|
+
}
|
1593
|
+
|
1594
|
+
template <class real_t, class real_t_>
|
1595
|
+
double find_split_std_gain_t(real_t_ *restrict x, real_t_ xmean, size_t ix_arr[], size_t st, size_t end, double *restrict sd_arr,
|
1596
|
+
GainCriterion criterion, double min_gain, double &restrict split_point, size_t &restrict split_ix)
|
1597
|
+
{
|
1598
|
+
real_t full_sd = calc_sd_right_to_left<real_t>(x, xmean, ix_arr, st, end, sd_arr);
|
1599
|
+
real_t running_mean = 0;
|
1600
|
+
real_t running_ssq = 0;
|
1601
|
+
real_t mean_prev = x[ix_arr[st]] - xmean;
|
1602
|
+
real_t best_gain = -HUGE_VAL;
|
1603
|
+
real_t n = (real_t)(end - st + 1);
|
1604
|
+
real_t this_sd, this_gain;
|
1605
|
+
split_ix = st;
|
1606
|
+
for (size_t row = st; row < end; row++)
|
1607
|
+
{
|
1608
|
+
running_mean += ((x[ix_arr[row]] - xmean) - running_mean) / (real_t)(row-st+1);
|
1609
|
+
running_ssq += ((x[ix_arr[row]] - xmean) - running_mean) * ((x[ix_arr[row]] - xmean) - mean_prev);
|
1610
|
+
mean_prev = running_mean;
|
1611
|
+
if (x[ix_arr[row]] == x[ix_arr[row+1]])
|
1612
|
+
continue;
|
1613
|
+
|
1614
|
+
this_sd = (row == st)? 0. : std::sqrt(running_ssq / (real_t)(row-st+1));
|
1615
|
+
this_gain = (criterion == Pooled)?
|
1616
|
+
pooled_gain(full_sd, n, this_sd, sd_arr[row-st+1], row-st+1, end-row)
|
1617
|
+
:
|
1618
|
+
sd_gain(full_sd, this_sd, sd_arr[row-st+1]);
|
1619
|
+
if (this_gain > best_gain && this_gain > min_gain)
|
1620
|
+
{
|
1621
|
+
best_gain = this_gain;
|
1622
|
+
split_ix = row;
|
1623
|
+
}
|
1624
|
+
}
|
1625
|
+
|
1626
|
+
if (best_gain > -HUGE_VAL)
|
1627
|
+
split_point = midpoint(x[ix_arr[split_ix]], x[ix_arr[split_ix+1]]);
|
1628
|
+
|
1629
|
+
return best_gain;
|
1630
|
+
}
|
1631
|
+
|
1632
|
+
template <class real_t_, class ldouble_safe>
|
1633
|
+
double find_split_std_gain(real_t_ *restrict x, real_t_ xmean, size_t ix_arr[], size_t st, size_t end, double *restrict sd_arr,
|
1634
|
+
GainCriterion criterion, double min_gain, double &restrict split_point, size_t &restrict split_ix)
|
1635
|
+
{
|
1636
|
+
if ((end-st+1) < THRESHOLD_LONG_DOUBLE)
|
1637
|
+
return find_split_std_gain_t<double, real_t_>(x, xmean, ix_arr, st, end, sd_arr, criterion, min_gain, split_point, split_ix);
|
1638
|
+
else
|
1639
|
+
return find_split_std_gain_t<ldouble_safe, real_t_>(x, xmean, ix_arr, st, end, sd_arr, criterion, min_gain, split_point, split_ix);
|
1640
|
+
}
|
1641
|
+
|
1642
|
+
template <class real_t, class mapping, class ldouble_safe>
|
1643
|
+
double find_split_std_gain_weighted(real_t *restrict x, real_t xmean, size_t ix_arr[], size_t st, size_t end, double *restrict sd_arr,
|
1644
|
+
GainCriterion criterion, double min_gain, double &restrict split_point, size_t &restrict split_ix, mapping &restrict w)
|
1645
|
+
{
|
1646
|
+
ldouble_safe cumw;
|
1647
|
+
double full_sd = calc_sd_right_to_left_weighted(x, xmean, ix_arr, st, end, sd_arr, w, cumw);
|
1648
|
+
ldouble_safe running_mean = 0;
|
1649
|
+
ldouble_safe running_ssq = 0;
|
1650
|
+
ldouble_safe mean_prev = x[ix_arr[st]] - xmean;
|
1651
|
+
double best_gain = -HUGE_VAL;
|
1652
|
+
ldouble_safe currw = 0;
|
1653
|
+
double this_sd, this_gain;
|
1654
|
+
double w_this;
|
1655
|
+
split_ix = st;
|
1656
|
+
|
1657
|
+
for (size_t row = st; row < end; row++)
|
1658
|
+
{
|
1659
|
+
w_this = w[ix_arr[row]];
|
1660
|
+
currw += w_this;
|
1661
|
+
running_mean += w_this * ((x[ix_arr[row]] - xmean) - running_mean) / currw;
|
1662
|
+
running_ssq += w_this * (((x[ix_arr[row]] - xmean) - running_mean) * ((x[ix_arr[row]] - xmean) - mean_prev));
|
1663
|
+
mean_prev = running_mean;
|
1664
|
+
if (x[ix_arr[row]] == x[ix_arr[row+1]])
|
1665
|
+
continue;
|
1666
|
+
|
1667
|
+
this_sd = (row == st)? 0. : std::sqrt(running_ssq / currw);
|
1668
|
+
this_gain = (criterion == Pooled)?
|
1669
|
+
pooled_gain(full_sd, cumw, this_sd, sd_arr[row-st+1], currw, cumw-currw)
|
1670
|
+
:
|
1671
|
+
sd_gain(full_sd, this_sd, sd_arr[row-st+1]);
|
1672
|
+
if (this_gain > best_gain && this_gain > min_gain)
|
1673
|
+
{
|
1674
|
+
best_gain = this_gain;
|
1675
|
+
split_ix = row;
|
1676
|
+
}
|
1677
|
+
}
|
1678
|
+
|
1679
|
+
if (best_gain > -HUGE_VAL)
|
1680
|
+
split_point = midpoint(x[ix_arr[split_ix]], x[ix_arr[split_ix+1]]);
|
1681
|
+
|
1682
|
+
return best_gain;
|
1683
|
+
}
|
1684
|
+
|
1685
|
+
#ifndef _FOR_R
|
1686
|
+
#if defined(__clang__)
|
1687
|
+
#pragma clang diagnostic push
|
1688
|
+
#pragma clang diagnostic ignored "-Wunknown-attributes"
|
1689
|
+
#endif
|
1690
|
+
#endif
|
1691
|
+
|
1692
|
+
#ifndef _FOR_R
|
1693
|
+
[[gnu::optimize("Ofast")]]
|
1694
|
+
#endif
|
1695
|
+
static inline void xpy1(double *restrict x, double *restrict y, size_t n)
|
1696
|
+
{
|
1697
|
+
for (size_t ix = 0; ix < n; ix++) y[ix] += x[ix];
|
1698
|
+
}
|
1699
|
+
|
1700
|
+
#ifndef _FOR_R
|
1701
|
+
[[gnu::optimize("Ofast")]]
|
1702
|
+
#endif
|
1703
|
+
static inline void axpy1(const double a, double *restrict x, double *restrict y, size_t n)
|
1704
|
+
{
|
1705
|
+
for (size_t ix = 0; ix < n; ix++) y[ix] = std::fma(a, x[ix], y[ix]);
|
1706
|
+
}
|
1707
|
+
|
1708
|
+
#ifndef _FOR_R
|
1709
|
+
[[gnu::optimize("Ofast")]]
|
1710
|
+
#endif
|
1711
|
+
static inline void xpy1(double *restrict xval, size_t ind[], size_t nnz, double *restrict y)
|
1712
|
+
{
|
1713
|
+
for (size_t ix = 0; ix < nnz; ix++) y[ind[ix]] += xval[ix];
|
1714
|
+
}
|
1715
|
+
|
1716
|
+
#ifndef _FOR_R
|
1717
|
+
[[gnu::optimize("Ofast")]]
|
1718
|
+
#endif
|
1719
|
+
static inline void axpy1(const double a, double *restrict xval, size_t ind[], size_t nnz, double *restrict y)
|
1720
|
+
{
|
1721
|
+
for (size_t ix = 0; ix < nnz; ix++) y[ind[ix]] = std::fma(a, xval[ix], y[ind[ix]]);
|
1722
|
+
}
|
1723
|
+
|
1724
|
+
#ifndef _FOR_R
|
1725
|
+
#if defined(__clang__)
|
1726
|
+
#pragma clang diagnostic pop
|
1727
|
+
#endif
|
1728
|
+
#endif
|
1729
|
+
|
1730
|
+
template <class real_t, class ldouble_safe>
|
1731
|
+
double find_split_full_gain(real_t *restrict x, size_t st, size_t end, size_t *restrict ix_arr,
|
1732
|
+
size_t *restrict cols_use, size_t ncols_use, bool force_cols_use,
|
1733
|
+
double *restrict X_row_major, size_t ncols,
|
1734
|
+
double *restrict Xr, size_t *restrict Xr_ind, size_t *restrict Xr_indptr,
|
1735
|
+
double *restrict buffer_sum_left, double *restrict buffer_sum_tot,
|
1736
|
+
size_t &restrict split_ix, double &restrict split_point,
|
1737
|
+
bool x_uses_ix_arr)
|
1738
|
+
{
|
1739
|
+
if (end <= st) return -HUGE_VAL;
|
1740
|
+
if (cols_use != NULL && ncols_use && (double)ncols_use / (double)ncols < 0.1)
|
1741
|
+
force_cols_use = true;
|
1742
|
+
|
1743
|
+
memset(buffer_sum_tot, 0, (force_cols_use? ncols_use : ncols)*sizeof(double));
|
1744
|
+
if (Xr_indptr == NULL)
|
1745
|
+
{
|
1746
|
+
if (force_cols_use)
|
1747
|
+
{
|
1748
|
+
double *restrict ptr_row;
|
1749
|
+
for (size_t row = st; row <= end; row++)
|
1750
|
+
{
|
1751
|
+
ptr_row = X_row_major + ix_arr[row]*ncols;
|
1752
|
+
for (size_t col = 0; col < ncols_use; col++)
|
1753
|
+
buffer_sum_tot[col] += ptr_row[cols_use[col]];
|
1754
|
+
}
|
1755
|
+
}
|
1756
|
+
|
1757
|
+
else
|
1758
|
+
{
|
1759
|
+
for (size_t row = st; row <= end; row++)
|
1760
|
+
xpy1(X_row_major + ix_arr[row]*ncols, buffer_sum_tot, ncols);
|
1761
|
+
}
|
1762
|
+
}
|
1763
|
+
|
1764
|
+
else
|
1765
|
+
{
|
1766
|
+
if (force_cols_use)
|
1767
|
+
{
|
1768
|
+
size_t *curr_begin;
|
1769
|
+
size_t *row_end;
|
1770
|
+
size_t *curr_col;
|
1771
|
+
double *restrict Xr_this;
|
1772
|
+
size_t *cols_end = cols_use + ncols_use;
|
1773
|
+
for (size_t row = st; row <= end; row++)
|
1774
|
+
{
|
1775
|
+
curr_begin = Xr_ind + Xr_indptr[ix_arr[row]];
|
1776
|
+
row_end = Xr_ind + Xr_indptr[ix_arr[row] + 1];
|
1777
|
+
if (curr_begin == row_end) continue;
|
1778
|
+
curr_col = cols_use;
|
1779
|
+
Xr_this = Xr + Xr_indptr[ix_arr[row]];
|
1780
|
+
|
1781
|
+
while (curr_col < cols_end && curr_begin < row_end)
|
1782
|
+
{
|
1783
|
+
if (*curr_begin == *curr_col)
|
1784
|
+
{
|
1785
|
+
buffer_sum_tot[std::distance(cols_use, curr_col)] += Xr_this[std::distance(curr_begin, row_end)];
|
1786
|
+
curr_col++;
|
1787
|
+
curr_begin++;
|
1788
|
+
}
|
1789
|
+
|
1790
|
+
else
|
1791
|
+
{
|
1792
|
+
if (*curr_begin > *curr_col)
|
1793
|
+
curr_col = std::lower_bound(curr_col, cols_end, *curr_begin);
|
1794
|
+
else
|
1795
|
+
curr_begin = std::lower_bound(curr_begin, row_end, *curr_col);
|
1796
|
+
}
|
1797
|
+
}
|
1798
|
+
}
|
1799
|
+
}
|
1800
|
+
|
1801
|
+
else
|
1802
|
+
{
|
1803
|
+
size_t ptr_this;
|
1804
|
+
for (size_t row = st; row <= end; row++)
|
1805
|
+
{
|
1806
|
+
ptr_this = Xr_indptr[ix_arr[row]];
|
1807
|
+
xpy1(Xr + ptr_this, Xr_ind + ptr_this, Xr_indptr[ix_arr[row]+1] - ptr_this, buffer_sum_tot);
|
1808
|
+
}
|
1809
|
+
}
|
1810
|
+
}
|
1811
|
+
|
1812
|
+
double best_gain = -HUGE_VAL;
|
1813
|
+
double this_gain;
|
1814
|
+
double sl, sr;
|
1815
|
+
double dl, dr;
|
1816
|
+
double vleft, vright;
|
1817
|
+
memset(buffer_sum_left, 0, (force_cols_use? ncols_use : ncols)*sizeof(double));
|
1818
|
+
if (Xr_indptr == NULL)
|
1819
|
+
{
|
1820
|
+
if (!force_cols_use)
|
1821
|
+
{
|
1822
|
+
for (size_t row = st; row < end; row++)
|
1823
|
+
{
|
1824
|
+
xpy1(X_row_major + ix_arr[row]*ncols, buffer_sum_left, ncols);
|
1825
|
+
if (x_uses_ix_arr) {
|
1826
|
+
if (unlikely(x[ix_arr[row]] == x[ix_arr[row+1]])) continue;
|
1827
|
+
}
|
1828
|
+
else {
|
1829
|
+
if (unlikely(x[row] == x[row+1])) continue;
|
1830
|
+
}
|
1831
|
+
|
1832
|
+
vleft = 0;
|
1833
|
+
vright = 0;
|
1834
|
+
dl = (double)(row-st+1);
|
1835
|
+
dr = (double)(end-row);
|
1836
|
+
for (size_t col = 0; col < ncols; col++)
|
1837
|
+
{
|
1838
|
+
sl = buffer_sum_left[col];
|
1839
|
+
vleft += sl * (sl / dl);
|
1840
|
+
sr = buffer_sum_tot[col] - sl;
|
1841
|
+
vright += sr * (sr / dr);
|
1842
|
+
}
|
1843
|
+
|
1844
|
+
this_gain = vleft + vright;
|
1845
|
+
if (this_gain > best_gain)
|
1846
|
+
{
|
1847
|
+
best_gain = this_gain;
|
1848
|
+
split_ix = row;
|
1849
|
+
}
|
1850
|
+
}
|
1851
|
+
}
|
1852
|
+
|
1853
|
+
else
|
1854
|
+
{
|
1855
|
+
double *restrict ptr_row;
|
1856
|
+
for (size_t row = st; row < end; row++)
|
1857
|
+
{
|
1858
|
+
ptr_row = X_row_major + ix_arr[row]*ncols;
|
1859
|
+
for (size_t col = 0; col < ncols_use; col++)
|
1860
|
+
buffer_sum_left[col] += ptr_row[cols_use[col]];
|
1861
|
+
if (x_uses_ix_arr) {
|
1862
|
+
if (unlikely(x[ix_arr[row]] == x[ix_arr[row+1]])) continue;
|
1863
|
+
}
|
1864
|
+
else {
|
1865
|
+
if (unlikely(x[row] == x[row+1])) continue;
|
1866
|
+
}
|
1867
|
+
|
1868
|
+
vleft = 0;
|
1869
|
+
vright = 0;
|
1870
|
+
dl = (double)(row-st+1);
|
1871
|
+
dr = (double)(end-row);
|
1872
|
+
for (size_t col = 0; col < ncols_use; col++)
|
1873
|
+
{
|
1874
|
+
sl = buffer_sum_left[col];
|
1875
|
+
vleft += sl * (sl / dl);
|
1876
|
+
sr = buffer_sum_tot[col] - sl;
|
1877
|
+
vright += sr * (sr / dr);
|
1878
|
+
}
|
1879
|
+
|
1880
|
+
this_gain = vleft + vright;
|
1881
|
+
if (this_gain > best_gain)
|
1882
|
+
{
|
1883
|
+
best_gain = this_gain;
|
1884
|
+
split_ix = row;
|
1885
|
+
}
|
1886
|
+
}
|
1887
|
+
}
|
1888
|
+
}
|
1889
|
+
|
1890
|
+
else
|
1891
|
+
{
|
1892
|
+
if (!force_cols_use)
|
1893
|
+
{
|
1894
|
+
size_t ptr_this;
|
1895
|
+
for (size_t row = st; row < end; row++)
|
1896
|
+
{
|
1897
|
+
ptr_this = Xr_indptr[ix_arr[row]];
|
1898
|
+
xpy1(Xr + ptr_this, Xr_ind + ptr_this, Xr_indptr[ix_arr[row]+1] - ptr_this, buffer_sum_left);
|
1899
|
+
if (x_uses_ix_arr) {
|
1900
|
+
if (unlikely(x[ix_arr[row]] == x[ix_arr[row+1]])) continue;
|
1901
|
+
}
|
1902
|
+
else {
|
1903
|
+
if (unlikely(x[row] == x[row+1])) continue;
|
1904
|
+
}
|
1905
|
+
|
1906
|
+
vleft = 0;
|
1907
|
+
vright = 0;
|
1908
|
+
dl = (double)(row-st+1);
|
1909
|
+
dr = (double)(end-row);
|
1910
|
+
for (size_t col = 0; col < ncols; col++)
|
1911
|
+
{
|
1912
|
+
sl = buffer_sum_left[col];
|
1913
|
+
vleft += sl * (sl / dl);
|
1914
|
+
sr = buffer_sum_tot[col] - sl;
|
1915
|
+
vright += sr * (sr / dr);
|
1916
|
+
}
|
1917
|
+
|
1918
|
+
this_gain = vleft + vright;
|
1919
|
+
if (this_gain > best_gain)
|
1920
|
+
{
|
1921
|
+
best_gain = this_gain;
|
1922
|
+
split_ix = row;
|
1923
|
+
}
|
1924
|
+
}
|
1925
|
+
}
|
1926
|
+
|
1927
|
+
else
|
1928
|
+
{
|
1929
|
+
size_t *curr_begin;
|
1930
|
+
size_t *row_end;
|
1931
|
+
size_t *curr_col;
|
1932
