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 +1,2401 @@
|
|
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
|
+
/* Fit Isolation Forest model, or variant of it such as SCiForest
|
66
|
+
*
|
67
|
+
* Parameters:
|
68
|
+
* ===========
|
69
|
+
* - model_outputs (out)
|
70
|
+
* Pointer to already allocated isolation forest model object for single-variable splits.
|
71
|
+
* If fitting the extended model, pass NULL (must pass 'model_outputs_ext'). Can later add
|
72
|
+
* additional trees through function 'add_tree'.
|
73
|
+
* - model_outputs_ext (out)
|
74
|
+
* Pointer to already allocated extended isolation forest model object (for multiple-variable splits).
|
75
|
+
* Note that if 'ndim' = 1, must use instead the single-variable model object.
|
76
|
+
* If fitting the single-variable model, pass NULL (must pass 'model_outputs'). Can later add
|
77
|
+
* additional trees through function 'add_tree'.
|
78
|
+
* - numeric_data[nrows * ncols_numeric]
|
79
|
+
* Pointer to numeric data to which to fit the model. Must be ordered by columns like Fortran,
|
80
|
+
* not ordered by rows like C (i.e. entries 1..n contain column 0, n+1..2n column 1, etc.).
|
81
|
+
* Pass NULL if there are no dense numeric columns (must also pass 'ncols_numeric' = 0 if there's
|
82
|
+
* no sparse numeric data either).
|
83
|
+
* Can only pass one of 'numeric_data' or 'Xc' + 'Xc_ind' + 'Xc_indptr'.
|
84
|
+
* - ncols_numeric
|
85
|
+
* Number of numeric columns in the data (whether they come in a sparse matrix or dense array).
|
86
|
+
* - categ_data[nrows * ncols_categ]
|
87
|
+
* Pointer to categorical data to which to fit the model. Must be ordered by columns like Fortran,
|
88
|
+
* not ordered by rows like C (i.e. entries 1..n contain column 0, n+1..2n column 1, etc.).
|
89
|
+
* Pass NULL if there are no categorical columns (must also pass 'ncols_categ' = 0).
|
90
|
+
* Each category should be represented as an integer, and these integers must start at zero and
|
91
|
+
* be in consecutive order - i.e. if category '3' is present, category '2' must also be present
|
92
|
+
* (note that they are not treated as being ordinal, this is just an encoding). Missing values
|
93
|
+
* should be encoded as negative numbers such as (-1).
|
94
|
+
* - ncols_categ
|
95
|
+
* Number of categorical columns in the data.
|
96
|
+
* - ncat[ncols_categ]
|
97
|
+
* Number of categories in each categorical column. E.g. if the highest code for a column is '4',
|
98
|
+
* the number of categories for that column is '5' (zero is one category).
|
99
|
+
* - Xc[nnz]
|
100
|
+
* Pointer to numeric data in sparse numeric matrix in CSC format (column-compressed).
|
101
|
+
* Pass NULL if there are no sparse numeric columns.
|
102
|
+
* Can only pass one of 'numeric_data' or 'Xc' + 'Xc_ind' + 'Xc_indptr'.
|
103
|
+
* - Xc_ind[nnz]
|
104
|
+
* Pointer to row indices to which each non-zero entry in 'Xc' corresponds.
|
105
|
+
* Must be in sorted order, otherwise results will be incorrect.
|
106
|
+
* The largest value here should be smaller than the largest possible value of 'size_t'.
|
107
|
+
* Pass NULL if there are no sparse numeric columns.
|
108
|
+
* - Xc_indptr[ncols_numeric + 1]
|
109
|
+
* Pointer to column index pointers that tell at entry [col] where does column 'col'
|
110
|
+
* start and at entry [col + 1] where does column 'col' end.
|
111
|
+
* Pass NULL if there are no sparse numeric columns.
|
112
|
+
* - ndim
|
113
|
+
* How many dimensions (columns) to use for making a split. Must pass 'ndim' = 1 for
|
114
|
+
* the single-variable model. Note that the model object pointer passed must also
|
115
|
+
* agree with the value passed to 'ndim'.
|
116
|
+
* - ntry
|
117
|
+
* When using any of 'prob_pick_by_gain_pl', 'prob_pick_by_gain_avg', 'prob_pick_by_full_gain', 'prob_pick_by_dens', how many variables (with 'ndim=1')
|
118
|
+
* or linear combinations (with 'ndim>1') to try for determining the best one according to gain.
|
119
|
+
* Recommended value in reference [4] is 10 (with 'prob_pick_by_gain_avg', for outlier detection), while the
|
120
|
+
* recommended value in reference [11] is 1 (with 'prob_pick_by_gain_pl', for outlier detection), and the
|
121
|
+
* recommended value in reference [9] is 10 to 20 (with 'prob_pick_by_gain_pl', for missing value imputations).
|
122
|
+
* - coef_type
|
123
|
+
* For the extended model, whether to sample random coefficients according to a normal distribution ~ N(0, 1)
|
124
|
+
* (as proposed in [4]) or according to a uniform distribution ~ Unif(-1, +1) as proposed in [3]. Ignored for the
|
125
|
+
* single-variable model.
|
126
|
+
* - sample_weights[nrows]
|
127
|
+
* Weights for the rows when building a tree, either as sampling importances when using
|
128
|
+
* sub-samples for each tree (i.e. passing weight '2' makes a row twice as likely to be included
|
129
|
+
* in a random sub-sample), or as density measurement (i.e. passing weight '2' is the same as if
|
130
|
+
* the row appeared twice, thus it's less of an outlier) - how this is taken is determined
|
131
|
+
* through parameter 'weight_as_sample'.
|
132
|
+
* Pass NULL if the rows all have uniform weights.
|
133
|
+
* - with_replacement
|
134
|
+
* Whether to sample rows with replacement or not (not recommended). Note that distance calculations,
|
135
|
+
* if desired, don't work well with duplicate rows.
|
136
|
+
* - weight_as_sample
|
137
|
+
* If passing sample (row) weights when fitting the model, whether to consider those weights as row
|
138
|
+
* sampling weights (i.e. the higher the weights, the more likely the observation will end up included
|
139
|
+
* in each tree sub-sample), or as distribution density weights (i.e. putting a weight of two is the same
|
140
|
+
* as if the row appeared twice, thus higher weight makes it less of an outlier, but does not give it a
|
141
|
+
* higher chance of being sampled if the data uses sub-sampling).
|
142
|
+
* - nrows
|
143
|
+
* Number of rows in 'numeric_data', 'Xc', 'categ_data'.
|
144
|
+
* - sample_size
|
145
|
+
* Sample size of the data sub-samples with which each binary tree will be built. When a terminal node has more than
|
146
|
+
* 1 observation, the remaining isolation depth for them is estimated assuming the data and splits are both uniformly
|
147
|
+
* random (separation depth follows a similar process with expected value calculated as in [6]). If passing zero,
|
148
|
+
* will set it to 'nrows'. Recommended value in [1], [2], [3] is 256, while the default value in the author's code
|
149
|
+
* in [5] is 'nrows' here.
|
150
|
+
* - ntrees
|
151
|
+
* Number of binary trees to build for the model. Recommended value in [1] is 100, while the default value in the
|
152
|
+
* author's code in [5] is 10.
|
153
|
+
* - max_depth
|
154
|
+
* Maximum depth of the binary trees to grow. Will get overwritten if passing 'limit_depth' = 'true'.
|
155
|
+
* Models that use 'prob_pick_by_gain_pl' or 'prob_pick_by_gain_avg' are likely to benefit from
|
156
|
+
* deeper trees (larger 'max_depth'), but deeper trees can result in much slower model fitting and
|
157
|
+
* predictions.
|
158
|
+
* Note that models that use 'prob_pick_by_gain_pl' or 'prob_pick_by_gain_avg' are likely to benefit from
|
159
|
+
* deeper trees (larger 'max_depth'), but deeper trees can result in much slower model fitting and
|
160
|
+
* predictions.
|
161
|
+
* If using pooled gain, one might want to substitute 'max_depth' with 'min_gain'.
|
162
|
+
* - ncols_per_tree
|
163
|
+
* Number of columns to use (have as potential candidates for splitting at each iteration) in each tree,
|
164
|
+
* similar to the 'mtry' parameter of random forests.
|
165
|
+
* In general, this is only relevant when using non-random splits and/or weighted column choices.
|
166
|
+
* If passing zero, will use the full number of available columns.
|
167
|
+
* Recommended value: 0.
|
168
|
+
* - limit_depth
|
169
|
+
* Whether to automatically set the maximum depth to the corresponding depth of a balanced binary tree with number of
|
170
|
+
* terminal nodes corresponding to the sub-sample size (the reason being that, if trying to detect outliers, an outlier
|
171
|
+
* will only be so if it turns out to be isolated with shorter average depth than usual, which corresponds to a balanced
|
172
|
+
* tree depth). Default setting for [1], [2], [3], [4] is 'true', but it's recommended to pass 'false' here
|
173
|
+
* and higher values for 'max_depth' if using the model for purposes other than outlier detection.
|
174
|
+
* Note that, if passing 'limit_depth=true', then 'max_depth' is ignored.
|
175
|
+
* - penalize_range
|
176
|
+
* Whether to penalize (add -1 to the terminal depth) observations at prediction time that have a value
|
177
|
+
* of the chosen split variable (linear combination in extended model) that falls outside of a pre-determined
|
178
|
+
* reasonable range in the data being split (given by 2 * range in data and centered around the split point),
|
179
|
+
* as proposed in [4] and implemented in the authors' original code in [5]. Not used in single-variable model
|
180
|
+
* when splitting by categorical variables. Note that this can make a very large difference in the results
|
181
|
+
* when using 'prob_pick_by_gain_pl'.
|
182
|
+
* This option is not supported when using density-based outlier scoring metrics.
|
183
|
+
* - standardize_data
|
184
|
+
* Whether to standardize the features at each node before creating a linear combination of them as suggested
|
185
|
+
* in [4]. This is ignored when using 'ndim=1'.
|
186
|
+
* - scoring_metric
|
187
|
+
* Metric to use for determining outlier scores (see reference [13]).
|
188
|
+
* If passing 'Depth', will use isolation depth as proposed in reference [1]. This is typically the safest choice
|
189
|
+
* and plays well with all model types offered by this library.