|
+
double *restrict Xr_this;
|
1933
|
+
size_t *cols_end = cols_use + ncols_use;
|
1934
|
+
for (size_t row = st; row < end; row++)
|
1935
|
+
{
|
1936
|
+
curr_begin = Xr_ind + Xr_indptr[ix_arr[row]];
|
1937
|
+
row_end = Xr_ind + Xr_indptr[ix_arr[row] + 1];
|
1938
|
+
if (curr_begin == row_end) goto skip_sum;
|
1939
|
+
curr_col = cols_use;
|
1940
|
+
Xr_this = Xr + Xr_indptr[ix_arr[row]];
|
1941
|
+
while (curr_col < cols_end && curr_begin < row_end)
|
1942
|
+
{
|
1943
|
+
if (*curr_begin == *curr_col)
|
1944
|
+
{
|
1945
|
+
buffer_sum_left[std::distance(cols_use, curr_col)] += Xr_this[std::distance(curr_begin, row_end)];
|
1946
|
+
curr_col++;
|
1947
|
+
curr_begin++;
|
1948
|
+
}
|
1949
|
+
|
1950
|
+
else
|
1951
|
+
{
|
1952
|
+
if (*curr_begin > *curr_col)
|
1953
|
+
curr_col = std::lower_bound(curr_col, cols_end, *curr_begin);
|
1954
|
+
else
|
1955
|
+
curr_begin = std::lower_bound(curr_begin, row_end, *curr_col);
|
1956
|
+
}
|
1957
|
+
}
|
1958
|
+
|
1959
|
+
skip_sum:
|
1960
|
+
if (x_uses_ix_arr) {
|
1961
|
+
if (unlikely(x[ix_arr[row]] == x[ix_arr[row+1]])) continue;
|
1962
|
+
}
|
1963
|
+
else {
|
1964
|
+
if (unlikely(x[row] == x[row+1])) continue;
|
1965
|
+
}
|
1966
|
+
|
1967
|
+
vleft = 0;
|
1968
|
+
vright = 0;
|
1969
|
+
dl = (double)(row-st+1);
|
1970
|
+
dr = (double)(end-row);
|
1971
|
+
for (size_t col = 0; col < ncols_use; col++)
|
1972
|
+
{
|
1973
|
+
sl = buffer_sum_left[col];
|
1974
|
+
vleft += sl * (sl / dl);
|
1975
|
+
sr = buffer_sum_tot[col] - sl;
|
1976
|
+
vright += sr * (sr / dr);
|
1977
|
+
}
|
1978
|
+
|
1979
|
+
this_gain = vleft + vright;
|
1980
|
+
if (this_gain > best_gain)
|
1981
|
+
{
|
1982
|
+
best_gain = this_gain;
|
1983
|
+
split_ix = row;
|
1984
|
+
}
|
1985
|
+
}
|
1986
|
+
}
|
1987
|
+
}
|
1988
|
+
|
1989
|
+
if (best_gain <= -HUGE_VAL) return best_gain;
|
1990
|
+
|
1991
|
+
if (x_uses_ix_arr)
|
1992
|
+
split_point = midpoint(x[ix_arr[split_ix]], x[ix_arr[split_ix+1]]);
|
1993
|
+
else
|
1994
|
+
split_point = midpoint(x[split_ix], x[split_ix+1]);
|
1995
|
+
return best_gain / (ldouble_safe)(end - st + 1);
|
1996
|
+
}
|
1997
|
+
|
1998
|
+
template <class real_t, class mapping, class ldouble_safe>
|
1999
|
+
double find_split_full_gain_weighted(real_t *restrict x, size_t st, size_t end, size_t *restrict ix_arr,
|
2000
|
+
size_t *restrict cols_use, size_t ncols_use, bool force_cols_use,
|
2001
|
+
double *restrict X_row_major, size_t ncols,
|
2002
|
+
double *restrict Xr, size_t *restrict Xr_ind, size_t *restrict Xr_indptr,
|
2003
|
+
double *restrict buffer_sum_left, double *restrict buffer_sum_tot,
|
2004
|
+
size_t &restrict split_ix, double &restrict split_point,
|
2005
|
+
bool x_uses_ix_arr,
|
2006
|
+
mapping &restrict w)
|
2007
|
+
{
|
2008
|
+
if (end <= st) return -HUGE_VAL;
|
2009
|
+
if (cols_use != NULL && ncols_use && (double)ncols_use / (double)ncols < 0.1)
|
2010
|
+
force_cols_use = true;
|
2011
|
+
|
2012
|
+
double wtot = 0;
|
2013
|
+
if (x_uses_ix_arr)
|
2014
|
+
{
|
2015
|
+
for (size_t row = st; row <= end; row++)
|
2016
|
+
wtot += w[ix_arr[row]];
|
2017
|
+
}
|
2018
|
+
|
2019
|
+
else
|
2020
|
+
{
|
2021
|
+
for (size_t row = st; row <= end; row++)
|
2022
|
+
wtot += w[row];
|
2023
|
+
}
|
2024
|
+
|
2025
|
+
memset(buffer_sum_tot, 0, (force_cols_use? ncols_use : ncols)*sizeof(double));
|
2026
|
+
if (Xr_indptr == NULL)
|
2027
|
+
{
|
2028
|
+
if (!force_cols_use)
|
2029
|
+
{
|
2030
|
+
if (x_uses_ix_arr)
|
2031
|
+
{
|
2032
|
+
for (size_t row = st; row <= end; row++)
|
2033
|
+
axpy1(w[ix_arr[row]], X_row_major + ix_arr[row]*ncols, buffer_sum_tot, ncols);
|
2034
|
+
}
|
2035
|
+
|
2036
|
+
else
|
2037
|
+
{
|
2038
|
+
for (size_t row = st; row <= end; row++)
|
2039
|
+
axpy1(w[row], X_row_major + ix_arr[row]*ncols, buffer_sum_tot, ncols);
|
2040
|
+
}
|
2041
|
+
}
|
2042
|
+
|
2043
|
+
else
|
2044
|
+
{
|
2045
|
+
double *restrict ptr_row;
|
2046
|
+
double w_row;
|
2047
|
+
|
2048
|
+
if (x_uses_ix_arr)
|
2049
|
+
{
|
2050
|
+
for (size_t row = st; row <= end; row++)
|
2051
|
+
{
|
2052
|
+
ptr_row = X_row_major + ix_arr[row]*ncols;
|
2053
|
+
w_row = w[ix_arr[row]];
|
2054
|
+
for (size_t col = 0; col < ncols_use; col++)
|
2055
|
+
buffer_sum_tot[col] = std::fma(w_row, ptr_row[cols_use[col]], buffer_sum_tot[col]);
|
2056
|
+
}
|
2057
|
+
}
|
2058
|
+
|
2059
|
+
else
|
2060
|
+
{
|
2061
|
+
for (size_t row = st; row <= end; row++)
|
2062
|
+
{
|
2063
|
+
ptr_row = X_row_major + ix_arr[row]*ncols;
|
2064
|
+
w_row = w[row];
|
2065
|
+
for (size_t col = 0; col < ncols_use; col++)
|
2066
|
+
buffer_sum_tot[col] = std::fma(w_row, ptr_row[cols_use[col]], buffer_sum_tot[col]);
|
2067
|
+
}
|
2068
|
+
}
|
2069
|
+
}
|
2070
|
+
}
|
2071
|
+
|
2072
|
+
else
|
2073
|
+
{
|
2074
|
+
if (!force_cols_use)
|
2075
|
+
{
|
2076
|
+
size_t ptr_this;
|
2077
|
+
if (x_uses_ix_arr)
|
2078
|
+
{
|
2079
|
+
for (size_t row = st; row <= end; row++)
|
2080
|
+
{
|
2081
|
+
ptr_this = Xr_indptr[ix_arr[row]];
|
2082
|
+
axpy1(w[ix_arr[row]], Xr + ptr_this, Xr_ind + ptr_this, Xr_indptr[ix_arr[row]+1] - ptr_this, buffer_sum_tot);
|
2083
|
+
}
|
2084
|
+
}
|
2085
|
+
|
2086
|
+
else
|
2087
|
+
{
|
2088
|
+
for (size_t row = st; row <= end; row++)
|
2089
|
+
{
|
2090
|
+
ptr_this = Xr_indptr[ix_arr[row]];
|
2091
|
+
axpy1(w[row], Xr + ptr_this, Xr_ind + ptr_this, Xr_indptr[ix_arr[row]+1] - ptr_this, buffer_sum_tot);
|
2092
|
+
}
|
2093
|
+
}
|
2094
|
+
}
|
2095
|
+
|
2096
|
+
else
|
2097
|
+
{
|
2098
|
+
size_t *curr_begin;
|
2099
|
+
size_t *row_end;
|
2100
|
+
size_t *curr_col;
|
2101
|
+
double *restrict Xr_this;
|
2102
|
+
size_t *cols_end = cols_use + ncols_use;
|
2103
|
+
double w_row;
|
2104
|
+
for (size_t row = st; row <= end; row++)
|
2105
|
+
{
|
2106
|
+
curr_begin = Xr_ind + Xr_indptr[ix_arr[row]];
|
2107
|
+
row_end = Xr_ind + Xr_indptr[ix_arr[row] + 1];
|
2108
|
+
if (curr_begin == row_end) continue;
|
2109
|
+
curr_col = cols_use;
|
2110
|
+
Xr_this = Xr + Xr_indptr[ix_arr[row]];
|
2111
|
+
w_row = w[x_uses_ix_arr? ix_arr[row] : row];
|
2112
|
+
size_t dtemp;
|
2113
|
+
|
2114
|
+
while (curr_col < cols_end && curr_begin < row_end)
|
2115
|
+
{
|
2116
|
+
if (*curr_begin == *curr_col)
|
2117
|
+
{
|
2118
|
+
dtemp = std::distance(cols_use, curr_col);
|
2119
|
+
buffer_sum_tot[dtemp]
|
2120
|
+
=
|
2121
|
+
std::fma(w_row, Xr_this[std::distance(curr_begin, row_end)], buffer_sum_tot[dtemp]);
|
2122
|
+
curr_col++;
|
2123
|
+
curr_begin++;
|
2124
|
+
}
|
2125
|
+
|
2126
|
+
else
|
2127
|
+
{
|
2128
|
+
if (*curr_begin > *curr_col)
|
2129
|
+
curr_col = std::lower_bound(curr_col, cols_end, *curr_begin);
|
2130
|
+
else
|
2131
|
+
curr_begin = std::lower_bound(curr_begin, row_end, *curr_col);
|
2132
|
+
}
|
2133
|
+
}
|
2134
|
+
}
|
2135
|
+
}
|
2136
|
+
}
|
2137
|
+
|
2138
|
+
double best_gain = -HUGE_VAL;
|
2139
|
+
double this_gain;
|
2140
|
+
double sl, sr;
|
2141
|
+
double vleft, vright;
|
2142
|
+
double wleft = 0;
|
2143
|
+
double w_row;
|
2144
|
+
double wright;
|
2145
|
+
memset(buffer_sum_left, 0, (force_cols_use? ncols_use : ncols)*sizeof(double));
|
2146
|
+
if (Xr_indptr == NULL)
|
2147
|
+
{
|
2148
|
+
if (!force_cols_use)
|
2149
|
+
{
|
2150
|
+
for (size_t row = st; row < end; row++)
|
2151
|
+
{
|
2152
|
+
w_row = w[x_uses_ix_arr? ix_arr[row] : row];
|
2153
|
+
wleft += w_row;
|
2154
|
+
axpy1(w_row, X_row_major + ix_arr[row]*ncols, buffer_sum_left, ncols);
|
2155
|
+
if (x_uses_ix_arr) {
|
2156
|
+
if (unlikely(x[ix_arr[row]] == x[ix_arr[row+1]])) continue;
|
2157
|
+
}
|
2158
|
+
else {
|
2159
|
+
if (unlikely(x[row] == x[row+1])) continue;
|
2160
|
+
}
|
2161
|
+
|
2162
|
+
vleft = 0;
|
2163
|
+
vright = 0;
|
2164
|
+
wright = wtot - wleft;
|
2165
|
+
for (size_t col = 0; col < ncols; col++)
|
2166
|
+
{
|
2167
|
+
sl = buffer_sum_left[col];
|
2168
|
+
vleft += sl * (sl / wleft);
|
2169
|
+
sr = buffer_sum_tot[col] - sl;
|
2170
|
+
vright += sr * (sr / wright);
|
2171
|
+
}
|
2172
|
+
|
2173
|
+
this_gain = vleft + vright;
|
2174
|
+
if (this_gain > best_gain)
|
2175
|
+
{
|
2176
|
+
best_gain = this_gain;
|
2177
|
+
split_ix = row;
|
2178
|
+
}
|
2179
|
+
}
|
2180
|
+
}
|
2181
|
+
|
2182
|
+
else
|
2183
|
+
{
|
2184
|
+
double *restrict ptr_row;
|
2185
|
+
double w_row;
|
2186
|
+
for (size_t row = st; row < end; row++)
|
2187
|
+
{
|
2188
|
+
w_row = w[x_uses_ix_arr? ix_arr[row] : row];
|
2189
|
+
wleft += w_row;
|
2190
|
+
|
2191
|
+
ptr_row = X_row_major + ix_arr[row]*ncols;
|
2192
|
+
for (size_t col = 0; col < ncols_use; col++)
|
2193
|
+
buffer_sum_left[col] = std::fma(w_row, ptr_row[cols_use[col]], buffer_sum_left[col]);
|
2194
|
+
if (x_uses_ix_arr) {
|
2195
|
+
if (unlikely(x[ix_arr[row]] == x[ix_arr[row+1]])) continue;
|
2196
|
+
}
|
2197
|
+
else {
|
2198
|
+
if (unlikely(x[row] == x[row+1])) continue;
|
2199
|
+
}
|
2200
|
+
|
2201
|
+
vleft = 0;
|
2202
|
+
vright = 0;
|
2203
|
+
wright = wtot - wleft;
|
2204
|
+
for (size_t col = 0; col < ncols_use; col++)
|
2205
|
+
{
|
2206
|
+
sl = buffer_sum_left[col];
|
2207
|
+
vleft += sl * (sl / wleft);
|
2208
|
+
sr = buffer_sum_tot[col] - sl;
|
2209
|
+
vright += sr * (sr / wright);
|
2210
|
+
}
|
2211
|
+
|
2212
|
+
this_gain = vleft + vright;
|
2213
|
+
if (this_gain > best_gain)
|
2214
|
+
{
|
2215
|
+
best_gain = this_gain;
|
2216
|
+
split_ix = row;
|
2217
|
+
}
|
2218
|
+
}
|
2219
|
+
}
|
2220
|
+
}
|
2221
|
+
|
2222
|
+
else
|
2223
|
+
{
|
2224
|
+
if (!force_cols_use)
|
2225
|
+
{
|
2226
|
+
size_t ptr_this;
|
2227
|
+
double w_row;
|
2228
|
+
for (size_t row = st; row < end; row++)
|
2229
|
+
{
|
2230
|
+
w_row= w[x_uses_ix_arr? ix_arr[row] : row];
|
2231
|
+
wleft += w_row;
|
2232
|
+
ptr_this = Xr_indptr[ix_arr[row]];
|
2233
|
+
axpy1(w_row, Xr + ptr_this, Xr_ind + ptr_this, Xr_indptr[ix_arr[row]+1] - ptr_this, buffer_sum_left);
|
2234
|
+
if (x_uses_ix_arr) {
|
2235
|
+
if (unlikely(x[ix_arr[row]] == x[ix_arr[row+1]])) continue;
|
2236
|
+
}
|
2237
|
+
else {
|
2238
|
+
if (unlikely(x[row] == x[row+1])) continue;
|
2239
|
+
}
|
2240
|
+
|
2241
|
+
vleft = 0;
|
2242
|
+
vright = 0;
|
2243
|
+
wright = wtot - wleft;
|
2244
|
+
for (size_t col = 0; col < ncols; col++)
|
2245
|
+
{
|
2246
|
+
sl = buffer_sum_left[col];
|
2247
|
+
vleft += sl * (sl / wleft);
|
2248
|
+
sr = buffer_sum_tot[col] - sl;
|
2249
|
+
vright += sr * (sr / wright);
|
2250
|
+
}
|
2251
|
+
|
2252
|
+
this_gain = vleft + vright;
|
2253
|
+
if (this_gain > best_gain)
|
2254
|
+
{
|
2255
|
+
best_gain = this_gain;
|
2256
|
+
split_ix = row;
|
2257
|
+
}
|
2258
|
+
}
|
2259
|
+
}
|
2260
|
+
|
2261
|
+
else
|
2262
|
+
{
|
2263
|
+
size_t *curr_begin;
|
2264
|
+
size_t *row_end;
|
2265
|
+
size_t *curr_col;
|
2266
|
+
double *restrict Xr_this;
|
2267
|
+
size_t *cols_end = cols_use + ncols_use;
|
2268
|
+
double w_row;
|
2269
|
+
size_t dtemp;
|
2270
|
+
for (size_t row = st; row < end; row++)
|
2271
|
+
{
|
2272
|
+
w_row = w[x_uses_ix_arr? ix_arr[row] : row];
|
2273
|
+
wleft += w_row;
|
2274
|
+
|
2275
|
+
curr_begin = Xr_ind + Xr_indptr[ix_arr[row]];
|
2276
|
+
row_end = Xr_ind + Xr_indptr[ix_arr[row] + 1];
|
2277
|
+
if (curr_begin == row_end) goto skip_sum;
|
2278
|
+
curr_col = cols_use;
|
2279
|
+
Xr_this = Xr + Xr_indptr[ix_arr[row]];
|
2280
|
+
while (curr_col < cols_end && curr_begin < row_end)
|
2281
|
+
{
|
2282
|
+
if (*curr_begin == *curr_col)
|
2283
|
+
{
|
2284
|
+
dtemp = std::distance(cols_use, curr_col);
|
2285
|
+
buffer_sum_left[dtemp]
|
2286
|
+
=
|
2287
|
+
std::fma(w_row, Xr_this[std::distance(curr_begin, row_end)], buffer_sum_left[dtemp]);
|
2288
|
+
curr_col++;
|
2289
|
+
curr_begin++;
|
2290
|
+
}
|
2291
|
+
|
2292
|
+
else
|
2293
|
+
{
|
2294
|
+
if (*curr_begin > *curr_col)
|
2295
|
+
curr_col = std::lower_bound(curr_col, cols_end, *curr_begin);
|
2296
|
+
else
|
2297
|
+
curr_begin = std::lower_bound(curr_begin, row_end, *curr_col);
|
2298
|
+
}
|
2299
|
+
}
|
2300
|
+
|
2301
|
+
skip_sum:
|
2302
|
+
if (x_uses_ix_arr) {
|
2303
|
+
if (unlikely(x[ix_arr[row]] == x[ix_arr[row+1]])) continue;
|
2304
|
+
}
|
2305
|
+
else {
|
2306
|
+
if (unlikely(x[row] == x[row+1])) continue;
|
2307
|
+
}
|
2308
|
+
|
2309
|
+
vleft = 0;
|
2310
|
+
vright = 0;
|
2311
|
+
wright = wtot - wleft;
|
2312
|
+
for (size_t col = 0; col < ncols_use; col++)
|
2313
|
+
{
|
2314
|
+
sl = buffer_sum_left[col];
|
2315
|
+
vleft += sl * (sl / wleft);
|
2316
|
+
sr = buffer_sum_tot[col] - sl;
|
2317
|
+
vright += sr * (sr / wright);
|
2318
|
+
}
|
2319
|
+
|
2320
|
+
this_gain = vleft + vright;
|
2321
|
+
if (this_gain > best_gain)
|
2322
|
+
{
|
2323
|
+
best_gain = this_gain;
|
2324
|
+
split_ix = row;
|
2325
|
+
}
|
2326
|
+
}
|
2327
|
+
}
|
2328
|
+
}
|
2329
|
+
|
2330
|
+
if (best_gain <= -HUGE_VAL) return best_gain;
|
2331
|
+
|
2332
|
+
split_point = midpoint(x[ix_arr[split_ix]], x[ix_arr[split_ix+1]]);
|
2333
|
+
return best_gain / wtot;
|
2334
|
+
}
|
2335
|
+
|
2336
|
+
template <class real_t_, class real_t>
|
2337
|
+
double find_split_dens_shortform_t(real_t *restrict x, size_t n, double &restrict split_point)
|
2338
|
+
{
|
2339
|
+
double best_gain = -HUGE_VAL;
|
2340
|
+
size_t n_minus_one = n - 1;
|
2341
|
+
real_t_ xmin = x[0];
|
2342
|
+
real_t_ xmax = x[n-1];
|
2343
|
+
real_t_ xleft, xright;
|
2344
|
+
real_t_ xmid;
|
2345
|
+
double this_gain;
|
2346
|
+
size_t split_ix = 0;
|
2347
|
+
|
2348
|
+
for (size_t ix = 0; ix < n_minus_one; ix++)
|
2349
|
+
{
|
2350
|
+
if (x[ix] == x[ix+1]) continue;
|
2351
|
+
xmid = (real_t_)x[ix] + ((real_t_)x[ix+1] - (real_t_)x[ix]) / (real_t_)2;
|
2352
|
+
xleft = xmid - xmin;
|
2353
|
+
xright = xmax - xmid;
|
2354
|
+
if (unlikely(!xleft || !xright)) continue;
|
2355
|
+
this_gain = (real_t_)square(ix+1) / xleft + (real_t_)square(n_minus_one - ix) / xright;
|
2356
|
+
if (this_gain > best_gain)
|
2357
|
+
{
|
2358
|
+
best_gain = this_gain;
|
2359
|
+
split_ix = ix;
|
2360
|
+
}
|
2361
|
+
}
|
2362
|
+
|
2363
|
+
if (best_gain <= -HUGE_VAL) return best_gain;
|
2364
|
+
|
2365
|
+
real_t_ xtot = (real_t_)xmax - (real_t_)xmin;
|
2366
|
+
real_t_ nleft = (real_t_)(split_ix+1);
|
2367
|
+
real_t_ nright = (real_t_)(n_minus_one - split_ix);
|
2368
|
+
split_point = midpoint(x[split_ix], x[split_ix+1]);
|
2369
|
+
real_t_ rpct_left = split_point / xtot;
|
2370
|
+
rpct_left = std::fmax(rpct_left, std::numeric_limits<double>::min());
|
2371
|
+
real_t_ rpct_right = (real_t_)1 - rpct_left;
|
2372
|
+
rpct_right = std::fmax(rpct_right, std::numeric_limits<double>::min());
|
2373
|
+
|
2374
|
+
real_t_ nl_sq = nleft / (real_t_)n; nl_sq = square(nl_sq);
|
2375
|
+
real_t_ nr_sq = nright / (real_t_)n; nl_sq = square(nr_sq);
|
2376
|
+
|
2377
|
+
return nl_sq / rpct_left + nr_sq / rpct_right;
|
2378
|
+
}
|
2379
|
+
|
2380
|
+
template <class real_t, class ldouble_safe>
|
2381
|
+
double find_split_dens_shortform(real_t *restrict x, size_t n, double &restrict split_point)
|
2382
|
+
{
|
2383
|
+
if (n < INT32_MAX)
|
2384
|
+
return find_split_dens_shortform_t<double, real_t>(x, n, split_point);
|
2385
|
+
else
|
2386
|
+
return find_split_dens_shortform_t<ldouble_safe, real_t>(x, n, split_point);
|
2387
|
+
}
|
2388
|
+
|
2389
|
+
template <class real_t_, class real_t, class mapping>
|
2390
|
+
double find_split_dens_shortform_weighted_t(real_t *restrict x, size_t n, double &restrict split_point, mapping &restrict w, size_t *restrict buffer_indices)
|
2391
|
+
{
|
2392
|
+
double best_gain = -HUGE_VAL;
|
2393
|
+
size_t n_minus_one = n - 1;
|
2394
|
+
real_t_ xmin = x[buffer_indices[0]];
|
2395
|
+
real_t_ xmax = x[buffer_indices[n-1]];
|
2396
|
+
real_t_ xleft, xright;
|
2397
|
+
real_t_ xmid;
|
2398
|
+
double this_gain;
|
2399
|
+
|
2400
|
+
real_t_ wtot = 0;
|
2401
|
+
for (size_t ix = 0; ix < n; ix++)
|
2402
|
+
wtot += w[buffer_indices[ix]];
|
2403
|
+
real_t_ w_left = 0;
|
2404
|
+
real_t_ w_right;
|
2405
|
+
real_t_ best_w = 0;
|
2406
|
+
size_t split_ix = 0;
|
2407
|
+
|
2408
|
+
for (size_t ix = 0; ix < n_minus_one; ix++)
|
2409
|
+
{
|
2410
|
+
w_left += w[buffer_indices[ix]];
|
2411
|
+
if (x[buffer_indices[ix]] == x[buffer_indices[ix+1]]) continue;
|
2412
|
+
xmid = (real_t_)x[buffer_indices[ix]] + ((real_t_)x[buffer_indices[ix+1]] - (real_t_)x[buffer_indices[ix]]) / (real_t_)2;
|
2413
|
+
xleft = xmid - xmin;
|
2414
|
+
xright = xmax - xmid;
|
2415
|
+
if (unlikely(!xleft || !xright)) continue;
|
2416
|
+
|
2417
|
+
w_right = wtot - w_left;
|
2418
|
+
this_gain = square(w_left) / xleft + square(w_right) / xright;
|
2419
|
+
if (this_gain > best_gain)
|
2420
|
+
{
|
2421
|
+
best_gain = this_gain;
|
2422
|
+
best_w = w_left;
|
2423
|
+
split_ix = xmid;
|
2424
|
+
}
|
2425
|
+
}
|
2426
|
+
|
2427
|
+
if (best_gain <= -HUGE_VAL) return best_gain;
|
2428
|
+
|
2429
|
+
real_t_ xtot = xmax - xmin;
|
2430
|
+
w_left = best_w;
|
2431
|
+
w_right = wtot - w_left;
|
2432
|
+
w_left = std::fmax(w_left, std::numeric_limits<double>::min());
|
2433
|
+
w_right = std::fmax(w_right, std::numeric_limits<double>::min());
|
2434
|
+
split_point = midpoint(x[buffer_indices[split_ix]], x[buffer_indices[split_ix+1]]);
|
2435
|
+
real_t_ rpct_left = split_point / xtot;
|
2436
|
+
rpct_left = std::fmax(rpct_left, std::numeric_limits<double>::min());
|
2437
|
+
real_t_ rpct_right = (real_t_)1 - rpct_left;
|
2438
|
+
rpct_right = std::fmax(rpct_right, std::numeric_limits<double>::min());
|
2439
|
+
|
2440
|
+
real_t_ wl_sq = w_left / wtot; wl_sq = square(wl_sq);
|
2441
|
+
real_t_ wr_sq = w_right / wtot; wl_sq = square(wr_sq);
|
2442
|
+
|
2443
|
+
return wl_sq / rpct_left + wr_sq / rpct_right;
|
2444
|
+
}
|
2445
|
+
|
2446
|
+
template <class real_t, class mapping, class ldouble_safe>
|
2447
|
+
double find_split_dens_shortform_weighted(real_t *restrict x, size_t n, double &restrict split_point, mapping &restrict w, size_t *restrict buffer_indices)
|
2448
|
+
{
|
2449
|
+
if (n < INT32_MAX)
|
2450
|
+
return find_split_dens_shortform_weighted_t<double, real_t, mapping>(x, n, split_point, w, buffer_indices);
|
2451
|
+
else
|
2452
|
+
return find_split_dens_shortform_weighted_t<ldouble_safe, real_t, mapping>(x, n, split_point, w, buffer_indices);
|
2453
|
+
}
|
2454
|
+
|
2455
|
+
template <class real_t>
|
2456
|
+
double find_split_dens_shortform(real_t *restrict x, size_t *restrict ix_arr, size_t st, size_t end,
|
2457
|
+
double &restrict split_point, size_t &restrict split_ix)
|
2458
|
+
{
|
2459
|
+
double best_gain = -HUGE_VAL;
|
2460
|
+
real_t xmin = x[ix_arr[st]];
|
2461
|
+
real_t xmax = x[ix_arr[end]];
|
2462
|
+
real_t xleft, xright;
|
2463
|
+
real_t xmid;
|
2464
|
+
double this_gain;
|
2465
|
+
|
2466
|
+
for (size_t row = st; row < end; row++)
|
2467
|
+
{
|
2468
|
+
if (x[ix_arr[row]] == x[ix_arr[row+1]]) continue;
|
2469
|
+
xmid = x[ix_arr[row]] + (x[ix_arr[row+1]] - x[ix_arr[row]]) / (real_t)2;
|
2470
|
+
xleft = xmid - xmin;
|
2471
|
+
xright = xmax - xmid;
|
2472
|
+
if (unlikely(!xleft || !xright)) continue;
|
2473
|
+
this_gain = square(row-st+1) / xleft + square(end-row) / xright;
|
2474
|
+
if (this_gain > best_gain)
|
2475
|
+
{
|
2476
|
+
best_gain = this_gain;
|
2477
|
+
split_ix = row;
|
2478
|
+
}
|
2479
|
+
}
|
2480
|
+
|
2481
|
+
if (best_gain <= -HUGE_VAL) return best_gain;
|
2482
|
+
|
2483
|
+
double xtot = (double)xmax - (double)xmin;
|
2484
|
+
double nleft = (double)(split_ix-st+1);
|
2485
|
+
double nright = (double)(end - split_ix);
|
2486
|
+
split_point = midpoint(x[ix_arr[split_ix]], x[ix_arr[split_ix+1]]);
|
2487
|
+
double rpct_left = split_point / xtot;
|
2488
|
+
rpct_left = std::fmax(rpct_left, std::numeric_limits<double>::min());
|
2489
|
+
double rpct_right = 1. - rpct_left;
|
2490
|
+
rpct_right = std::fmax(rpct_right, std::numeric_limits<double>::min());
|
2491
|
+
double ntot = (double)(end - st + 1);
|
2492
|
+
|
2493
|
+
double nl_sq = nleft / ntot; nl_sq = square(nl_sq);
|
2494
|
+
double nr_sq = nright / ntot; nl_sq = square(nr_sq);
|
2495
|
+
|
2496
|
+
return nl_sq / rpct_left + nr_sq / rpct_right;
|
2497
|
+
}
|
2498
|
+
|
2499
|
+
template <class real_t, class mapping>
|
2500
|
+
double find_split_dens_shortform_weighted(real_t *restrict x, size_t *restrict ix_arr, size_t st, size_t end,
|
2501
|
+
double &restrict split_point, size_t &restrict split_ix, mapping &restrict w)
|
2502
|
+
{
|
2503
|
+
double best_gain = -HUGE_VAL;
|
2504
|
+
real_t xmin = x[ix_arr[st]];
|
2505
|
+
real_t xmax = x[ix_arr[end]];
|
2506
|
+
real_t xleft, xright;
|
2507
|
+
real_t xmid;
|
2508
|
+
double this_gain;
|
2509
|
+
|
2510
|
+
double wtot = 0;
|
2511
|
+
for (size_t row = st; row <= end; row++)
|
2512
|
+
wtot += w[ix_arr[row]];
|
2513
|
+
double w_left = 0;
|
2514
|
+
double w_right;
|
2515
|
+
double best_w = 0;
|
2516
|
+
|
2517
|
+
for (size_t row = st; row < end; row++)
|
2518
|
+
{
|
2519
|
+
w_left += w[ix_arr[row]];
|
2520
|
+
if (x[ix_arr[row]] == x[ix_arr[row+1]]) continue;
|
2521
|
+
xmid = x[ix_arr[row]] + (x[ix_arr[row+1]] - x[ix_arr[row]]) / (real_t)2;
|
2522
|
+
xleft = xmid - xmin;
|
2523
|
+
xright = xmax - xmid;
|
2524
|
+
if (unlikely(!xleft || !xright)) continue;
|
2525
|
+
|
2526
|
+
w_right = wtot - w_left;
|
2527
|
+
this_gain = square(w_left) / xleft + square(w_right) / xright;
|
2528
|
+
if (this_gain > best_gain)
|
2529
|
+
{
|
2530
|
+
best_gain = this_gain;
|
2531
|
+
best_w = w_left;
|
2532
|
+
split_ix = row;
|
2533
|
+
}
|
2534
|
+
}
|
2535
|
+
|
2536
|
+
if (best_gain <= -HUGE_VAL) return best_gain;
|
2537
|
+
|
2538
|
+
double xtot = (double)xmax - (double)xmin;
|
2539
|
+
w_left = best_w;
|
2540
|
+
w_right = wtot - w_left;
|
2541
|
+
w_left = std::fmax(w_left, std::numeric_limits<double>::min());
|
2542
|
+
w_right = std::fmax(w_right, std::numeric_limits<double>::min());
|
2543
|
+
split_point = midpoint(x[split_ix], x[split_ix+1]);
|
2544
|
+
double rpct_left = split_point / xtot;
|
2545
|
+
rpct_left = std::fmax(rpct_left, std::numeric_limits<double>::min());
|
2546
|
+
double rpct_right = 1. - rpct_left;
|
2547
|
+
rpct_right = std::fmax(rpct_right, std::numeric_limits<double>::min());
|
2548
|
+
|
2549
|
+
double wl_sq = w_left / wtot; wl_sq = square(wl_sq);
|
2550
|
+
double wr_sq = w_right / wtot; wl_sq = square(wr_sq);
|
2551
|
+
|
2552
|
+
return wl_sq / rpct_left + wr_sq / rpct_right;
|
2553
|
+
}
|
2554
|
+
|
2555
|
+
/* This is a slower but more numerically-robust form */
|
2556
|
+
template <class real_t, class ldouble_safe>
|
2557
|
+
double find_split_dens_longform(real_t *restrict x, size_t *restrict ix_arr, size_t st, size_t end,
|
2558
|
+
double &restrict split_point, size_t &restrict split_ix)
|
2559
|
+
{
|
2560
|
+
double best_gain = -HUGE_VAL;
|
2561
|
+
real_t xmin = x[ix_arr[st]];
|
2562
|
+
real_t xmax = x[ix_arr[end]];
|
2563
|
+
real_t xleft, xright;
|
2564
|
+
real_t xmid;
|
2565
|
+
ldouble_safe pct_left, pct_right;
|
2566
|
+
ldouble_safe rpct_left, rpct_right;
|
2567
|
+
ldouble_safe n_tot = end - st + 1;
|
2568
|
+
ldouble_safe xtot = (ldouble_safe)xmax - (ldouble_safe)xmin;
|
2569
|
+
ldouble_safe cnt_left;
|
2570
|
+
double this_gain;
|
2571
|
+
|
2572
|
+
for (size_t row = st; row < end; row++)
|
2573
|
+
{
|
2574
|
+
if (x[ix_arr[row]] == x[ix_arr[row+1]]) continue;
|
2575
|
+
xmid = midpoint(x[ix_arr[row]], x[ix_arr[row+1]]);
|
2576
|
+
xleft = xmid - xmin;
|
2577
|
+
xright = xmax - xmid;
|
2578
|
+
if (unlikely(!xleft || !xright)) continue;
|
2579
|
+
|
2580
|
+
cnt_left = (ldouble_safe)(row-st+1);
|
2581
|
+
|
2582
|
+
xleft = std::fmax(xleft, (real_t)std::numeric_limits<real_t>::min());
|
2583
|
+
xright = std::fmax(xright, (real_t)std::numeric_limits<real_t>::min());
|
2584
|
+
pct_left = cnt_left / n_tot;
|
2585
|
+
pct_right = (ldouble_safe)1 - pct_left;
|
2586
|
+
rpct_left = (ldouble_safe)xleft / xtot;
|
2587
|
+
rpct_right = (ldouble_safe)xright / xtot;
|
2588
|
+
|
2589
|
+
this_gain = square(pct_left) / rpct_left + square(pct_right) / rpct_right;
|
2590
|
+
if (unlikely(is_na_or_inf(this_gain))) continue;
|
2591
|
+
if (this_gain > best_gain)
|
2592
|
+
{
|
2593
|
+
best_gain = this_gain;
|
2594
|
+
split_point = xmid;
|
2595
|
+
split_ix = row;
|
2596
|
+
}
|
2597
|
+
}
|
2598
|
+
|
2599
|
+
return best_gain;
|
2600
|
+
}
|
2601
|
+
|
2602
|
+
template <class real_t, class mapping, class ldouble_safe>
|
2603
|
+
double find_split_dens_longform_weighted(real_t *restrict x, size_t *restrict ix_arr, size_t st, size_t end,
|
2604
|
+
double &restrict split_point, size_t &restrict split_ix, mapping &restrict w)
|
2605
|
+
{
|
2606
|
+
double best_gain = -HUGE_VAL;
|
2607
|
+
real_t xmin = x[ix_arr[st]];
|
2608
|
+
real_t xmax = x[ix_arr[end]];
|
2609
|
+
real_t xleft, xright;
|
2610
|
+
real_t xmid;
|
2611
|
+
ldouble_safe pct_left, pct_right;
|
2612
|
+
ldouble_safe rpct_left, rpct_right;
|
2613
|
+