|
190
|
+
* If passing 'Density', will set scores for each terminal node as the ratio between the fraction of points in the sub-sample
|
191
|
+
* that end up in that node and the fraction of the volume in the feature space which defines
|
192
|
+
* the node according to the splits that lead to it.
|
193
|
+
* If using 'ndim=1', for categorical variables, 'Density' is defined in terms
|
194
|
+
* of number of categories that go towards each side of the split divided by number of categories
|
195
|
+
* in the observations that reached that node.
|
196
|
+
* The standardized outlier score from 'Density' for a given observation is calculated as the
|
197
|
+
* negative of the logarithm of the geometric mean from the per-tree densities, which unlike
|
198
|
+
* the standardized score produced from 'Depth', is unbounded, but just like the standardized
|
199
|
+
* score form 'Depth', has a natural threshold for definining outlierness, which in this case
|
200
|
+
* is zero is instead of 0.5. The non-standardized outlier score for 'Density' is calculated as the
|
201
|
+
* geometric mean, while the per-tree scores are calculated as the density values.
|
202
|
+
* 'Density' might lead to better predictions when using 'ndim=1', particularly in the presence
|
203
|
+
* of categorical variables. Note however that using 'Density' requires more trees for convergence
|
204
|
+
* of scores (i.e. good results) compared to isolation-based metrics.
|
205
|
+
* 'Density' is incompatible with 'penalize_range=true'.
|
206
|
+
* If passing 'AdjDepth', will use an adjusted isolation depth that takes into account the number of points that
|
207
|
+
* go to each side of a given split vs. the fraction of the range of that feature that each
|
208
|
+
* side of the split occupies, by a metric as follows: 'd = 2/ (1 + 1/(2*p))'
|
209
|
+
* where 'p' is defined as 'p = (n_s / n_t) / (r_s / r_t)
|
210
|
+
* with 'n_t' being the number of points that reach a given node, 'n_s' the
|
211
|
+
* number of points that are sent to a given side of the split/branch at that node,
|
212
|
+
* 'r_t' being the range (maximum minus minimum) of the splitting feature or
|
213
|
+
* linear combination among the points that reached the node, and 'r_s' being the
|
214
|
+
* range of the same feature or linear combination among the points that are sent to this
|
215
|
+
* same side of the split/branch. This makes each split add a number between zero and two
|
216
|
+
* to the isolation depth, with this number's probabilistic distribution being centered
|
217
|
+
* around 1 and thus the expected isolation depth remaing the same as in the original
|
218
|
+
* 'Depth' metric, but having more variability around the extremes.
|
219
|
+
* Scores (standardized, non-standardized, per-tree) for 'AdjDepth' are aggregated in the same way
|
220
|
+
* as for 'Depth'.
|
221
|
+
* 'AdjDepth' might lead to better predictions when using 'ndim=1', particularly in the prescence
|
222
|
+
* of categorical variables and for smaller datasets, and for smaller datasets, might make
|
223
|
+
* sense to combine it with 'penalize_range=true'.
|
224
|
+
* If passing 'AdjDensity', will use the same metric from 'AdjDepth', but applied multiplicatively instead
|
225
|
+
* of additively. The expected value for 'AdjDepth' is not strictly the same
|
226
|
+
* as for isolation, but using the expected isolation depth as standardizing criterion
|
227
|
+
* tends to produce similar standardized score distributions (centered around 0.5).
|
228
|
+
* Scores (standardized, non-standardized, per-tree) from 'AdjDensity' are aggregated in the same way
|
229
|
+
* as for 'Depth'.
|
230
|
+
* 'AdjDepth' is incompatible with 'penalize_range=true'.
|
231
|
+
* If passing 'BoxedRatio', will set the scores for each terminal node as the ratio between the volume of the boxed
|
232
|
+
* feature space for the node as defined by the smallest and largest values from the split
|
233
|
+
* conditions for each column (bounded by the variable ranges in the sample) and the
|
234
|
+
* variable ranges in the tree sample.
|
235
|
+
* If using 'ndim=1', for categorical variables 'BoxedRatio' is defined in terms of number of categories.
|
236
|
+
* If using 'ndim=>1', 'BoxedRatio' is defined in terms of the maximum achievable value for the
|
237
|
+
* splitting linear combination determined from the minimum and maximum values for each
|
238
|
+
* variable among the points in the sample, and as such, it has a rather different meaning
|
239
|
+
* compared to the score obtained with 'ndim=1' - 'BoxedRatio' scores with 'ndim>1'
|
240
|
+
* typically provide very poor quality results and this metric is thus not recommended to
|
241
|
+
* use in the extended model. With 'ndim>1', 'BoxedRatio' also has a tendency of producing too small
|
242
|
+
* values which round to zero.
|
243
|
+
* The standardized outlier score from 'BoxedRatio' for a given observation is calculated
|
244
|
+
* simply as the the average from the per-tree boxed ratios. 'BoxedRatio' metric
|
245
|
+
* has a lower bound of zero and a theorical upper bound of one, but in practice the scores
|
246
|
+
* tend to be very small numbers close to zero, and its distribution across
|
247
|
+
* different datasets is rather unpredictable. In order to keep rankings comparable with
|
248
|
+
* the rest of the metrics, the non-standardized outlier scores for 'BoxedRatio' are calculated as the
|
249
|
+
* negative of the average instead. The per-tree 'BoxedRatio' scores are calculated as the ratios.
|
250
|
+
* 'BoxedRatio' can be calculated in a fast-but-not-so-precise way, and in a low-but-precise
|
251
|
+
* way, which is controlled by parameter 'fast_bratio'. Usually, both should give the
|
252
|
+
* same results, but in some fatasets, the fast way can lead to numerical inaccuracies
|
253
|
+
* due to roundoffs very close to zero.
|
254
|
+
* 'BoxedRatio' might lead to better predictions in datasets with many rows when using 'ndim=1'
|
255
|
+
* and a relatively small 'sample_size'. Note that more trees are required for convergence
|
256
|
+
* of scores when using 'BoxedRatio'. In some datasets, 'BoxedRatio' metric might result in very bad
|
257
|
+
* predictions, to the point that taking its inverse produces a much better ranking of outliers.
|
258
|
+
* 'BoxedRatio' option is incompatible with 'penalize_range'.
|
259
|
+
* If passing 'BoxedDensity2', will set the score as the ratio between the fraction of points within the sample that
|
260
|
+
* end up in a given terminal node and the 'BoxedRatio' metric.
|
261
|
+
* Aggregation of scores (standardized, non-standardized, per-tree) for 'BoxedDensity2' is done in the same
|
262
|
+
* way as for 'Density', and it also has a natural threshold at zero for determining
|
263
|
+
* outliers and inliers.
|
264
|
+
* 'BoxedDensity2' is typically usable with 'ndim>1', but tends to produce much bigger values
|
265
|
+
* compared to 'ndim=1'.
|
266
|
+
* Albeit unintuitively, in many datasets, one can usually get better results with metric
|
267
|
+
* 'BoxedDensity' instead.
|
268
|
+
* The calculation of 'BoxedDensity2' is also controlled by 'fast_bratio'.
|
269
|
+
* 'BoxedDensity2' incompatible with 'penalize_range'.
|
270
|
+
* If passing 'BoxedDensity', will set the score as the ratio between the fraction of points within the sample that
|
271
|
+
* end up in a given terminal node and the ratio between the boxed volume of the feature
|
272
|
+
* space in the sample and the boxed volume of a node given by the split conditions (inverse
|
273
|
+
* as in 'BoxedDensity2'). This metric does not have any theoretical or intuitive
|
274
|
+
* justification behind its existence, and it is perhaps ilogical to use it as a
|
275
|
+
* scoring metric, but tends to produce good results in some datasets.
|
276
|
+
* The standardized outlier scores for 'BoxedDensity' are defined as the negative of the geometric mean,
|
277
|
+
* while the non-standardized scores are the geometric mean, and the per-tree scores are simply the 'density' values.
|
278
|
+
* The calculation of 'BoxedDensity' is also controlled by 'fast_bratio'.
|
279
|
+
* 'BoxedDensity' option is incompatible with 'penalize_range'.
|
280
|
+
* - fast_bratio
|
281
|
+
* When using "boxed" metrics for scoring, whether to calculate them in a fast way through
|
282
|
+
* cumulative sum of logarithms of ratios after each split, or in a slower way as sum of
|
283
|
+
* logarithms of a single ratio per column for each terminal node.
|
284
|
+
* Usually, both methods should give the same results, but in some datasets, particularly
|
285
|
+
* when variables have too small or too large ranges, the first method can be prone to
|
286
|
+
* numerical inaccuracies due to roundoff close to zero.
|
287
|
+
* Note that this does not affect calculations for models with 'ndim>1', since given the
|
288
|
+
* split types, the calculation for them is different.
|
289
|
+
* - standardize_dist
|
290
|
+
* If passing 'tmat' (see documentation for it), whether to standardize the resulting average separation
|
291
|
+
* depths in order to produce a distance metric or not, in the same way this is done for the outlier score.
|
292
|
+
* - tmat[nrows * (nrows - 1) / 2]
|
293
|
+
* Array in which to calculate average separation depths or standardized distance metric (see documentation
|
294
|
+
* for 'standardize_dist') as the model is being fit. Pass NULL to avoid doing these calculations alongside
|
295
|
+
* the regular model process. If passing this output argument, the sample size must be the same as the number
|
296
|
+
* of rows, and there cannot be sample weights. If not NULL, must already be initialized to zeros. As the
|
297
|
+
* output is a symmetric matrix, this function will only fill in the upper-triangular part, in which
|
298
|
+
* entry 0 <= i < j < n will be located at position
|
299
|
+
* p(i,j) = (i * (n - (i+1)/2) + j - i - 1).
|
300
|
+
* Can be converted to a dense square matrix through function 'tmat_to_dense'.
|
301
|
+
* - output_depths[nrows]
|
302
|
+
* Array in which to calculate average path depths or standardized outlierness metric (see documentation
|
303
|
+
* for 'standardize_depth') as the model is being fit. Pass NULL to avoid doing these calculations alongside
|
304
|
+
* the regular model process. If passing this output argument, the sample size must be the same as the number
|
305
|
+
* of rows. If not NULL, must already be initialized to zeros.