ldouble_safe xtot = (ldouble_safe)xmax - (ldouble_safe)xmin;
|
2614
|
+
double this_gain;
|
2615
|
+
|
2616
|
+
ldouble_safe wtot = 0;
|
2617
|
+
for (size_t row = st; row <= end; row++)
|
2618
|
+
wtot += w[ix_arr[row]];
|
2619
|
+
ldouble_safe w_left = 0;
|
2620
|
+
|
2621
|
+
for (size_t row = st; row < end; row++)
|
2622
|
+
{
|
2623
|
+
w_left += w[ix_arr[row]];
|
2624
|
+
if (x[ix_arr[row]] == x[ix_arr[row+1]]) continue;
|
2625
|
+
xmid = midpoint(x[ix_arr[row]], x[ix_arr[row+1]]);
|
2626
|
+
xleft = xmid - xmin;
|
2627
|
+
xright = xmax - xmid;
|
2628
|
+
if (unlikely(!xleft || !xright)) continue;
|
2629
|
+
|
2630
|
+
xleft = std::fmax(xleft, (real_t)std::numeric_limits<real_t>::min());
|
2631
|
+
xright = std::fmax(xright, (real_t)std::numeric_limits<real_t>::min());
|
2632
|
+
pct_left = w_left / wtot;
|
2633
|
+
pct_right = (ldouble_safe)1 - pct_left;
|
2634
|
+
rpct_left = (ldouble_safe)xleft / xtot;
|
2635
|
+
rpct_right = (ldouble_safe)xright / xtot;
|
2636
|
+
|
2637
|
+
this_gain = square(pct_left) / rpct_left + square(pct_right) / rpct_right;
|
2638
|
+
if (unlikely(is_na_or_inf(this_gain))) continue;
|
2639
|
+
if (this_gain > best_gain)
|
2640
|
+
{
|
2641
|
+
best_gain = this_gain;
|
2642
|
+
split_point = xmid;
|
2643
|
+
split_ix = row;
|
2644
|
+
}
|
2645
|
+
}
|
2646
|
+
|
2647
|
+
return best_gain;
|
2648
|
+
}
|
2649
|
+
|
2650
|
+
template <class real_t, class ldouble_safe>
|
2651
|
+
double find_split_dens(real_t *restrict x, size_t *restrict ix_arr, size_t st, size_t end,
|
2652
|
+
double &restrict split_point, size_t &restrict split_ix)
|
2653
|
+
{
|
2654
|
+
if (end - st + 1 < INT32_MAX && x[ix_arr[end]] - x[ix_arr[st]] >= 1)
|
2655
|
+
return find_split_dens_shortform<real_t>(x, ix_arr, st, end, split_point, split_ix);
|
2656
|
+
else
|
2657
|
+
return find_split_dens_longform<real_t, ldouble_safe>(x, ix_arr, st, end, split_point, split_ix);
|
2658
|
+
}
|
2659
|
+
|
2660
|
+
template <class real_t, class mapping, class ldouble_safe>
|
2661
|
+
double find_split_dens_weighted(real_t *restrict x, size_t *restrict ix_arr, size_t st, size_t end,
|
2662
|
+
double &restrict split_point, size_t &restrict split_ix, mapping &restrict w)
|
2663
|
+
{
|
2664
|
+
if (end - st + 1 < INT32_MAX && x[ix_arr[end]] - x[ix_arr[st]] >= 1)
|
2665
|
+
return find_split_dens_shortform_weighted<real_t, mapping>(x, ix_arr, st, end, split_point, split_ix, w);
|
2666
|
+
else
|
2667
|
+
return find_split_dens_longform_weighted<real_t, mapping, ldouble_safe>(x, ix_arr, st, end, split_point, split_ix, w);
|
2668
|
+
}
|
2669
|
+
|
2670
|
+
template <class int_t, class ldouble_safe>
|
2671
|
+
double find_split_dens_longform(int *restrict x, int ncat, size_t *restrict ix_arr, size_t st, size_t end,
|
2672
|
+
CategSplit cat_split_type, MissingAction missing_action,
|
2673
|
+
int &restrict chosen_cat, signed char *restrict split_categ, int *restrict saved_cat_mode,
|
2674
|
+
size_t *restrict buffer_cnt, int_t *restrict buffer_indices)
|
2675
|
+
{
|
2676
|
+
if (st >= end || ncat <= 1) return -HUGE_VAL;
|
2677
|
+
size_t n_nas = 0;
|
2678
|
+
int xval;
|
2679
|
+
|
2680
|
+
/* count categories */
|
2681
|
+
memset(buffer_cnt, 0, sizeof(size_t) * ncat);
|
2682
|
+
if (missing_action == Fail)
|
2683
|
+
{
|
2684
|
+
for (size_t row = st; row <= end; row++)
|
2685
|
+
if (likely(x[ix_arr[row]] >= 0))
|
2686
|
+
buffer_cnt[x[ix_arr[row]]]++;
|
2687
|
+
}
|
2688
|
+
|
2689
|
+
else if (missing_action == Impute)
|
2690
|
+
{
|
2691
|
+
for (size_t row = st; row <= end; row++)
|
2692
|
+
{
|
2693
|
+
xval = x[ix_arr[row]];
|
2694
|
+
if (unlikely(xval < 0))
|
2695
|
+
n_nas++;
|
2696
|
+
else
|
2697
|
+
buffer_cnt[xval]++;
|
2698
|
+
}
|
2699
|
+
|
2700
|
+
if (unlikely(n_nas >= end-st)) return -HUGE_VAL;
|
2701
|
+
|
2702
|
+
if (n_nas)
|
2703
|
+
{
|
2704
|
+
auto idxmax = std::max_element(buffer_cnt, buffer_cnt + ncat);
|
2705
|
+
*idxmax += n_nas;
|
2706
|
+
*saved_cat_mode = (int)std::distance(buffer_cnt, idxmax);
|
2707
|
+
}
|
2708
|
+
}
|
2709
|
+
|
2710
|
+
else
|
2711
|
+
{
|
2712
|
+
for (size_t row = st; row <= end; row++)
|
2713
|
+
{
|
2714
|
+
xval = x[ix_arr[row]];
|
2715
|
+
if (likely(xval >= 0)) buffer_cnt[xval]++;
|
2716
|
+
}
|
2717
|
+
}
|
2718
|
+
|
2719
|
+
std::iota(buffer_indices, buffer_indices + ncat, (int_t)0);
|
2720
|
+
std::sort(buffer_indices, buffer_indices + ncat,
|
2721
|
+
[&buffer_cnt](const int_t a, const int_t b)
|
2722
|
+
{return buffer_cnt[a] < buffer_cnt[b];});
|
2723
|
+
|
2724
|
+
int curr = 0;
|
2725
|
+
if (split_categ != NULL)
|
2726
|
+
{
|
2727
|
+
while (buffer_cnt[buffer_indices[curr]] == 0)
|
2728
|
+
{
|
2729
|
+
split_categ[buffer_indices[curr]] = -1;
|
2730
|
+
curr++;
|
2731
|
+
}
|
2732
|
+
}
|
2733
|
+
|
2734
|
+
else
|
2735
|
+
{
|
2736
|
+
while (buffer_cnt[buffer_indices[curr]] == 0) curr++;
|
2737
|
+
}
|
2738
|
+
|
2739
|
+
int ncat_present = ncat - curr;
|
2740
|
+
if (ncat_present <= 1) return -HUGE_VAL;
|
2741
|
+
if (ncat_present == 2)
|
2742
|
+
{
|
2743
|
+
switch (cat_split_type)
|
2744
|
+
{
|
2745
|
+
case SingleCateg:
|
2746
|
+
{
|
2747
|
+
chosen_cat = buffer_indices[curr];
|
2748
|
+
break;
|
2749
|
+
}
|
2750
|
+
|
2751
|
+
case SubSet:
|
2752
|
+
{
|
2753
|
+
split_categ[buffer_indices[curr]] = 1;
|
2754
|
+
split_categ[buffer_indices[curr+1]] = 0;
|
2755
|
+
break;
|
2756
|
+
}
|
2757
|
+
}
|
2758
|
+
|
2759
|
+
ldouble_safe pct_left
|
2760
|
+
=
|
2761
|
+
(ldouble_safe)buffer_cnt[buffer_indices[curr]]
|
2762
|
+
/
|
2763
|
+
(ldouble_safe)(
|
2764
|
+
buffer_cnt[buffer_indices[curr]]
|
2765
|
+
+
|
2766
|
+
buffer_cnt[buffer_indices[curr+1]]
|
2767
|
+
);
|
2768
|
+
|
2769
|
+
return ((ldouble_safe)buffer_cnt[buffer_indices[curr]] * (2. * pct_left)
|
2770
|
+
+
|
2771
|
+
(ldouble_safe)buffer_cnt[buffer_indices[curr+1]] * (2. - 2.*pct_left))
|
2772
|
+
/
|
2773
|
+
(ldouble_safe)(buffer_cnt[buffer_indices[curr]] + buffer_cnt[buffer_indices[curr+1]]);
|
2774
|
+
}
|
2775
|
+
|
2776
|
+
size_t ntot;
|
2777
|
+
if (missing_action == Impute)
|
2778
|
+
ntot = end - st + 1;
|
2779
|
+
else
|
2780
|
+
ntot = std::accumulate(buffer_cnt, buffer_cnt + ncat, (size_t)0);
|
2781
|
+
if (unlikely(ntot <= 1)) unexpected_error();
|
2782
|
+
ldouble_safe ntot_ = (ldouble_safe)ntot;
|
2783
|
+
|
2784
|
+
switch (cat_split_type)
|
2785
|
+
{
|
2786
|
+
case SingleCateg:
|
2787
|
+
{
|
2788
|
+
double pct_one_cat = 1. / (double)ncat_present;
|
2789
|
+
double pct_left_smallest = (ldouble_safe)buffer_cnt[buffer_indices[curr]] / ntot_;
|
2790
|
+
double gain_smallest
|
2791
|
+
=
|
2792
|
+
(ldouble_safe)buffer_cnt[buffer_indices[curr]] * (pct_left_smallest / pct_one_cat)
|
2793
|
+
+
|
2794
|
+
(ldouble_safe)(ntot - buffer_cnt[buffer_indices[curr]]) * ((1. - pct_left_smallest) / (1. - pct_one_cat))
|
2795
|
+
;
|
2796
|
+
|
2797
|
+
double pct_left_biggest = (ldouble_safe)buffer_cnt[buffer_indices[ncat-1]] / ntot_;
|
2798
|
+
double gain_biggest
|
2799
|
+
=
|
2800
|
+
(ldouble_safe)buffer_cnt[buffer_indices[ncat-1]] * (pct_left_biggest / pct_one_cat)
|
2801
|
+
+
|
2802
|
+
(ldouble_safe)(ntot - buffer_cnt[buffer_indices[ncat-1]]) * ((1. - pct_left_biggest) / (1. - pct_one_cat))
|
2803
|
+
;
|
2804
|
+
|
2805
|
+
if (gain_smallest >= gain_biggest)
|
2806
|
+
{
|
2807
|
+
chosen_cat = buffer_indices[curr];
|
2808
|
+
return gain_smallest / ntot_;
|
2809
|
+
}
|
2810
|
+
|
2811
|
+
else
|
2812
|
+
{
|
2813
|
+
chosen_cat = buffer_indices[ncat-1];
|
2814
|
+
return gain_biggest / ntot_;
|
2815
|
+
}
|
2816
|
+
break;
|
2817
|
+
}
|
2818
|
+
|
2819
|
+
case SubSet:
|
2820
|
+
{
|
2821
|
+
size_t cnt_left = 0;
|
2822
|
+
size_t cnt_right;
|
2823
|
+
int st_cat = curr - 1;
|
2824
|
+
double this_gain;
|
2825
|
+
double best_gain = -HUGE_VAL;
|
2826
|
+
int best_cat = 0;
|
2827
|
+
ldouble_safe pct_left;
|
2828
|
+
double pct_cat_left;
|
2829
|
+
double ncat_present_ = (double)ncat_present;
|
2830
|
+
for (; curr < ncat; curr++)
|
2831
|
+
{
|
2832
|
+
cnt_left += buffer_cnt[buffer_indices[curr]];
|
2833
|
+
cnt_right = ntot - cnt_left;
|
2834
|
+
pct_left = (ldouble_safe)cnt_left / ntot_;
|
2835
|
+
pct_cat_left = (double)(curr - st_cat) / ncat_present_;
|
2836
|
+
this_gain
|
2837
|
+
=
|
2838
|
+
(ldouble_safe)cnt_left * (pct_left / pct_cat_left)
|
2839
|
+
+
|
2840
|
+
(ldouble_safe)cnt_right * (((ldouble_safe)1 - pct_left) / (1. - pct_cat_left))
|
2841
|
+
;
|
2842
|
+
if (this_gain > best_gain)
|
2843
|
+
{
|
2844
|
+
best_gain = this_gain;
|
2845
|
+
best_cat = curr;
|
2846
|
+
}
|
2847
|
+
}
|
2848
|
+
|
2849
|
+
if (best_gain <= -HUGE_VAL) return best_gain;
|
2850
|
+
st_cat++;
|
2851
|
+
for (; st_cat <= best_cat; st_cat++)
|
2852
|
+
split_categ[buffer_indices[st_cat]] = 1;
|
2853
|
+
for (; st_cat < ncat; st_cat++)
|
2854
|
+
split_categ[buffer_indices[st_cat]] = 0;
|
2855
|
+
return best_gain / ntot_;
|
2856
|
+
break;
|
2857
|
+
}
|
2858
|
+
}
|
2859
|
+
|
2860
|
+
/* This will not be reached, but CRAN might complain otherwise */
|
2861
|
+
return -HUGE_VAL;
|
2862
|
+
}
|
2863
|
+
|
2864
|
+
template <class mapping, class int_t, class ldouble_safe>
|
2865
|
+
double find_split_dens_longform_weighted(int *restrict x, int ncat, size_t *restrict ix_arr, size_t st, size_t end,
|
2866
|
+
CategSplit cat_split_type, MissingAction missing_action,
|
2867
|
+
int &restrict chosen_cat, signed char *restrict split_categ, int *restrict saved_cat_mode,
|
2868
|
+
int_t *restrict buffer_indices, mapping &restrict w)
|
2869
|
+
{
|
2870
|
+
if (st >= end || ncat <= 1) return -HUGE_VAL;
|
2871
|
+
ldouble_safe w_missing = 0;
|
2872
|
+
int xval;
|
2873
|
+
size_t ix_;
|
2874
|
+
|
2875
|
+
/* count categories */
|
2876
|
+
/* TODO: allocate this buffer externally */
|
2877
|
+
std::vector<ldouble_safe> buffer_cnt(ncat, (ldouble_safe)0);
|
2878
|
+
if (missing_action == Fail)
|
2879
|
+
{
|
2880
|
+
for (size_t row = st; row <= end; row++)
|
2881
|
+
{
|
2882
|
+
ix_ = ix_arr[row];
|
2883
|
+
if (unlikely(x[ix_]) < 0) continue;
|
2884
|
+
buffer_cnt[x[ix_]] += w[ix_];
|
2885
|
+
}
|
2886
|
+
}
|
2887
|
+
|
2888
|
+
else if (missing_action == Impute)
|
2889
|
+
{
|
2890
|
+
for (size_t row = st; row <= end; row++)
|
2891
|
+
{
|
2892
|
+
ix_ = ix_arr[row];
|
2893
|
+
xval = x[ix_];
|
2894
|
+
if (unlikely(xval < 0))
|
2895
|
+
w_missing += w[ix_];
|
2896
|
+
else
|
2897
|
+
buffer_cnt[xval] += w[ix_];
|
2898
|
+
}
|
2899
|
+
|
2900
|
+
if (w_missing)
|
2901
|
+
{
|
2902
|
+
auto idxmax = std::max_element(buffer_cnt.begin(), buffer_cnt.end());
|
2903
|
+
*idxmax += w_missing;
|
2904
|
+
*saved_cat_mode = (int)std::distance(buffer_cnt.begin(), idxmax);
|
2905
|
+
}
|
2906
|
+
}
|
2907
|
+
|
2908
|
+
else
|
2909
|
+
{
|
2910
|
+
for (size_t row = st; row <= end; row++)
|
2911
|
+
{
|
2912
|
+
ix_ = ix_arr[row];
|
2913
|
+
xval = x[ix_];
|
2914
|
+
if (likely(xval >= 0)) buffer_cnt[xval] += w[ix_];
|
2915
|
+
}
|
2916
|
+
}
|
2917
|
+
|
2918
|
+
std::iota(buffer_indices, buffer_indices + ncat, (int_t)0);
|
2919
|
+
std::sort(buffer_indices, buffer_indices + ncat,
|
2920
|
+
[&buffer_cnt](const int_t a, const int_t b)
|
2921
|
+
{return buffer_cnt[a] < buffer_cnt[b];});
|
2922
|
+
|
2923
|
+
int curr = 0;
|
2924
|
+
if (split_categ != NULL)
|
2925
|
+
{
|
2926
|
+
while (buffer_cnt[buffer_indices[curr]] == 0)
|
2927
|
+
{
|
2928
|
+
split_categ[buffer_indices[curr]] = -1;
|
2929
|
+
curr++;
|
2930
|
+
}
|
2931
|
+
}
|
2932
|
+
|
2933
|
+
else
|
2934
|
+
{
|
2935
|
+
while (buffer_cnt[buffer_indices[curr]] == 0) curr++;
|
2936
|
+
}
|
2937
|
+
|
2938
|
+
int ncat_present = ncat - curr;
|
2939
|
+
if (ncat_present <= 1) return -HUGE_VAL;
|
2940
|
+
if (ncat_present == 2)
|
2941
|
+
{
|
2942
|
+
switch (cat_split_type)
|
2943
|
+
{
|
2944
|
+
case SingleCateg:
|
2945
|
+
{
|
2946
|
+
chosen_cat = buffer_indices[curr];
|
2947
|
+
break;
|
2948
|
+
}
|
2949
|
+
|
2950
|
+
case SubSet:
|
2951
|
+
{
|
2952
|
+
split_categ[buffer_indices[curr]] = 1;
|
2953
|
+
split_categ[buffer_indices[curr+1]] = 0;
|
2954
|
+
break;
|
2955
|
+
}
|
2956
|
+
}
|
2957
|
+
|
2958
|
+
ldouble_safe pct_left
|
2959
|
+
=
|
2960
|
+
buffer_cnt[buffer_indices[curr]]
|
2961
|
+
/
|
2962
|
+
(
|
2963
|
+
buffer_cnt[buffer_indices[curr]]
|
2964
|
+
+
|
2965
|
+
buffer_cnt[buffer_indices[curr+1]]
|
2966
|
+
);
|
2967
|
+
|
2968
|
+
return (buffer_cnt[buffer_indices[curr]] * (pct_left * 2.)