|
306
|
+
* - standardize_depth
|
307
|
+
* If passing 'output_depths', whether to standardize the results as proposed in [1], in order to obtain
|
308
|
+
* a metric in which the more outlier is an observation, the closer this standardized metric will be to 1,
|
309
|
+
* with average observations obtaining 0.5. If passing 'false' here, the numbers in 'output_depths' will be
|
310
|
+
* the average depth of each row across all trees.
|
311
|
+
* - col_weights[ncols_numeric + ncols_categ]
|
312
|
+
* Sampling weights for each column, assuming all the numeric columns come before the categorical columns.
|
313
|
+
* Ignored when picking columns by deterministic criterion.
|
314
|
+
* If passing NULL, each column will have a uniform weight. If used along with kurtosis weights, the
|
315
|
+
* effect is multiplicative.
|
316
|
+
* - weigh_by_kurt
|
317
|
+
* Whether to weigh each column according to the kurtosis obtained in the sub-sample that is selected
|
318
|
+
* for each tree as briefly proposed in [1]. Note that this is only done at the beginning of each tree
|
319
|
+
* sample. For categorical columns, will calculate expected kurtosis if the column were converted to
|
320
|
+
* numerical by assigning to each category a random number ~ Unif(0, 1).
|
321
|
+
* This is intended as a cheap feature selector, while the parameter 'prob_pick_col_by_kurt'
|
322
|
+
* provides the option to do this at each node in the tree for a different overall type of model.
|
323
|
+
* If passing column weights or weighted column choices ('prob_pick_col_by_range', 'prob_pick_col_by_var'),
|
324
|
+
* the effect will be multiplicative. This option is not compatible with 'prob_pick_col_by_kurt'.
|
325
|
+
* If passing 'missing_action=fail' and the data has infinite values, columns with rows
|
326
|
+
* having infinite values will get a weight of zero. If passing a different value for missing
|
327
|
+
* action, infinite values will be ignored in the kurtosis calculation.
|
328
|
+
* If using 'missing_action=Impute', the calculation of kurtosis will not use imputed values
|
329
|
+
* in order not to favor columns with missing values (which would increase kurtosis by all having
|
330
|
+
* the same central value).
|
331
|
+
* - prob_pick_by_gain_pl
|
332
|
+
* This parameter indicates the probability of choosing the threshold on which to split a variable
|
333
|
+
* (with 'ndim=1') or a linear combination of variables (when using 'ndim>1') as the threshold
|
334
|
+
* that maximizes a pooled standard deviation gain criterion (see references [9] and [11]) on the
|
335
|
+
* same variable or linear combination, similarly to regression trees such as CART.
|
336
|
+
* If using 'ntry>1', will try several variables or linear combinations thereof and choose the one
|
337
|
+
* in which the largest standardized gain can be achieved.
|
338
|
+
* For categorical variables with 'ndim=1', will use shannon entropy instead (like in [7]).
|
339
|
+
* Compared to a simple averaged gain, this tends to result in more evenly-divided splits and more clustered
|
340
|
+
* groups when they are smaller. Recommended to pass higher values when used for imputation of missing values.
|
341
|
+
* When used for outlier detection, datasets with multimodal distributions usually see better performance
|
342
|
+
* under this type of splits.
|
343
|
+
* Note that, since this makes the trees more even and thus it takes more steps to produce isolated nodes,
|
344
|
+
* the resulting object will be heavier. When splits are not made according to any of 'prob_pick_by_gain_avg',
|
345
|
+
* 'prob_pick_by_gain_pl', 'prob_pick_by_full_gain', 'prob_pick_by_dens', both the column and the split point are decided at random.
|
346
|
+
* Note that, if passing value 1 (100%) with no sub-sampling and using the single-variable model,
|
347
|
+
* every single tree will have the exact same splits.
|
348
|
+
* Be aware that 'penalize_range' can also have a large impact when using 'prob_pick_by_gain_pl'.
|
349
|
+
* Be aware also that, if passing a value of 1 (100%) with no sub-sampling and using the single-variable
|
350
|
+
* model, every single tree will have the exact same splits.
|
351
|
+
* Under this option, models are likely to produce better results when increasing 'max_depth'.
|
352
|
+
* Alternatively, one can also control the depth through 'min_gain' (for which one might want to
|
353
|
+
* set 'max_depth=0').
|
354
|
+
* Important detail: if using any of 'prob_pick_by_gain_avg', 'prob_pick_by_gain_pl', 'prob_pick_by_full_gain',
|
355
|
+
* 'prob_pick_by_dens', the distribution of outlier scores is unlikely to be centered around 0.5.
|
356
|
+
* - prob_pick_by_gain_avg
|
357
|
+
* This parameter indicates the probability of choosing the threshold on which to split a variable
|
358
|
+
* (with 'ndim=1') or a linear combination of variables (when using 'ndim>1') as the threshold
|
359
|
+
* that maximizes an averaged standard deviation gain criterion (see references [4] and [11]) on the
|
360
|
+
* same variable or linear combination.
|
361
|
+
* If using 'ntry>1', will try several variables or linear combinations thereof and choose the one
|
362
|
+
* in which the largest standardized gain can be achieved.
|
363
|
+
* For categorical variables with 'ndim=1', will take the expected standard deviation that would be
|
364
|
+
* gotten if the column were converted to numerical by assigning to each category a random
|
365
|
+
* number ~ Unif(0, 1) and calculate gain with those assumed standard deviations.
|
366
|
+
* Compared to a pooled gain, this tends to result in more cases in which a single observation or very
|
367
|
+
* few of them are put into one branch. Typically, datasets with outliers defined by extreme values in
|
368
|
+
* some column more or less independently of the rest, usually see better performance under this type
|
369
|
+
* of split. Recommended to use sub-samples (parameter 'sample_size') when
|
370
|
+
* passing this parameter. Note that, since this will create isolated nodes faster, the resulting object
|
371
|
+
* will be lighter (use less memory).
|
372
|
+
* When splits are not made according to any of 'prob_pick_by_gain_avg', 'prob_pick_by_gain_pl',
|
373
|
+
* 'prob_pick_by_full_gain', 'prob_pick_by_dens', both the column and the split point are decided at random.
|
374
|
+
* Default setting for [1], [2], [3] is zero, and default for [4] is 1.
|
375
|
+
* This is the randomization parameter that can be passed to the author's original code in [5],
|
376
|
+
* but note that the code in [5] suffers from a mathematical error in the calculation of running standard deviations,
|
377
|
+
* so the results from it might not match with this library's.
|
378
|
+
* Be aware that, if passing a value of 1 (100%) with no sub-sampling and using the single-variable model,
|
379
|
+
* every single tree will have the exact same splits.
|
380
|
+
* Under this option, models are likely to produce better results when increasing 'max_depth'.
|
381
|
+
* Important detail: if using any of 'prob_pick_by_gain_avg', 'prob_pick_by_gain_pl',
|
382
|
+
* 'prob_pick_by_full_gain', 'prob_pick_by_dens', the distribution of outlier scores is unlikely to be centered around 0.5.
|
383
|
+
* - prob_pick_by_full_gain
|
384
|
+
* This parameter indicates the probability of choosing the threshold on which to split a variable
|
385
|
+
* (with 'ndim=1') or a linear combination of variables (when using 'ndim>1') as the threshold
|
386
|
+
* that minimizes the pooled sums of variances of all columns (or a subset of them if using
|
387
|
+
* 'ncols_per_tree').
|
388
|
+
* In general, 'prob_pick_by_full_gain' is much slower to evaluate than the other gain types, and does not tend to
|
389
|
+
* lead to better results. When using 'prob_pick_by_full_gain', one might want to use a different scoring
|
390
|
+
* metric (particulatly 'Density', 'BoxedDensity2' or 'BoxedRatio'). Note that
|
391
|
+
* the variance calculations are all done through the (exact) sorted-indices approach, while is much
|
392
|
+
* slower than the (approximate) histogram approach used by other decision tree software.
|
393
|
+
* Be aware that the data is not standardized in any way for the range calculations, thus the scales
|
394
|
+
* of features will make a large difference under 'prob_pick_by_full_gain', which might not make it suitable for
|
395
|
+
* all types of data.
|
396
|
+
* 'prob_pick_by_full_gain' is not compatible with categorical data, and 'min_gain' does not apply to it.
|
397
|
+
* When splits are not made according to any of 'prob_pick_by_gain_avg', 'prob_pick_by_gain_pl',
|
398
|
+
* 'prob_pick_by_full_gain', 'prob_pick_by_dens', both the column and the split point are decided at random.
|
399
|
+
* Default setting for [1], [2], [3], [4] is zero.
|
400
|
+
* - prob_pick_dens
|
401
|
+
* This parameter indicates the probability of choosing the threshold on which to split a variable
|
402
|
+
* (with 'ndim=1') or a linear combination of variables (when using 'ndim>1') as the threshold
|
403
|
+
* that maximizes the pooled densities of the branch distributions.
|
404
|
+
* The 'min_gain' option does not apply to this type of splits.
|
405
|
+
* When splits are not made according to any of 'prob_pick_by_gain_avg', 'prob_pick_by_gain_pl',
|
406
|
+
* 'prob_pick_by_full_gain', 'prob_pick_by_dens', both the column and the split point are decided at random.
|
407
|
+
* Default setting for [1], [2], [3], [4] is zero.
|
408
|
+
* - prob_pick_col_by_range
|
409
|
+
* When using 'ndim=1', this denotes the probability of choosing the column to split with a probability
|
410
|
+
* proportional to the range spanned by each column within a node as proposed in reference [12].
|
411
|
+
* When using 'ndim>1', this denotes the probability of choosing columns to create a hyperplane with a
|
412
|
+
* probability proportional to the range spanned by each column within a node.
|
413
|
+
* This option is not compatible with categorical data. If passing column weights, the
|
414
|
+
* effect will be multiplicative.
|
415
|
+
* Be aware that the data is not standardized in any way for the range calculations, thus the scales
|
416
|
+
* of features will make a large difference under this option, which might not make it suitable for
|
417
|
+
* all types of data.