|
2969
|
+
+
|
2970
|
+
buffer_cnt[buffer_indices[curr+1]] * (2. - 2.*pct_left))
|
2971
|
+
/
|
2972
|
+
(buffer_cnt[buffer_indices[curr]] + buffer_cnt[buffer_indices[curr+1]]);
|
2973
|
+
}
|
2974
|
+
|
2975
|
+
ldouble_safe ntot = std::accumulate(buffer_cnt.begin(), buffer_cnt.end(), (ldouble_safe)0);
|
2976
|
+
if (unlikely(ntot <= 0)) unexpected_error();
|
2977
|
+
|
2978
|
+
switch (cat_split_type)
|
2979
|
+
{
|
2980
|
+
case SingleCateg:
|
2981
|
+
{
|
2982
|
+
double pct_one_cat = 1. / (double)ncat_present;
|
2983
|
+
double pct_left_smallest = buffer_cnt[buffer_indices[curr]] / ntot;
|
2984
|
+
double gain_smallest
|
2985
|
+
=
|
2986
|
+
buffer_cnt[buffer_indices[curr]] * (pct_left_smallest / pct_one_cat)
|
2987
|
+
+
|
2988
|
+
(ntot - buffer_cnt[buffer_indices[curr]]) * ((1. - pct_left_smallest) / (1. - pct_one_cat))
|
2989
|
+
;
|
2990
|
+
|
2991
|
+
double pct_left_biggest = buffer_cnt[buffer_indices[ncat-1]] / ntot;
|
2992
|
+
double gain_biggest
|
2993
|
+
=
|
2994
|
+
buffer_cnt[buffer_indices[ncat-1]] * (pct_left_biggest / pct_one_cat)
|
2995
|
+
+
|
2996
|
+
(ntot - buffer_cnt[buffer_indices[ncat-1]]) * ((1. - pct_left_biggest) / (1. - pct_one_cat))
|
2997
|
+
;
|
2998
|
+
|
2999
|
+
if (gain_smallest >= gain_biggest)
|
3000
|
+
{
|
3001
|
+
chosen_cat = buffer_indices[curr];
|
3002
|
+
return gain_smallest / ntot;
|
3003
|
+
}
|
3004
|
+
|
3005
|
+
else
|
3006
|
+
{
|
3007
|
+
chosen_cat = buffer_indices[ncat-1];
|
3008
|
+
return gain_biggest / ntot;
|
3009
|
+
}
|
3010
|
+
break;
|
3011
|
+
}
|
3012
|
+
|
3013
|
+
case SubSet:
|
3014
|
+
{
|
3015
|
+
ldouble_safe cnt_left = 0;
|
3016
|
+
ldouble_safe cnt_right;
|
3017
|
+
int st_cat = curr - 1;
|
3018
|
+
double this_gain;
|
3019
|
+
double best_gain = -HUGE_VAL;
|
3020
|
+
int best_cat = 0;
|
3021
|
+
ldouble_safe pct_left;
|
3022
|
+
double pct_cat_left;
|
3023
|
+
double ncat_present_ = (double)ncat_present;
|
3024
|
+
for (; curr < ncat; curr++)
|
3025
|
+
{
|
3026
|
+
cnt_left += buffer_cnt[buffer_indices[curr]];
|
3027
|
+
cnt_right = ntot - cnt_left;
|
3028
|
+
pct_left = cnt_left / ntot;
|
3029
|
+
pct_cat_left = (double)(curr - st_cat) / ncat_present_;
|
3030
|
+
this_gain
|
3031
|
+
=
|
3032
|
+
(ldouble_safe)cnt_left * (pct_left / pct_cat_left)
|
3033
|
+
+
|
3034
|
+
(ldouble_safe)cnt_right * (((ldouble_safe)1 - pct_left) / (1. - pct_cat_left))
|
3035
|
+
;
|
3036
|
+
if (this_gain > best_gain)
|
3037
|
+
{
|
3038
|
+
best_gain = this_gain;
|
3039
|
+
best_cat = curr;
|
3040
|
+
}
|
3041
|
+
}
|
3042
|
+
|
3043
|
+
if (best_gain <= -HUGE_VAL) return best_gain;
|
3044
|
+
st_cat++;
|
3045
|
+
for (; st_cat <= best_cat; st_cat++)
|
3046
|
+
split_categ[buffer_indices[st_cat]] = 1;
|
3047
|
+
for (; st_cat < ncat; st_cat++)
|
3048
|
+
split_categ[buffer_indices[st_cat]] = 0;
|
3049
|
+
return best_gain / ntot;
|
3050
|
+
break;
|
3051
|
+
}
|
3052
|
+
}
|
3053
|
+
|
3054
|
+
/* This will not be reached, but CRAN might complain otherwise */
|
3055
|
+
return -HUGE_VAL;
|
3056
|
+
}
|
3057
|
+
|
3058
|
+
/* for split-criterion in hyperplanes (see below for version aimed at single-variable splits) */
|
3059
|
+
template <class ldouble_safe>
|
3060
|
+
double eval_guided_crit(double *restrict x, size_t n, GainCriterion criterion,
|
3061
|
+
double min_gain, bool as_relative_gain, double *restrict buffer_sd,
|
3062
|
+
double &restrict split_point, double &restrict xmin, double &restrict xmax,
|
3063
|
+
size_t *restrict ix_arr_plus_st,
|
3064
|
+
size_t *restrict cols_use, size_t ncols_use, bool force_cols_use,
|
3065
|
+
double *restrict X_row_major, size_t ncols,
|
3066
|
+
double *restrict Xr, size_t *restrict Xr_ind, size_t *restrict Xr_indptr)
|
3067
|
+
{
|
3068
|
+
/* Note: the input 'x' is supposed to be a linear combination of standardized variables, so
|
3069
|
+
all numbers are assumed to be small and in the same scale */
|
3070
|
+
double gain = 0;
|
3071
|
+
if (criterion == DensityCrit || criterion == FullGain) min_gain = 0;
|
3072
|
+
|
3073
|
+
/* here it's assumed the 'x' vector matches exactly with 'ix_arr' + 'st' */
|
3074
|
+
if (unlikely(n == 2))
|
3075
|
+
{
|
3076
|
+
if (x[0] == x[1]) return -HUGE_VAL;
|
3077
|
+
split_point = midpoint_with_reorder(x[0], x[1]);
|
3078
|
+
gain = 1.;
|
3079
|
+
if (gain > min_gain)
|
3080
|
+
return gain;
|
3081
|
+
else
|
3082
|
+
return 0.;
|
3083
|
+
}
|
3084
|
+
|
3085
|
+
if (criterion == FullGain)
|
3086
|
+
{
|
3087
|
+
/* TODO: these buffers should be allocated externally */
|
3088
|
+
std::vector<size_t> argsorted(n);
|
3089
|
+
std::iota(argsorted.begin(), argsorted.end(), (size_t)0);
|
3090
|
+
std::sort(argsorted.begin(), argsorted.end(),
|
3091
|
+
[&x](const size_t a, const size_t b){return x[a] < x[b];});
|
3092
|
+
if (x[argsorted[0]] == x[argsorted[n-1]]) return -HUGE_VAL;
|
3093
|
+
std::vector<double> temp_buffer(n + mult2(ncols));
|
3094
|
+
for (size_t ix = 0; ix < n; ix++) temp_buffer[ix] = x[argsorted[ix]];
|
3095
|
+
for (size_t ix = 0; ix < n; ix++)
|
3096
|
+
argsorted[ix] = ix_arr_plus_st[argsorted[ix]];
|
3097
|
+
size_t ignored;
|
3098
|
+
return find_split_full_gain<double, ldouble_safe>(
|
3099
|
+
temp_buffer.data(), (size_t)0, n-1, argsorted.data(),
|
3100
|
+
cols_use, ncols_use, force_cols_use,
|
3101
|
+
X_row_major, ncols,
|
3102
|
+
Xr, Xr_ind, Xr_indptr,
|
3103
|
+
temp_buffer.data() + n, temp_buffer.data() + n + ncols,
|
3104
|
+
ignored, split_point,
|
3105
|
+
false);
|
3106
|
+
}
|
3107
|
+
|
3108
|
+
/* sort in ascending order */
|
3109
|
+
std::sort(x, x + n);
|
3110
|
+
xmin = x[0]; xmax = x[n-1];
|
3111
|
+
if (x[0] == x[n-1]) return -HUGE_VAL;
|
3112
|
+
|
3113
|
+
if (criterion == Pooled && as_relative_gain && min_gain <= 0)
|
3114
|
+
gain = find_split_rel_gain<double, ldouble_safe>(x, n, split_point);
|
3115
|
+
else if (criterion == Pooled || criterion == Averaged)
|
3116
|
+
gain = find_split_std_gain<double, ldouble_safe>(x, n, buffer_sd, criterion, min_gain, split_point);
|
3117
|
+
else if (criterion == DensityCrit)
|
3118
|
+
gain = find_split_dens_shortform<double, ldouble_safe>(x, n, split_point);
|
3119
|
+
/* Note: a gain of -Inf signals that the data is unsplittable. Zero signals it's below the minimum. */
|
3120
|
+
return std::fmax(0., gain);
|
3121
|
+
}
|
3122
|
+
|
3123
|
+
template <class ldouble_safe>
|
3124
|
+
double eval_guided_crit_weighted(double *restrict x, size_t n, GainCriterion criterion,
|
3125
|
+
double min_gain, bool as_relative_gain, double *restrict buffer_sd,
|
3126
|
+
double &restrict split_point, double &restrict xmin, double &restrict xmax,
|
3127
|
+
double *restrict w, size_t *restrict buffer_indices,
|
3128
|
+
size_t *restrict ix_arr_plus_st,
|
3129
|
+
size_t *restrict cols_use, size_t ncols_use, bool force_cols_use,
|
3130
|
+
double *restrict X_row_major, size_t ncols,
|
3131
|
+
double *restrict Xr, size_t *restrict Xr_ind, size_t *restrict Xr_indptr)
|
3132
|
+
{
|
3133
|
+
/* Note: the input 'x' is supposed to be a linear combination of standardized variables, so
|
3134
|
+
all numbers are assumed to be small and in the same scale */
|
3135
|
+
double gain = 0;
|
3136
|
+
if (criterion == DensityCrit || criterion == FullGain) min_gain = 0;
|
3137
|
+
|
3138
|
+
/* here it's assumed the 'x' vector matches exactly with 'ix_arr' + 'st' */
|
3139
|
+
if (unlikely(n == 2))
|
3140
|
+
{
|
3141
|
+
if (x[0] == x[1]) return -HUGE_VAL;
|
3142
|
+
split_point = midpoint_with_reorder(x[0], x[1]);
|
3143
|
+
gain = 1.;
|
3144
|
+
if (gain > min_gain)
|
3145
|
+
return gain;
|
3146
|
+
else
|
3147
|
+
return 0.;
|
3148
|
+
}
|
3149
|
+
|
3150
|
+
/* sort in ascending order */
|
3151
|
+
std::iota(buffer_indices, buffer_indices + n, (size_t)0);
|
3152
|
+
std::sort(buffer_indices, buffer_indices + n,
|
3153
|
+
[&x](const size_t a, const size_t b){return x[a] < x[b];});
|
3154
|
+
xmin = x[buffer_indices[0]]; xmax = x[buffer_indices[n-1]];
|
3155
|
+
if (xmin == xmax) return -HUGE_VAL;
|
3156
|
+
|
3157
|
+
if (criterion == Pooled || criterion == Averaged)
|
3158
|
+
gain = find_split_std_gain_weighted<double, ldouble_safe>(x, n, buffer_sd, criterion, min_gain, split_point, w, buffer_indices);
|
3159
|
+
else if (criterion == DensityCrit)
|
3160
|
+
gain = find_split_dens_shortform_weighted<double, double *restrict, ldouble_safe>(x, n, split_point, w, buffer_indices);
|
3161
|
+
else if (criterion == FullGain)
|
3162
|
+
{
|
3163
|
+
std::vector<size_t> argsorted(n);
|
3164
|
+
std::iota(argsorted.begin(), argsorted.end(), (size_t)0);
|
3165
|
+
std::sort(argsorted.begin(), argsorted.end(),
|
3166
|
+
[&x](const size_t a, const size_t b){return x[a] < x[b];});
|
3167
|
+
if (x[argsorted[0]] == x[argsorted[n-1]]) return -HUGE_VAL;
|
3168
|
+
std::vector<double> temp_buffer(n + mult2(ncols));
|
3169
|
+
for (size_t ix = 0; ix < n; ix++) temp_buffer[ix] = x[argsorted[ix]];
|
3170
|
+
for (size_t ix = 0; ix < n; ix++)
|
3171
|
+
argsorted[ix] = ix_arr_plus_st[argsorted[ix]];
|
3172
|
+
size_t ignored;
|
3173
|
+
gain = find_split_full_gain_weighted<double, double *restrict, ldouble_safe>(
|
3174
|
+
temp_buffer.data(), (size_t)0, n-1, argsorted.data(),
|
3175
|
+
cols_use, ncols_use, force_cols_use,
|
3176
|
+
X_row_major, ncols,
|
3177
|
+
Xr, Xr_ind, Xr_indptr,
|
3178
|
+
temp_buffer.data() + n, temp_buffer.data() + n + ncols,
|
3179
|
+
ignored, split_point,
|
3180
|
+
false,
|
3181
|
+
w);
|
3182
|
+
}
|
3183
|
+
/* Note: a gain of -Inf signals that the data is unsplittable. Zero signals it's below the minimum. */
|
3184
|
+
return std::fmax(0., gain);
|
3185
|
+
}
|
3186
|
+
|
3187
|
+
/* for split-criterion in single-variable splits */
|
3188
|
+
template <class real_t_, class ldouble_safe>
|
3189
|
+
double eval_guided_crit(size_t *restrict ix_arr, size_t st, size_t end, real_t_ *restrict x,
|
3190
|
+
double *restrict buffer_sd, bool as_relative_gain,
|
3191
|
+
double *restrict buffer_imputed_x, double *restrict saved_xmedian,
|
3192
|
+
size_t &split_ix, double &restrict split_point, double &restrict xmin, double &restrict xmax,
|
3193
|
+
GainCriterion criterion, double min_gain, MissingAction missing_action,
|
3194
|
+
size_t *restrict cols_use, size_t ncols_use, bool force_cols_use,
|
3195
|
+
double *restrict X_row_major, size_t ncols,
|
3196
|
+
double *restrict Xr, size_t *restrict Xr_ind, size_t *restrict Xr_indptr)
|
3197
|
+
{
|
3198
|
+
size_t st_orig = st;
|
3199
|
+
double gain = 0;
|
3200
|
+
if (criterion == DensityCrit || criterion == FullGain) min_gain = 0;
|
3201
|
+
|
3202
|
+
/* move NAs to the front if there's any, exclude them from calculations */
|
3203
|
+
if (missing_action != Fail)
|
3204
|
+
st = move_NAs_to_front(ix_arr, st, end, x);
|
3205
|
+
|
3206
|
+
if (unlikely(st >= end)) return -HUGE_VAL;
|
3207
|
+
else if (unlikely(st == (end-1)))
|
3208
|
+
{
|
3209
|
+
if (x[ix_arr[st]] == x[ix_arr[end]])
|
3210
|
+
return -HUGE_VAL;
|
3211
|
+
split_point = midpoint_with_reorder(x[ix_arr[st]], x[ix_arr[end]]);
|
3212
|
+
split_ix = st;
|
3213
|
+
gain = 1.;
|
3214
|
+
if (gain > min_gain)
|
3215
|
+
return gain;
|
3216
|
+
else
|
3217
|
+
return 0.;
|
3218
|
+
}
|
3219
|
+
|
3220
|
+
/* sort in ascending order */
|
3221
|
+
std::sort(ix_arr + st, ix_arr + end + 1, [&x](const size_t a, const size_t b){return x[a] < x[b];});
|
3222
|
+
if (x[ix_arr[st]] == x[ix_arr[end]]) return -HUGE_VAL;
|
3223
|
+
xmin = x[ix_arr[st]]; xmax = x[ix_arr[end]];
|
3224
|
+
|
3225
|
+
/* unlike the previous case for the extended model, the data here has not been centered,
|
3226
|
+
which could make the standard deviations have poor precision. It's nevertheless not
|
3227
|
+
necessary for this mean to have good precision, since it's only meant for centering,
|
3228
|
+
so it can be calculated inexactly with simd instructions. */
|
3229
|
+
real_t_ xmean = 0;
|
3230
|
+
if (criterion == Pooled || criterion == Averaged)
|
3231
|
+
{
|
3232
|
+
for (size_t ix = st; ix <= end; ix++)
|
3233
|
+
xmean += x[ix_arr[ix]];
|
3234
|
+
xmean /= (real_t_)(end - st + 1);
|
3235
|
+
}
|
3236
|
+
|
3237
|
+
if (missing_action == Impute && st > st_orig)
|
3238
|
+
{
|
3239
|
+
missing_action = Fail;
|
3240
|
+
fill_NAs_with_median(ix_arr, st_orig, st, end, x, buffer_imputed_x, saved_xmedian);
|
3241
|
+
if (criterion == Pooled && as_relative_gain && min_gain <= 0)
|
3242
|
+
gain = find_split_rel_gain<double, ldouble_safe>(buffer_imputed_x, (double)xmean, ix_arr, st_orig, end, split_point, split_ix);
|
3243
|
+
else if (criterion == Pooled || criterion == Averaged)
|
3244
|
+
gain = find_split_std_gain<double, ldouble_safe>(buffer_imputed_x, (double)xmean, ix_arr, st_orig, end, buffer_sd, criterion, min_gain, split_point, split_ix);
|
3245
|
+
else if (criterion == DensityCrit)
|
3246
|
+
gain = find_split_dens<double, ldouble_safe>(buffer_imputed_x, ix_arr, st_orig, end, split_point, split_ix);
|
3247
|
+
else if (criterion == FullGain)
|
3248
|
+
{
|
3249
|
+
/* TODO: this buffer should be allocated from outside */
|
3250
|
+
std::vector<double> temp_buffer(mult2(ncols));
|
3251
|
+
gain = find_split_full_gain<double, ldouble_safe>(
|
3252
|
+
buffer_imputed_x, st_orig, end, ix_arr,
|
3253
|
+
cols_use, ncols_use, force_cols_use,
|
3254
|
+
X_row_major, ncols,
|
3255
|
+
Xr, Xr_ind, Xr_indptr,
|
3256
|
+
temp_buffer.data(), temp_buffer.data() + ncols,
|
3257
|
+
split_ix, split_point, true);
|
3258
|
+
}
|
3259
|
+
|
3260
|
+
/* Note: in theory, it should be possible to use a faster version assuming a contiguous array for 'x',