|
418
|
+
* Note that the proposed RRCF model from [12] uses a different scoring metric for producing anomaly
|
419
|
+
* scores, while this library uses isolation depth regardless of how columns are chosen, thus results
|
420
|
+
* are likely to be different from those of other software implementations. Nevertheless, as explored
|
421
|
+
* in [11], isolation depth as a scoring metric typically provides better results than the
|
422
|
+
* "co-displacement" metric from [12] under these split types.
|
423
|
+
* - prob_pick_col_by_var
|
424
|
+
* When using 'ndim=1', this denotes the probability of choosing the column to split with a probability
|
425
|
+
* proportional to the variance of each column within a node.
|
426
|
+
* When using 'ndim>1', this denotes the probability of choosing columns to create a hyperplane with a
|
427
|
+
* probability proportional to the variance of each column within a node.
|
428
|
+
* For categorical data, it will calculate the expected variance if the column were converted to
|
429
|
+
* numerical by assigning to each category a random number ~ Unif(0, 1), which depending on the number of
|
430
|
+
* categories and their distribution, produces numbers typically a bit smaller than standardized numerical
|
431
|
+
* variables.
|
432
|
+
* Note that when using sparse matrices, the calculation of variance will rely on a procedure that
|
433
|
+
* uses sums of squares, which has less numerical precision than the
|
434
|
+
* calculation used for dense inputs, and as such, the results might differ slightly.
|
435
|
+
* Be aware that this calculated variance is not standardized in any way, so the scales of
|
436
|
+
* features will make a large difference under this option.
|
437
|
+
* If there are infinite values, all columns having infinite values will be treated as having the
|
438
|
+
* same weight, and will be chosen before every other column with non-infinite values.
|
439
|
+
* If passing column weights , the effect will be multiplicative.
|
440
|
+
* If passing a 'missing_action' different than 'fail', infinite values will be ignored for the
|
441
|
+
* variance calculation. Otherwise, all columns with infinite values will have the same probability
|
442
|
+
* and will be chosen before columns with non-infinite values.
|
443
|
+
* - prob_pick_col_by_kurt
|
444
|
+
* When using 'ndim=1', this denotes the probability of choosing the column to split with a probability
|
445
|
+
* proportional to the kurtosis of each column **within a node** (unlike the option 'weigh_by_kurtosis'
|
446
|
+
* which calculates this metric only at the root).
|
447
|
+
* When using 'ndim>1', this denotes the probability of choosing columns to create a hyperplane with a
|
448
|
+
* probability proportional to the kurtosis of each column within a node.
|
449
|
+
* For categorical data, it will calculate the expected kurtosis if the column were converted to
|
450
|
+
* numerical by assigning to each category a random number ~ Unif(0, 1).
|
451
|
+
* Note that when using sparse matrices, the calculation of kurtosis will rely on a procedure that
|
452
|
+
* uses sums of squares and higher-power numbers, which has less numerical precision than the
|
453
|
+
* calculation used for dense inputs, and as such, the results might differ slightly.
|
454
|
+
* If passing column weights, the effect will be multiplicative. This option is not compatible
|
455
|
+
* with 'weigh_by_kurtosis'.
|
456
|
+
* If passing a 'missing_action' different than 'fail', infinite values will be ignored for the
|
457
|
+
* variance calculation. Otherwise, all columns with infinite values will have the same probability
|
458
|
+
* and will be chosen before columns with non-infinite values.
|
459
|
+
* If using 'missing_action=Impute', the calculation of kurtosis will not use imputed values
|
460
|
+
* in order not to favor columns with missing values (which would increase kurtosis by all having
|
461
|
+
* the same central value).
|
462
|
+
* Be aware that kurtosis can be a rather slow metric to calculate.
|
463
|
+
* - min_gain
|
464
|
+
* Minimum gain that a split threshold needs to produce in order to proceed with a split.
|
465
|
+
* Only used when the splits are decided by a variance gain criterion ('prob_pick_by_gain_pl' or
|
466
|
+
* 'prob_pick_by_gain_avg', but not 'prob_pick_by_full_gain' nor 'prob_pick_by_dens').
|
467
|
+
* If the highest possible gain in the evaluated splits at a node is below this threshold,
|
468
|
+
* that node becomes a terminal node.
|
469
|
+
* This can be used as a more sophisticated depth control when using pooled gain (note that 'max_depth'
|
470
|
+
* still applies on top of this heuristic).
|
471
|
+
* - missing_action
|
472
|
+
* How to handle missing data at both fitting and prediction time. Options are a) 'Divide' (for the single-variable
|
473
|
+
* model only, recommended), which will follow both branches and combine the result with the weight given by the fraction of
|
474
|
+
* the data that went to each branch when fitting the model, b) 'Impute', which will assign observations to the
|
475
|
+
* branch with the most observations in the single-variable model (but imputed values will also be used for
|
476
|
+
* gain calculations), or fill in missing values with the median of each column of the sample from which the
|
477
|
+
* split was made in the extended model (recommended) (but note that the calculation of medians does not take
|
478
|
+
* into account sample weights when using 'weights_as_sample_prob=false', and note that when using a gain
|
479
|
+
* criterion for splits with 'ndim=1', it will use the imputed values in the calculation), c) 'Fail' which will
|
480
|
+
* assume that there are no missing values and will trigger undefined behavior if it encounters any.
|
481
|
+
* In the extended model, infinite values will be treated as missing.