|
3261
|
+
but such an approach might give inexact split points. Better to avoid such inexactness at the
|
3262
|
+
expense of more computations. */
|
3263
|
+
}
|
3264
|
+
|
3265
|
+
else
|
3266
|
+
{
|
3267
|
+
if (criterion == Pooled && as_relative_gain && min_gain <= 0)
|
3268
|
+
gain = find_split_rel_gain<real_t_, ldouble_safe>(x, xmean, ix_arr, st, end, split_point, split_ix);
|
3269
|
+
else if (criterion == Pooled || criterion == Averaged)
|
3270
|
+
gain = find_split_std_gain<real_t_, ldouble_safe>(x, xmean, ix_arr, st, end, buffer_sd, criterion, min_gain, split_point, split_ix);
|
3271
|
+
else if (criterion == DensityCrit)
|
3272
|
+
gain = find_split_dens<real_t_, ldouble_safe>(x, ix_arr, st, end, split_point, split_ix);
|
3273
|
+
else if (criterion == FullGain)
|
3274
|
+
{
|
3275
|
+
/* TODO: this buffer should be allocated from outside */
|
3276
|
+
std::vector<double> temp_buffer(mult2(ncols));
|
3277
|
+
gain = find_split_full_gain<real_t_, ldouble_safe>(
|
3278
|
+
x, st, end, ix_arr,
|
3279
|
+
cols_use, ncols_use, force_cols_use,
|
3280
|
+
X_row_major, ncols,
|
3281
|
+
Xr, Xr_ind, Xr_indptr,
|
3282
|
+
temp_buffer.data(), temp_buffer.data() + ncols,
|
3283
|
+
split_ix, split_point, true);
|
3284
|
+
}
|
3285
|
+
}
|
3286
|
+
|
3287
|
+
/* Note: a gain of -Inf signals that the data is unsplittable. Zero signals it's below the minimum. */
|
3288
|
+
return std::fmax(0., gain);
|
3289
|
+
}
|
3290
|
+
|
3291
|
+
template <class real_t_, class mapping, class ldouble_safe>
|
3292
|
+
double eval_guided_crit_weighted(size_t *restrict ix_arr, size_t st, size_t end, real_t_ *restrict x,
|
3293
|
+
double *restrict buffer_sd, bool as_relative_gain,
|
3294
|
+
double *restrict buffer_imputed_x, double *restrict saved_xmedian,
|
3295
|
+
size_t &split_ix, double &restrict split_point, double &restrict xmin, double &restrict xmax,
|
3296
|
+
GainCriterion criterion, double min_gain, MissingAction missing_action,
|
3297
|
+
size_t *restrict cols_use, size_t ncols_use, bool force_cols_use,
|
3298
|
+
double *restrict X_row_major, size_t ncols,
|
3299
|
+
double *restrict Xr, size_t *restrict Xr_ind, size_t *restrict Xr_indptr,
|
3300
|
+
mapping &restrict w)
|
3301
|
+
{
|
3302
|
+
size_t st_orig = st;
|
3303
|
+
double gain = 0;
|
3304
|
+
if (criterion == DensityCrit || criterion == FullGain) min_gain = 0;
|
3305
|
+
|
3306
|
+
/* move NAs to the front if there's any, exclude them from calculations */
|
3307
|
+
if (missing_action != Fail)
|
3308
|
+
st = move_NAs_to_front(ix_arr, st, end, x);
|
3309
|
+
|
3310
|
+
if (unlikely(st >= end)) return -HUGE_VAL;
|
3311
|
+
else if (unlikely(st == (end-1)))
|
3312
|
+
{
|
3313
|
+
if (x[ix_arr[st]] == x[ix_arr[end]])
|
3314
|
+
return -HUGE_VAL;
|
3315
|
+
split_point = midpoint_with_reorder(x[ix_arr[st]], x[ix_arr[end]]);
|
3316
|
+
split_ix = st;
|
3317
|
+
gain = 1.;
|
3318
|
+
if (gain > min_gain)
|
3319
|
+
return gain;
|
3320
|
+
else
|
3321
|
+
return 0.;
|
3322
|
+
}
|
3323
|
+
|
3324
|
+
/* sort in ascending order */
|
3325
|
+
std::sort(ix_arr + st, ix_arr + end + 1, [&x](const size_t a, const size_t b){return x[a] < x[b];});
|
3326
|
+
if (x[ix_arr[st]] == x[ix_arr[end]]) return -HUGE_VAL;
|
3327
|
+
xmin = x[ix_arr[st]]; xmax = x[ix_arr[end]];
|
3328
|
+
|
3329
|
+
/* unlike the previous case for the extended model, the data here has not been centered,
|
3330
|
+
which could make the standard deviations have poor precision. It's nevertheless not
|
3331
|
+
necessary for this mean to have good precision, since it's only meant for centering,
|
3332
|
+
so it can be calculated inexactly with simd instructions. */
|
3333
|
+
real_t_ xmean = 0;
|
3334
|
+
real_t_ cnt = 0;
|
3335
|
+
if (criterion == Pooled || criterion == Averaged)
|
3336
|
+
{
|
3337
|
+
for (size_t ix = st; ix <= end; ix++)
|
3338
|
+
{
|
3339
|
+
xmean += x[ix_arr[ix]];
|
3340
|
+
cnt += w[ix_arr[ix]];
|
3341
|
+
}
|
3342
|
+
xmean /= cnt;
|
3343
|
+
}
|
3344
|
+
|
3345
|
+
if (missing_action == Impute && st > st_orig)
|
3346
|
+
{
|
3347
|
+
missing_action = Fail;
|
3348
|
+
fill_NAs_with_median(ix_arr, st_orig, st, end, x, buffer_imputed_x, saved_xmedian);
|
3349
|
+
if (criterion == Pooled && as_relative_gain && min_gain <= 0)
|
3350
|
+
gain = find_split_rel_gain_weighted<double, mapping, ldouble_safe>(buffer_imputed_x, (double)xmean, ix_arr, st_orig, end, split_point, split_ix, w);
|
3351
|
+
else if (criterion == Pooled || criterion == Averaged)
|
3352
|
+
gain = find_split_std_gain_weighted<double, mapping, ldouble_safe>(buffer_imputed_x, (double)xmean, ix_arr, st_orig, end, buffer_sd, criterion, min_gain, split_point, split_ix, w);
|
3353
|
+
else if (criterion == DensityCrit)
|
3354
|
+
gain = find_split_dens_weighted<double, mapping, ldouble_safe>(buffer_imputed_x, ix_arr, st_orig, end, split_point, split_ix, w);
|
3355
|
+
else if (criterion == FullGain)
|
3356
|
+
{
|
3357
|
+
std::vector<double> temp_buffer(mult2(ncols));
|
3358
|
+
gain = find_split_full_gain_weighted<double, mapping, ldouble_safe>(
|
3359
|
+
buffer_imputed_x, st_orig, end, ix_arr,
|
3360
|
+
cols_use, ncols_use, force_cols_use,
|
3361
|
+
X_row_major, ncols,
|
3362
|
+
Xr, Xr_ind, Xr_indptr,
|
3363
|
+
temp_buffer.data(), temp_buffer.data() + ncols,
|
3364
|
+
split_ix, split_point, true,
|
3365
|
+
w);
|
3366
|
+
}
|
3367
|
+
}
|
3368
|
+
|
3369
|
+
else
|
3370
|
+
{
|
3371
|
+
if (criterion == Pooled && as_relative_gain && min_gain <= 0)
|
3372
|
+
gain = find_split_rel_gain_weighted<real_t_, mapping, ldouble_safe>(x, xmean, ix_arr, st, end, split_point, split_ix, w);
|
3373
|
+
else if (criterion == Pooled || criterion == Averaged)
|
3374
|
+
gain = find_split_std_gain_weighted<real_t_, mapping, ldouble_safe>(x, xmean, ix_arr, st, end, buffer_sd, criterion, min_gain, split_point, split_ix, w);
|
3375
|
+
else if (criterion == DensityCrit)
|
3376
|
+
gain = find_split_dens_weighted<real_t_, mapping, ldouble_safe>(x, ix_arr, st, end, split_point, split_ix, w);
|
3377
|
+
else if (criterion == FullGain)
|
3378
|
+
{
|
3379
|
+
std::vector<double> temp_buffer(mult2(ncols));
|
3380
|
+
gain = find_split_full_gain_weighted<real_t_, mapping, ldouble_safe>(
|
3381
|
+
x, st, end, ix_arr,
|
3382
|
+
cols_use, ncols_use, force_cols_use,
|
3383
|
+
X_row_major, ncols,
|
3384
|
+
Xr, Xr_ind, Xr_indptr,
|
3385
|
+
temp_buffer.data(), temp_buffer.data() + ncols,
|
3386
|
+
split_ix, split_point, true,
|
3387
|
+
w);
|
3388
|
+
}
|
3389
|
+
}
|
3390
|
+
|
3391
|
+
/* Note: a gain of -Inf signals that the data is unsplittable. Zero signals it's below the minimum. */
|
3392
|
+
return std::fmax(0., gain);
|
3393
|
+
}
|
3394
|
+
|
3395
|
+
/* TODO: here it should only need to look at the non-zero entries. It can then use the
|
3396
|
+
same algorithm as above, but putting an extra check to see where do the zeros fit in
|
3397
|
+
the sorted order of the non-zero entries while calculating gains and SDs, and then
|
3398
|
+
call the 'divide_subset' function after-the-fact to reach the same end result.
|
3399
|
+
It should be much faster than this if the non-zero entries are few. */
|
3400
|
+
template <class real_t_, class sparse_ix, class ldouble_safe>
|
3401
|
+
double eval_guided_crit(size_t ix_arr[], size_t st, size_t end,
|
3402
|
+
size_t col_num, real_t_ Xc[], sparse_ix Xc_ind[], sparse_ix Xc_indptr[],
|
3403
|
+
double buffer_arr[], size_t buffer_pos[], bool as_relative_gain,
|
3404
|
+
double *restrict saved_xmedian,
|
3405
|
+
double &split_point, double &xmin, double &xmax,
|
3406
|
+
GainCriterion criterion, double min_gain, MissingAction missing_action,
|
3407
|
+
size_t *restrict cols_use, size_t ncols_use, bool force_cols_use,
|
3408
|
+
double *restrict X_row_major, size_t ncols,
|
3409
|
+
double *restrict Xr, size_t *restrict Xr_ind, size_t *restrict Xr_indptr)
|
3410
|
+
{
|
3411
|
+
size_t ignored;
|
3412
|
+
|
3413
|
+
|
3414
|
+
todense(ix_arr, st, end,
|
3415
|
+
col_num, Xc, Xc_ind, Xc_indptr,
|
3416
|
+
buffer_arr);
|
3417
|
+
size_t tot = end - st + 1;
|
3418
|
+
std::iota(buffer_pos, buffer_pos + tot, (size_t)0);
|
3419
|
+
|
3420
|
+
if (missing_action == Impute)
|
3421
|
+
{
|
3422
|
+
missing_action = Fail;
|
3423
|
+
for (size_t ix = 0; ix < tot; ix++)
|
3424
|
+
{
|
3425
|
+
if (unlikely(is_na_or_inf(buffer_arr[ix])))
|
3426
|
+
{
|
3427
|
+
goto fill_missing;
|
3428
|
+
}
|
3429
|
+
}
|
3430
|
+
goto no_nas;
|
3431
|
+
|
3432
|
+
fill_missing:
|
3433
|
+
{
|
3434
|
+
size_t idx_half = div2(tot);
|
3435
|
+
std::nth_element(buffer_pos, buffer_pos + idx_half, buffer_pos + tot,
|
3436
|
+
[&buffer_arr](const size_t a, const size_t b){return buffer_arr[a] < buffer_arr[b];});
|
3437
|
+
*saved_xmedian = buffer_arr[buffer_pos[idx_half]];
|
3438
|
+
|
3439
|
+
if ((tot % 2) == 0)
|
3440
|
+
{
|
3441
|
+
double xlow = *std::max_element(buffer_pos, buffer_pos + idx_half);
|
3442
|
+
*saved_xmedian = xlow + ((*saved_xmedian)-xlow)/2.;
|
3443
|
+
}
|
3444
|
+
|
3445
|
+
for (size_t ix = 0; ix < tot; ix++)
|
3446
|
+
buffer_arr[ix] = is_na_or_inf(buffer_arr[ix])? (*saved_xmedian) : buffer_arr[ix];
|
3447
|
+
std::iota(buffer_pos, buffer_pos + tot, (size_t)0);
|
3448
|
+
}
|
3449
|
+
}
|
3450
|
+
|
3451
|
+
no_nas:
|
3452
|
+
return eval_guided_crit<double, ldouble_safe>(
|
3453
|
+
buffer_pos, 0, end - st, buffer_arr, buffer_arr + tot,
|
3454
|
+
as_relative_gain, saved_xmedian, (double*)NULL, ignored, split_point,
|
3455
|
+
xmin, xmax, criterion, min_gain, missing_action,
|
3456
|
+
cols_use, ncols_use, force_cols_use,
|
3457
|
+
X_row_major, ncols,
|
3458
|
+
Xr, Xr_ind, Xr_indptr);
|
3459
|
+
}
|
3460
|
+
|
3461
|
+
template <class real_t_, class sparse_ix, class mapping, class ldouble_safe>
|
3462
|
+
double eval_guided_crit_weighted(size_t ix_arr[], size_t st, size_t end,
|
3463
|
+
size_t col_num, real_t_ Xc[], sparse_ix Xc_ind[], sparse_ix Xc_indptr[],
|
3464
|
+
double buffer_arr[], size_t buffer_pos[], bool as_relative_gain,
|
3465
|
+
double *restrict saved_xmedian,
|
3466
|
+
double &restrict split_point, double &restrict xmin, double &restrict xmax,
|
3467
|
+
GainCriterion criterion, double min_gain, MissingAction missing_action,
|
3468
|
+
size_t *restrict cols_use, size_t ncols_use, bool force_cols_use,
|
3469
|
+
double *restrict X_row_major, size_t ncols,
|
3470
|
+
double *restrict Xr, size_t *restrict Xr_ind, size_t *restrict Xr_indptr,
|
3471
|
+
mapping &restrict w)
|
3472
|
+
{
|
3473
|
+
size_t ignored;
|
3474
|
+
|
3475
|
+
|
3476
|
+
todense(ix_arr, st, end,
|
3477
|
+
col_num, Xc, Xc_ind, Xc_indptr,
|
3478
|
+
buffer_arr);
|
3479
|
+
size_t tot = end - st + 1;
|
3480
|
+
std::iota(buffer_pos, buffer_pos + tot, (size_t)0);
|
3481
|
+
|
3482
|
+
|
3483
|
+
if (missing_action == Impute)
|
3484
|
+
{
|
3485
|
+
missing_action = Fail;
|
3486
|
+
for (size_t ix = 0; ix < tot; ix++)
|
3487
|
+
{
|
3488
|
+
if (unlikely(is_na_or_inf(buffer_arr[ix])))
|
3489
|
+
{
|
3490
|
+
goto fill_missing;
|
3491
|
+
}
|
3492
|
+
}
|
3493
|
+
goto no_nas;
|
3494
|
+
|
3495
|
+
fill_missing:
|
3496
|
+
{
|
3497
|
+
size_t idx_half = div2(tot);
|
3498
|
+
std::nth_element(buffer_pos, buffer_pos + idx_half, buffer_pos + tot,
|
3499
|
+
[&buffer_arr](const size_t a, const size_t b){return buffer_arr[a] < buffer_arr[b];});
|
3500
|
+
*saved_xmedian = buffer_arr[buffer_pos[idx_half]];
|
3501
|
+
|
3502
|
+
if ((tot % 2) == 0)
|
3503
|
+
{
|
3504
|
+
double xlow = *std::max_element(buffer_pos, buffer_pos + idx_half);
|
3505
|
+
*saved_xmedian = xlow + ((*saved_xmedian)-xlow)/2.;
|
3506
|
+
}
|
3507
|
+
|
3508
|
+
for (size_t ix = 0; ix < tot; ix++)
|
3509
|
+
buffer_arr[ix] = is_na_or_inf(buffer_arr[ix])? (*saved_xmedian) : buffer_arr[ix];
|
3510
|
+
std::iota(buffer_pos, buffer_pos + tot, (size_t)0);
|
3511
|
+
}
|
3512
|
+
}
|
3513
|
+
|
3514
|
+
|
3515
|
+
no_nas:
|
3516
|
+
/* TODO: allocate this buffer externally */
|
3517
|
+
std::vector<double> buffer_w(tot);
|
3518
|
+
for (size_t row = st; row <= end; row++)
|
3519
|
+
buffer_w[row-st] = w[ix_arr[row]];
|
3520
|
+
/* TODO: in this case, as the weights match with the order of the indices, could use a faster version
|
3521
|
+
with a weighted rel_gain function instead (not yet implemented). */
|
3522
|
+
return eval_guided_crit_weighted<double, std::vector<double>, ldouble_safe>(
|
3523
|
+
buffer_pos, 0, end - st, buffer_arr, buffer_arr + tot,
|
3524
|
+
as_relative_gain, saved_xmedian, (double*)NULL, ignored, split_point,
|
3525
|
+
xmin, xmax, criterion, min_gain, missing_action,
|
3526
|
+
cols_use, ncols_use, force_cols_use,
|
3527
|
+
X_row_major, ncols,
|
3528
|
+
Xr, Xr_ind, Xr_indptr,
|
3529
|
+
buffer_w);
|
3530
|
+
}
|
3531
|
+
|
3532
|
+
/* How this works:
|
3533
|
+
- For Averaged criterion, will take the expected standard deviation that would be gotten with the category counts
|
3534
|
+
if each category got assigned a real number at random ~ Unif(0,1) and the data were thus converted to
|
3535
|
+
numerical. In such case, the best split (highest sd gain) is always putting the second-highest count in one
|
3536
|
+
branch, so there is no point in doing a full search over other permutations. In order to get more reasonable
|
3537
|
+
splits, when using the option to split by subsets of categories, it will sort the counts and evaluate only
|
3538
|
+
splits in which the categories are grouped in sorted order - in such cases it tends to pick either the
|
3539
|
+
smallest or the largest category to assign to one branch, but sometimes picks groups too.
|
3540
|
+
- For Pooled criterion, will take shannon entropy, which tends to make a more even split. In the case of splitting
|
3541
|
+
by a single category, it always puts the largest category in a separate branch. In the case of subsets,
|
3542
|
+
it can either evaluate possible splits over all permutations (not feasible if there are too many categories),
|
3543
|
+
or look up for splits in sorted order just like for Averaged criterion.
|
3544
|
+
Splitting by averaged Gini gain (like with Averaged) also selects always the second-largest category to put in one branch,
|
3545
|
+
while splitting by weighted Gini (like with Pooled) usually selects the largest category to put in one branch. The
|
3546
|
+
Gini gain is not easily comparable to that of numerical columns, so it's not offered as an option here.
|
3547
|
+
*/
|
3548
|
+
/* https://math.stackexchange.com/questions/3343384/expected-variance-and-kurtosis-from-pmf-in-which-possible-discrete-values-are-dr */
|
3549
|
+
/* TODO: 'buffer_pos' doesn't need to be 'size_t', 'int' would suffice */
|
3550
|
+
template <class ldouble_safe>
|
3551
|
+
double eval_guided_crit(size_t *restrict ix_arr, size_t st, size_t end, int *restrict x, int ncat,
|
3552
|
+
int *restrict saved_cat_mode,
|
3553
|
+
size_t *restrict buffer_cnt, size_t *restrict buffer_pos, double *restrict buffer_prob,
|
3554
|
+
int &restrict chosen_cat, signed char *restrict split_categ, signed char *restrict buffer_split,
|
3555
|
+
GainCriterion criterion, double min_gain, bool all_perm,
|
3556
|
+
MissingAction missing_action, CategSplit cat_split_type)
|
3557
|
+
{
|
3558
|
+
if (criterion == DensityCrit)
|
3559
|
+
return find_split_dens_longform<size_t, ldouble_safe>(
|
3560
|
+
x, ncat, ix_arr, st, end,
|
3561
|
+
cat_split_type, missing_action,
|
3562
|
+
chosen_cat, split_categ, saved_cat_mode,
|
3563
|
+
buffer_cnt, buffer_pos);
|
3564
|
+
if (st >= end) return -HUGE_VAL;
|
3565
|
+
size_t n_nas = 0;
|
3566
|
+
int xval;
|
3567
|
+
|
3568
|
+
/* count categories */
|
3569
|
+
memset(buffer_cnt, 0, sizeof(size_t) * ncat);
|
3570
|
+
if (missing_action == Fail)
|
3571
|
+
{
|
3572
|
+
for (size_t row = st; row <= end; row++)
|
3573
|
+
if (likely(x[ix_arr[row]] >= 0))
|
3574
|
+
buffer_cnt[x[ix_arr[row]]]++;
|
3575
|
+
}
|
3576
|
+
|
3577
|
+
else if (missing_action == Impute)
|
3578
|
+
{
|
3579
|
+
for (size_t row = st; row <= end; row++)
|
3580
|
+
{
|
3581
|
+
xval = x[ix_arr[row]];
|
3582
|
+
if (unlikely(xval < 0))
|
3583
|
+
n_nas++;
|
3584
|
+
else
|
3585
|
+
buffer_cnt[xval]++;
|
3586
|
+
}
|
3587
|
+
|
3588
|
+
if (unlikely(n_nas >= end-st)) return -HUGE_VAL;
|
3589
|
+
|
3590
|
+
if (n_nas)
|
3591
|
+
{
|
3592
|
+
auto idxmax = std::max_element(buffer_cnt, buffer_cnt + ncat);
|
3593
|
+
*idxmax += n_nas;
|
3594
|
+
*saved_cat_mode = (int)std::distance(buffer_cnt, idxmax);
|
3595
|
+
}
|
3596
|
+
}
|
3597
|
+
|
3598
|
+
else
|
3599
|
+
{
|
3600
|
+
for (size_t row = st; row <= end; row++)
|
3601
|
+
{
|
3602
|
+
xval = x[ix_arr[row]];
|
3603
|
+
if (likely(xval >= 0)) buffer_cnt[xval]++;
|
3604
|
+
}
|
3605
|
+
}
|
3606
|
+
|
3607
|
+
double this_gain = -HUGE_VAL;
|
3608
|
+
double best_gain = -HUGE_VAL;
|
3609
|
+
std::iota(buffer_pos, buffer_pos + ncat, (size_t)0);
|
3610
|
+
size_t st_pos = 0;
|
3611
|
+
|
3612
|
+
switch (cat_split_type)
|
3613
|
+
{
|
3614
|
+
case SingleCateg:
|
3615
|
+
{
|
3616
|
+
size_t cnt = end - st + 1;
|
3617
|
+
ldouble_safe cnt_l = (ldouble_safe) cnt;
|
3618
|
+
size_t ncat_present = 0;
|
3619
|
+
|
3620
|
+
switch(criterion)
|
3621
|
+
{
|
3622
|
+
case Averaged:
|
3623
|
+
{
|
3624
|
+
/* move zero-counts to the beginning */
|
3625
|
+
size_t temp;
|
3626
|
+
for (int cat = 0; cat < ncat; cat++)
|
3627
|
+
{
|
3628
|
+
if (buffer_cnt[cat])
|
3629
|
+
{
|
3630
|
+
ncat_present++;
|
3631
|
+
buffer_prob[cat] = (ldouble_safe) buffer_cnt[cat] / cnt_l;
|
3632
|
+
}
|
3633
|
+
|
3634
|
+
else
|
3635
|
+
{
|
3636
|
+
temp = buffer_pos[st_pos];
|
3637
|
+
buffer_pos[st_pos] = buffer_pos[cat];
|
3638
|
+
buffer_pos[cat] = temp;
|
3639
|
+
st_pos++;
|
3640
|
+
}
|
3641
|
+
}
|
3642
|
+
|
3643
|
+
if (ncat_present <= 1) return -HUGE_VAL;
|
3644
|
+
|
3645
|
+
double sd_full = expected_sd_cat<size_t, ldouble_safe>(buffer_prob, ncat_present, buffer_pos + st_pos);
|
3646
|
+
|
3647
|
+
/* try isolating each category one at a time */
|
3648
|
+
for (size_t pos = st_pos; (int)pos < ncat; pos++)
|
3649
|
+
{
|
3650
|
+
this_gain = sd_gain(sd_full,
|
3651
|
+
0.0,
|
3652
|
+
(expected_sd_cat_single<size_t, size_t, ldouble_safe>(buffer_cnt, buffer_prob, ncat_present, buffer_pos + st_pos, pos - st_pos, cnt)));
|
3653
|
+
if (this_gain > min_gain && this_gain > best_gain)
|
3654
|
+
{
|
3655
|
+
best_gain = this_gain;
|
3656
|
+
chosen_cat = buffer_pos[pos];
|
3657
|
+
}
|
3658
|
+
}
|
3659
|
+
break;
|
3660
|
+
}
|
3661
|
+
|
3662
|
+
case Pooled:
|
3663
|
+
{
|
3664
|
+
/* here it will always pick the largest one */
|
3665
|
+
size_t ncat_present = 0;
|
3666
|
+
size_t cnt_max = 0;
|
3667
|
+
for (int cat = 0; cat < ncat; cat++)
|
3668
|
+
{
|
3669
|
+
if (buffer_cnt[cat])
|
3670
|
+
{
|
3671
|
+
ncat_present++;
|
3672
|
+
if (cnt_max < buffer_cnt[cat])
|
3673
|
+
{
|
3674
|
+
cnt_max = buffer_cnt[cat];
|
3675
|
+
chosen_cat = cat;
|
3676
|
+
}
|
3677
|
+
}
|
3678
|
+
}
|
3679
|
+
|
3680
|
+
if (ncat_present <= 1) return -HUGE_VAL;
|
3681
|
+
|
3682
|
+
ldouble_safe cnt_left = (ldouble_safe)((end - st + 1) - cnt_max);
|
3683
|
+
this_gain = (
|
3684
|
+
(ldouble_safe)cnt * std::log((ldouble_safe)cnt)
|
3685
|
+
- cnt_left * std::log(cnt_left)
|
3686
|
+
- (ldouble_safe)cnt_max * std::log((ldouble_safe)cnt_max)
|
3687
|
+
) / cnt;
|
3688
|
+
best_gain = (this_gain > min_gain)? this_gain : best_gain;
|
3689
|
+
break;
|
3690
|
+
}
|
3691
|
+
|
3692
|
+
default:
|
3693
|
+
{
|
3694
|
+
unexpected_error();
|
3695
|
+
break;
|
3696
|
+
}
|
3697
|
+
}
|
3698
|
+
break;
|
3699
|
+
}
|
3700
|
+
|
3701
|
+
case SubSet:
|
3702
|
+
{
|
3703
|
+
/* sort by counts */
|
3704
|
+
std::sort(buffer_pos, buffer_pos + ncat, [&buffer_cnt](const size_t a, const size_t b){return buffer_cnt[a] < buffer_cnt[b];});
|
3705
|
+
|
3706
|
+
/* set split as: (1):left (0):right (-1):not_present */
|
3707
|
+
memset(buffer_split, 0, ncat * sizeof(signed char));
|
3708
|
+
|
3709
|
+
ldouble_safe cnt = (ldouble_safe)(end - st + 1);
|
3710
|
+
|
3711
|
+
switch(criterion)
|
3712
|
+
{
|
3713
|
+
case Averaged:
|
3714
|
+
{
|
3715
|
+
/* determine first non-zero and convert to probabilities */
|
3716
|
+
double sd_full;
|
3717
|
+
for (int cat = 0; cat < ncat; cat++)
|
3718
|
+
{
|
3719
|
+
if (buffer_cnt[buffer_pos[cat]])
|
3720
|
+
{
|
3721
|
+
buffer_prob[buffer_pos[cat]] = (ldouble_safe)buffer_cnt[buffer_pos[cat]] / cnt;
|
3722
|
+
}
|
3723
|
+
|
3724
|
+
else
|
3725
|
+
{
|
3726
|
+
buffer_split[buffer_pos[cat]] = -1;
|
3727
|
+
st_pos++;
|
3728
|
+
}
|
3729
|
+
}
|
3730
|
+
|
3731
|
+
if ((int)st_pos >= (ncat-1)) return -HUGE_VAL;
|
3732
|
+
|
3733
|
+
/* calculate full SD assuming they take values randomly ~Unif(0, 1) */
|
3734
|
+
size_t ncat_present = (size_t)ncat - st_pos;
|
3735
|
+
sd_full = expected_sd_cat<size_t, ldouble_safe>(buffer_prob, ncat_present, buffer_pos + st_pos);
|
3736
|
+
if (ncat_present >= log2ceil(SIZE_MAX)) all_perm = false;
|
3737
|
+
|
3738
|
+
/* move categories one at a time */
|
3739
|
+
for (size_t pos = st_pos; pos < ((size_t)ncat - st_pos - 1); pos++)
|
3740
|
+
{
|
3741
|
+
buffer_split[buffer_pos[pos]] = 1;
|
3742
|
+
this_gain = sd_gain(sd_full,
|
3743
|
+
(expected_sd_cat<size_t, size_t, ldouble_safe>(buffer_cnt, buffer_prob, pos - st_pos + 1, buffer_pos + st_pos)),
|
3744
|
+
(expected_sd_cat<size_t, size_t, ldouble_safe>(buffer_cnt, buffer_prob, (size_t)ncat - pos - 1, buffer_pos + pos + 1))
|
3745
|
+
);
|
3746
|
+
if (this_gain > min_gain && this_gain > best_gain)
|
3747
|
+
{
|
3748
|
+
best_gain = this_gain;
|
3749
|
+
memcpy(split_categ, buffer_split, ncat * sizeof(signed char));
|
3750
|
+
}
|
3751
|
+
}
|
3752
|
+
|
3753
|
+
break;
|
3754
|
+
}
|
3755
|
+
|
3756
|
+
case Pooled:
|
3757
|
+
{
|
3758
|
+
ldouble_safe s = 0;
|
3759
|
+
|
3760
|
+
/* determine first non-zero and get base info */
|
3761
|
+
for (int cat = 0; cat < ncat; cat++)
|
3762
|
+
{
|
3763
|
+
if (buffer_cnt[buffer_pos[cat]])
|
3764
|
+
{
|
3765
|
+
s += (buffer_cnt[buffer_pos[cat]] <= 1)?
|
3766
|
+
0 : ((ldouble_safe) buffer_cnt[buffer_pos[cat]] * std::log((ldouble_safe)buffer_cnt[buffer_pos[cat]]));
|
3767
|
+
}
|
3768
|
+
|
3769
|
+
else
|
3770
|
+
{
|
3771
|
+
buffer_split[buffer_pos[cat]] = -1;
|
3772
|
+
st_pos++;
|
3773
|
+
}
|
3774
|
+
}
|
3775
|
+
|
3776
|
+
if ((int)st_pos >= (ncat-1)) return -HUGE_VAL;
|
3777
|
+
|
3778
|
+
/* calculate base info */
|
3779
|
+
ldouble_safe base_info = cnt * std::log(cnt) - s;
|
3780
|
+
|
3781
|
+
if (all_perm)
|
3782
|
+
{
|
3783
|
+
size_t cnt_left, cnt_right;
|
3784
|
+
double s_left, s_right;
|
3785
|
+
size_t ncat_present = (size_t)ncat - st_pos;
|
3786
|
+
size_t ncomb = pow2(ncat_present) - 1;
|
3787
|
+
size_t best_combin;
|
3788
|
+
|
3789
|
+
for (size_t combin = 1; combin < ncomb; combin++)
|
3790
|
+
{
|
3791
|
+
cnt_left = 0; cnt_right = 0;
|
3792
|
+
s_left = 0; s_right = 0;
|
3793
|
+
for (size_t pos = st_pos; (int)pos < ncat; pos++)
|
3794
|
+
{
|
3795
|
+
if (extract_bit(combin, pos))
|
3796
|
+
{
|
3797
|
+
cnt_left += buffer_cnt[buffer_pos[pos]];
|
3798
|
+
s_left += (buffer_cnt[buffer_pos[pos]] <= 1)?
|
3799
|
+
0 : ((ldouble_safe) buffer_cnt[buffer_pos[pos]]
|
3800
|
+
* std::log((ldouble_safe) buffer_cnt[buffer_pos[pos]]));
|
3801
|
+
}
|
3802
|
+
|
3803
|
+
else
|
3804
|
+
{
|
3805
|
+
cnt_right += buffer_cnt[buffer_pos[pos]];
|
3806
|
+
s_right += (buffer_cnt[buffer_pos[pos]] <= 1)?
|
3807
|
+
0 : ((ldouble_safe) buffer_cnt[buffer_pos[pos]]
|
3808
|
+
* std::log((ldouble_safe) buffer_cnt[buffer_pos[pos]]));
|
3809
|
+
}
|
3810
|
+
}
|
3811
|
+
|
3812
|
+
this_gain = categ_gain<size_t, ldouble_safe>(
|
3813
|
+
cnt_left, cnt_right,
|
3814
|
+
s_left, s_right,
|
3815
|
+
base_info, cnt);
|
3816
|
+
|
3817
|
+
if (this_gain > min_gain && this_gain > best_gain)
|
3818
|
+
{
|
3819
|
+
best_gain = this_gain;
|
3820
|
+
best_combin = combin;
|
3821
|
+
}
|
3822
|
+
|
3823
|
+
}
|
3824
|
+
|
3825
|
+
if (best_gain > min_gain)
|
3826
|
+
for (size_t pos = 0; pos < ncat_present; pos++)
|
3827
|
+
split_categ[buffer_pos[st_pos + pos]] = extract_bit(best_combin, pos);
|
3828
|
+
|
3829
|
+
}
|
3830
|
+
|
3831
|
+
else
|
3832
|
+
{
|
3833
|
+
/* try moving the categories one at a time */
|
3834
|
+
size_t cnt_left = 0;
|
3835
|
+
size_t cnt_right = end - st + 1;
|
3836
|
+
double s_left = 0;
|
3837
|
+
double s_right = s;
|
3838
|
+
|
3839
|
+
for (size_t pos = st_pos; pos < (ncat - st_pos - 1); pos++)
|
3840
|
+
{
|
3841
|
+
buffer_split[buffer_pos[pos]] = 1;
|
3842
|
+
s_left += (buffer_cnt[buffer_pos[pos]] <= 1)?
|
3843
|
+
0 : ((ldouble_safe)buffer_cnt[buffer_pos[pos]] * std::log((ldouble_safe)buffer_cnt[buffer_pos[pos]]));
|
3844
|
+
s_right -= (buffer_cnt[buffer_pos[pos]] <= 1)?
|
3845
|
+
0 : ((ldouble_safe)buffer_cnt[buffer_pos[pos]] * std::log((ldouble_safe)buffer_cnt[buffer_pos[pos]]));
|
3846
|
+
cnt_left += buffer_cnt[buffer_pos[pos]];
|
3847
|
+
cnt_right -= buffer_cnt[buffer_pos[pos]];
|
3848
|
+
|
3849
|
+
this_gain = categ_gain<size_t, ldouble_safe>(
|
3850
|
+
cnt_left, cnt_right,
|
3851
|
+
s_left, s_right,
|
3852
|
+
base_info, cnt);
|
3853
|
+
|
3854
|
+
if (this_gain > min_gain && this_gain > best_gain)
|
3855
|
+
{
|
3856
|
+
best_gain = this_gain;
|
3857
|
+
memcpy(split_categ, buffer_split, ncat * sizeof(signed char));
|
3858
|
+
}
|
3859
|
+
}
|
3860
|
+
}
|
3861
|
+
|
3862
|
+
break;
|
3863
|
+
}
|
3864
|
+
|
3865
|
+
default:
|
3866
|
+
{
|
3867
|
+
unexpected_error();
|
3868
|
+
break;
|
3869
|
+
}
|
3870
|
+
}
|
3871
|
+
}
|
3872
|
+
}
|
3873
|
+
|
3874
|
+
if (st == (end-1)) return 0;
|
3875
|
+
|
3876
|
+
if (best_gain <= -HUGE_VAL && this_gain <= min_gain && this_gain > -HUGE_VAL)
|
3877
|
+
return 0;
|
3878
|
+
else
|
3879
|
+
return best_gain;
|
3880
|
+
}
|
3881
|
+
|
3882
|
+
|
3883
|
+
template <class mapping, class ldouble_safe>
|
3884
|
+
double eval_guided_crit_weighted(size_t *restrict ix_arr, size_t st, size_t end, int *restrict x, int ncat,
|
3885
|
+
int *restrict saved_cat_mode,
|
3886
|
+
size_t *restrict buffer_pos, double *restrict buffer_prob,
|
3887
|
+
int &restrict chosen_cat, signed char *restrict split_categ, signed char *restrict buffer_split,
|
3888
|
+
GainCriterion criterion, double min_gain, bool all_perm,
|
3889
|
+
MissingAction missing_action, CategSplit cat_split_type,
|
3890
|
+
mapping &restrict w)
|
3891
|
+
{
|
3892
|
+
if (criterion == DensityCrit)
|
3893
|
+
return find_split_dens_longform_weighted<mapping, size_t, ldouble_safe>(
|
3894
|
+
x, ncat, ix_arr, st, end,
|
3895
|
+
cat_split_type, missing_action,
|
3896
|
+
chosen_cat, split_categ, saved_cat_mode,
|
3897
|
+
buffer_pos, w);
|
3898
|
+
if (st >= end) return -HUGE_VAL;
|
3899
|
+
ldouble_safe w_missing = 0;
|
3900
|
+
int xval;
|
3901
|
+
size_t ix_;
|
3902
|
+
|
3903
|
+
/* count categories */
|
3904
|
+
/* TODO: allocate this buffer externally */
|
3905
|
+
std::vector<ldouble_safe> buffer_cnt(ncat, (ldouble_safe)0);
|
3906
|
+
if (missing_action == Fail)
|
3907
|
+
{
|
3908
|
+
for (size_t row = st; row <= end; row++)
|
3909
|
+
{
|
3910
|
+
ix_ = ix_arr[row];
|
3911
|
+
if (unlikely(x[ix_]) < 0) continue;
|
3912
|
+
buffer_cnt[x[ix_]] += w[ix_];
|
3913
|
+
}
|
3914
|
+
}
|
3915
|
+
|
3916
|
+
else if (missing_action == Impute)
|
3917
|
+
{
|
3918
|
+
for (size_t row = st; row <= end; row++)
|
3919
|
+
{
|
3920
|
+
ix_ = ix_arr[row];
|
3921
|
+
xval = x[ix_];
|
3922
|
+
if (unlikely(xval < 0))
|
3923
|
+
w_missing += w[ix_];
|
3924
|
+
else
|
3925
|
+
buffer_cnt[xval] += w[ix_];
|
3926
|
+
}
|
3927
|
+
|
3928
|
+
if (w_missing)
|
3929
|
+
{
|
3930
|
+
auto idxmax = std::max_element(buffer_cnt.begin(), buffer_cnt.end());
|
3931
|
+
*idxmax += w_missing;
|
3932
|
+
*saved_cat_mode = (int)std::distance(buffer_cnt.begin(), idxmax);
|
3933
|
+
}
|
3934
|
+
}
|
3935
|
+
|
3936
|
+
else
|
3937
|
+
{
|
3938
|
+
for (size_t row = st; row <= end; row++)
|
3939
|
+
{
|
3940
|
+
ix_ = ix_arr[row];
|
3941
|
+
xval = x[ix_];
|
3942
|
+
if (likely(xval >= 0)) buffer_cnt[xval] += w[ix_];
|
3943
|
+
}
|
3944
|
+
}
|
3945
|
+
|
3946
|
+
ldouble_safe cnt = std::accumulate(buffer_cnt.begin(), buffer_cnt.end(), (ldouble_safe)0);
|
3947
|
+
|
3948
|
+
double this_gain = -HUGE_VAL;
|
3949
|
+
double best_gain = -HUGE_VAL;
|
3950
|
+
std::iota(buffer_pos, buffer_pos + ncat, (size_t)0);
|
3951
|
+
size_t st_pos = 0;
|
3952
|
+
|
3953
|
+
switch(cat_split_type)
|
3954
|
+
{
|
3955
|
+
case SingleCateg:
|
3956
|
+
{
|
3957
|
+
size_t ncat_present = 0;
|
3958
|
+
|
3959
|
+
switch(criterion)
|
3960
|
+
{
|
3961
|
+
case Averaged:
|
3962
|
+
{
|
3963
|
+
/* move zero-counts to the beginning */
|
3964
|
+
size_t temp;
|
3965
|
+
for (int cat = 0; cat < ncat; cat++)
|
3966
|
+
{
|
3967
|
+
if (buffer_cnt[cat])
|
3968
|
+
{
|
3969
|
+
ncat_present++;
|
3970
|
+
buffer_prob[cat] = buffer_cnt[cat] / cnt;
|
3971
|
+
}
|
3972
|
+
|
3973
|
+
else
|
3974
|
+
{
|
3975
|
+
temp = buffer_pos[st_pos];
|
3976
|
+
buffer_pos[st_pos] = buffer_pos[cat];
|
3977
|
+
buffer_pos[cat] = temp;
|
3978
|
+
st_pos++;
|
3979
|
+
}
|
3980
|
+
}
|
3981
|
+
|
3982
|
+
if (ncat_present <= 1) return -HUGE_VAL;
|
3983
|
+
|
3984
|
+
double sd_full = expected_sd_cat<size_t, ldouble_safe>(buffer_prob, ncat_present, buffer_pos + st_pos);
|
3985
|
+
|
3986
|
+
/* try isolating each category one at a time */
|
3987
|
+
for (size_t pos = st_pos; (int)pos < ncat; pos++)
|
3988
|
+
{
|
3989
|
+
this_gain = sd_gain(sd_full,
|
3990
|
+
0.0,
|
3991
|
+
(expected_sd_cat_single<ldouble_safe, size_t, ldouble_safe>(buffer_cnt.data(), buffer_prob, ncat_present, buffer_pos + st_pos, pos - st_pos, cnt))
|
3992
|
+
);
|
3993
|
+
if (this_gain > min_gain && this_gain > best_gain)
|
3994
|
+
{
|
3995
|
+
best_gain = this_gain;
|
3996
|
+
chosen_cat = buffer_pos[pos];
|
3997
|
+
}
|
3998
|
+
}
|
3999
|
+
break;
|
4000
|
+
}
|
4001
|
+
|
4002
|
+
case Pooled:
|
4003
|
+
{
|
4004
|
+
/* here it will always pick the largest one */
|
4005
|
+
size_t ncat_present = 0;
|
4006
|
+
ldouble_safe cnt_max = 0;
|
4007
|
+
for (int cat = 0; cat < ncat; cat++)
|
4008
|
+
{
|
4009
|
+
if (buffer_cnt[cat])
|
4010
|
+
{
|
4011
|
+
ncat_present++;
|
4012
|
+
if (cnt_max < buffer_cnt[cat])
|
4013
|
+
{
|
4014
|
+
cnt_max = buffer_cnt[cat];
|
4015
|
+
chosen_cat = cat;
|
4016
|
+
}
|
4017
|
+
}
|
4018
|
+
}
|
4019
|
+
|
4020
|
+
if (ncat_present <= 1) return -HUGE_VAL;
|
4021
|
+
|
4022
|
+
ldouble_safe cnt_left = (ldouble_safe)(cnt - cnt_max);
|
4023
|
+
|
4024
|
+
/* TODO: think of a better way of dealing with numbers between zero and one */
|
4025
|
+
this_gain = (
|
4026
|
+
std::fmax((ldouble_safe)1, cnt) * std::log(std::fmax((ldouble_safe)1, cnt))
|
4027
|
+
- std::fmax((ldouble_safe)1, cnt_left) * std::log(std::fmax((ldouble_safe)1, cnt_left))
|
4028
|
+
- std::fmax((ldouble_safe)1, cnt_max) * std::log(std::fmax((ldouble_safe)1, cnt_max))
|
4029
|
+
) / std::fmax((ldouble_safe)1, cnt);
|
4030
|
+
best_gain = (this_gain > min_gain)? this_gain : best_gain;
|
4031
|
+
break;
|
4032
|
+
}
|
4033
|
+
|
4034
|
+
default:
|
4035
|
+
{
|
4036
|
+
unexpected_error();
|
4037
|
+
break;
|
4038
|
+
}
|
4039
|
+
}
|
4040
|
+
break;
|
4041
|
+
}
|
4042
|
+
|
4043
|
+
case SubSet:
|
4044
|
+
{
|
4045
|
+
/* sort by counts */
|
4046
|
+
std::sort(buffer_pos, buffer_pos + ncat, [&buffer_cnt](const size_t a, const size_t b){return buffer_cnt[a] < buffer_cnt[b];});
|
4047
|
+
|
4048
|
+
/* set split as: (1):left (0):right (-1):not_present */
|
4049
|
+
memset(buffer_split, 0, ncat * sizeof(signed char));
|
4050
|
+
|
4051
|
+
|
4052
|
+
switch(criterion)
|
4053
|
+
{
|
4054
|
+
case Averaged:
|
4055
|
+
{
|
4056
|
+
/* determine first non-zero and convert to probabilities */
|
4057
|
+
double sd_full;
|
4058
|
+
for (int cat = 0; cat < ncat; cat++)
|
4059
|
+
{
|
4060
|
+
if (buffer_cnt[buffer_pos[cat]])
|
4061
|
+
{
|
4062
|
+
buffer_prob[buffer_pos[cat]] = (ldouble_safe)buffer_cnt[buffer_pos[cat]] / cnt;
|
4063
|
+
}
|
4064
|
+
|
4065
|
+
else
|
4066
|
+
{
|
4067
|
+
buffer_split[buffer_pos[cat]] = -1;
|
4068
|
+
st_pos++;
|
4069
|
+
}
|
4070
|
+
}
|
4071
|
+
|
4072
|
+
if ((int)st_pos >= (ncat-1)) return -HUGE_VAL;
|
4073
|
+
|
4074
|
+
/* calculate full SD assuming they take values randomly ~Unif(0, 1) */
|
4075
|
+
size_t ncat_present = (size_t)ncat - st_pos;
|
4076
|
+
sd_full = expected_sd_cat<size_t, ldouble_safe>(buffer_prob, ncat_present, buffer_pos + st_pos);
|
4077
|
+
if (ncat_present >= log2ceil(SIZE_MAX)) all_perm = false;
|
4078
|
+
|
4079
|
+
/* move categories one at a time */
|
4080
|
+
for (size_t pos = st_pos; pos < ((size_t)ncat - st_pos - 1); pos++)
|
4081
|
+
{
|
4082
|
+
buffer_split[buffer_pos[pos]] = 1;
|
4083
|
+
/* TODO: is this correct? */
|
4084
|
+
this_gain = sd_gain(sd_full,
|
4085
|
+
(expected_sd_cat<ldouble_safe, size_t, ldouble_safe>(buffer_cnt.data(), buffer_prob, pos - st_pos + 1, buffer_pos + st_pos)),
|
4086
|
+
(expected_sd_cat<ldouble_safe, size_t, ldouble_safe>(buffer_cnt.data(), buffer_prob, (size_t)ncat - pos - 1, buffer_pos + pos + 1))
|
4087
|
+
);
|
4088
|
+
if (this_gain > min_gain && this_gain > best_gain)
|
4089
|
+
{
|
4090
|
+
best_gain = this_gain;
|
4091
|
+
memcpy(split_categ, buffer_split, ncat * sizeof(signed char));
|
4092
|
+
}
|
4093
|
+
}
|
4094
|
+
|
4095
|
+
break;
|
4096
|
+
}
|
4097
|
+
|
4098
|
+
case Pooled:
|
4099
|
+
{
|
4100
|
+
ldouble_safe s = 0;
|
4101
|
+
|
4102
|
+
/* determine first non-zero and get base info */
|
4103
|
+
for (int cat = 0; cat < ncat; cat++)
|
4104
|
+
{
|
4105
|
+
if (buffer_cnt[buffer_pos[cat]])
|
4106
|
+
{
|
4107
|
+
s += (buffer_cnt[buffer_pos[cat]] <= 1)?
|
4108
|
+
(ldouble_safe)0
|
4109
|
+
:
|
4110
|
+
((ldouble_safe) buffer_cnt[buffer_pos[cat]] * std::log((ldouble_safe)buffer_cnt[buffer_pos[cat]]));
|
4111
|
+
}
|
4112
|
+
|
4113
|
+
else
|
4114
|
+
{
|
4115
|
+
buffer_split[buffer_pos[cat]] = -1;
|
4116
|
+
st_pos++;
|
4117
|
+
}
|
4118
|
+
}
|
4119
|
+
|
4120
|
+
if ((int)st_pos >= (ncat-1)) return -HUGE_VAL;
|
4121
|
+
|
4122
|
+
/* calculate base info */
|
4123
|
+
ldouble_safe base_info = std::fmax((ldouble_safe)1, cnt) * std::log(std::fmax((ldouble_safe)1, cnt)) - s;
|
4124
|
+
|
4125
|
+
if (all_perm)
|
4126
|
+
{
|
4127
|
+
size_t cnt_left, cnt_right;
|
4128
|
+
double s_left, s_right;
|
4129
|
+
size_t ncat_present = (size_t)ncat - st_pos;
|
4130
|
+
size_t ncomb = pow2(ncat_present) - 1;
|
4131
|
+
size_t best_combin;
|
4132
|
+
|
4133
|
+
for (size_t combin = 1; combin < ncomb; combin++)
|
4134
|
+
{
|
4135
|
+
cnt_left = 0; cnt_right = 0;
|
4136
|
+
s_left = 0; s_right = 0;
|
4137
|
+
for (size_t pos = st_pos; (int)pos < ncat; pos++)
|
4138
|
+
{
|
4139
|
+
if (extract_bit(combin, pos))
|
4140
|
+
{
|
4141
|
+
cnt_left += buffer_cnt[buffer_pos[pos]];
|
4142
|
+
s_left += (buffer_cnt[buffer_pos[pos]] <= 1)?
|
4143
|
+
(ldouble_safe)0
|
4144
|
+
:
|
4145
|
+
((ldouble_safe) buffer_cnt[buffer_pos[pos]]
|
4146
|
+
* std::log((ldouble_safe) buffer_cnt[buffer_pos[pos]]));
|
4147
|
+
}
|
4148
|
+
|
4149
|
+
else
|
4150
|
+
{
|
4151
|
+
cnt_right += buffer_cnt[buffer_pos[pos]];
|
4152
|
+
s_right += (buffer_cnt[buffer_pos[pos]] <= 1)?
|
4153
|
+
(ldouble_safe)0
|
4154
|
+
:
|
4155
|
+
((ldouble_safe) buffer_cnt[buffer_pos[pos]]
|
4156
|
+
* std::log((ldouble_safe) buffer_cnt[buffer_pos[pos]]));
|
4157
|
+
}
|
4158
|
+
}
|
4159
|
+
|
4160
|
+
this_gain = categ_gain<size_t, ldouble_safe>(
|
4161
|
+
cnt_left, cnt_right,
|
4162
|
+
s_left, s_right,
|
4163
|
+
base_info, cnt);
|
4164
|
+
|
4165
|
+
if (this_gain > min_gain && this_gain > best_gain)
|
4166
|
+
{
|
4167
|
+
best_gain = this_gain;
|
4168
|
+
best_combin = combin;
|
4169
|
+
}
|
4170
|
+
|
4171
|
+
}
|
4172
|
+
|
4173
|
+
if (best_gain > min_gain)
|
4174
|
+
for (size_t pos = 0; pos < ncat_present; pos++)
|
4175
|
+
split_categ[buffer_pos[st_pos + pos]] = extract_bit(best_combin, pos);
|
4176
|
+
|
4177
|
+
}
|
4178
|
+
|
4179
|
+
else
|
4180
|
+
{
|
4181
|
+
/* try moving the categories one at a time */
|
4182
|
+
size_t cnt_left = 0;
|
4183
|
+
size_t cnt_right = end - st + 1;
|
4184
|
+
double s_left = 0;
|
4185
|
+
double s_right = s;
|
4186
|
+
|
4187
|
+
for (size_t pos = st_pos; pos < (ncat - st_pos - 1); pos++)
|
4188
|
+
{
|
4189
|
+
buffer_split[buffer_pos[pos]] = 1;
|
4190
|
+
s_left += (buffer_cnt[buffer_pos[pos]] <= 1)?
|
4191
|
+
(ldouble_safe)0
|
4192
|
+
:
|
4193
|
+
((ldouble_safe)buffer_cnt[buffer_pos[pos]] * std::log((ldouble_safe) buffer_cnt[buffer_pos[pos]]));
|
4194
|
+
s_right -= (buffer_cnt[buffer_pos[pos]] <= 1)?
|
4195
|
+
(ldouble_safe)0
|
4196
|
+
:
|
4197
|
+
((ldouble_safe)buffer_cnt[buffer_pos[pos]] * std::log((ldouble_safe) buffer_cnt[buffer_pos[pos]]));
|
4198
|
+
cnt_left += buffer_cnt[buffer_pos[pos]];
|
4199
|
+
cnt_right -= buffer_cnt[buffer_pos[pos]];
|
4200
|
+
|
4201
|
+
this_gain = categ_gain<size_t, ldouble_safe>(
|
4202
|
+
cnt_left, cnt_right,
|
4203
|
+
s_left, s_right,
|
4204
|
+
base_info, cnt);
|
4205
|
+
|
4206
|
+
if (this_gain > min_gain && this_gain > best_gain)
|
4207
|
+
{
|
4208
|
+
best_gain = this_gain;
|
4209
|
+
memcpy(split_categ, buffer_split, ncat * sizeof(signed char));
|
4210
|
+
}
|
4211
|
+
}
|
4212
|
+
}
|
4213
|
+
|
4214
|
+
break;
|
4215
|
+
}
|
4216
|
+
|
4217
|
+
default:
|
4218
|
+
{
|
4219
|
+
unexpected_error();
|
4220
|
+
break;
|
4221
|
+
}
|
4222
|
+
}
|
4223
|
+
}
|
4224
|
+
}
|
4225
|
+
|
4226
|
+
if (st == (end-1)) return 0;
|
4227
|
+
|
4228
|
+
if (best_gain <= -HUGE_VAL && this_gain <= min_gain && this_gain > -HUGE_VAL)
|
4229
|
+
return 0;
|
4230
|
+
else
|
4231
|
+
return best_gain;
|
4232
|
+
}
|