|
482
|
+
* Note that passing 'Fail' might crash the process if there turn out to be missing values, but will otherwise
|
483
|
+
* produce faster fitting and prediction times along with decreased model object sizes.
|
484
|
+
* Models from [1], [2], [3], [4] correspond to 'Fail' here.
|
485
|
+
* - cat_split_type
|
486
|
+
* Whether to split categorical features by assigning sub-sets of them to each branch, or by assigning
|
487
|
+
* a single category to a branch and the rest to the other branch. For the extended model, whether to
|
488
|
+
* give each category a coefficient, or only one while the rest get zero.
|
489
|
+
* - new_cat_action
|
490
|
+
* What to do after splitting a categorical feature when new data that reaches that split has categories that
|
491
|
+
* the sub-sample from which the split was done did not have. Options are a) "Weighted" (recommended), which
|
492
|
+
* in the single-variable model will follow both branches and combine the result with weight given by the fraction of the
|
493
|
+
* data that went to each branch when fitting the model, and in the extended model will assign
|
494
|
+
* them the median value for that column that was added to the linear combination of features (but note that
|
495
|
+
* this median calculation does not use sample weights when using 'weights_as_sample_prob=false'),
|
496
|
+
* b) "Smallest", which will assign all observations with unseen categories in the split to the branch that
|
497
|
+
* had fewer observations when fitting the model, c) "Random", which will assing a branch (coefficient in the
|
498
|
+
* extended model) at random for each category beforehand, even if no observations had that category when
|
499
|
+
* fitting the model. Ignored when passing 'cat_split_type' = 'SingleCateg'.
|
500
|
+
* - all_perm
|
501
|
+
* When doing categorical variable splits by pooled gain with 'ndim=1' (regular model),
|
502
|
+
* whether to consider all possible permutations of variables to assign to each branch or not. If 'false',
|
503
|
+
* will sort the categories by their frequency and make a grouping in this sorted order. Note that the
|
504
|
+
* number of combinations evaluated (if 'true') is the factorial of the number of present categories in
|
505
|
+
* a given column (minus 2). For averaged gain, the best split is always to put the second most-frequent
|
506
|
+
* category in a separate branch, so not evaluating all permutations (passing 'false') will make it
|
507
|
+
* possible to select other splits that respect the sorted frequency order.
|
508
|
+
* The total number of combinations must be a number that can fit into a 'size_t' variable - for x64-64
|
509
|
+
* systems, this means no column can have more than 20 different categories if using 'all_perm=true',
|
510
|
+
* but note that this is not checked within the function.
|
511
|
+
* Ignored when not using categorical variables or not doing splits by pooled gain or using 'ndim>1'.
|
512
|
+
* - coef_by_prop
|
513
|
+
* In the extended model, whether to sort the randomly-generated coefficients for categories
|
514
|
+
* according to their relative frequency in the tree node. This might provide better results when using
|
515
|
+
* categorical variables with too many categories, but is not recommended, and not reflective of
|
516
|
+
* real "categorical-ness". Ignored for the regular model ('ndim=1') and/or when not using categorical
|
517
|
+
* variables.
|
518
|
+
* - imputer (out)
|
519
|
+
* Pointer to already-allocated imputer object, which can be used to produce missing value imputations
|
520
|
+
* in new data. Pass NULL if no missing value imputations are required. Note that this is not related to
|
521
|
+
* 'missing_action' as missing values inside the model are treated differently and follow their own imputation
|
522
|
+
* or division strategy.
|
523
|
+
* - min_imp_obs
|
524
|
+
* Minimum number of observations with which an imputation value can be produced. Ignored if passing
|
525
|
+
* 'build_imputer' = 'false'.
|
526
|
+
* - depth_imp
|
527
|
+
* How to weight observations according to their depth when used for imputing missing values. Passing
|
528
|
+
* "Higher" will weigh observations higher the further down the tree (away from the root node) the
|
529
|
+
* terminal node is, while "lower" will do the opposite, and "Sane" will not modify the weights according
|
530
|
+
* to node depth in the tree. Implemented for testing purposes and not recommended to change
|
531
|
+
* from the default. Ignored when not passing 'impute_nodes'.
|
532
|
+
* - weigh_imp_rows
|
533
|
+
* How to weight node sizes when used for imputing missing values. Passing "Inverse" will weigh
|
534
|
+
* a node inversely proportional to the number of observations that end up there, while "Proportional"
|
535
|
+
* will weight them heavier the more observations there are, and "Flat" will weigh all nodes the same
|
536
|
+
* in this regard regardless of how many observations end up there. Implemented for testing purposes
|
537
|
+
* and not recommended to change from the default. Ignored when not passing 'impute_nodes'.
|
538
|
+
* - impute_at_fit
|
539
|
+
* Whether to impute missing values in the input data as the model is being built. If passing 'true',
|
540
|
+
* then 'sample_size' must be equal to 'nrows'. Values in the arrays passed to 'numeric_data',
|
541
|
+
* 'categ_data', and 'Xc', will get overwritten with the imputations produced.
|
542
|
+
* - random_seed
|
543
|
+
* Seed that will be used to generate random numbers used by the model.
|
544
|
+
* - use_long_double
|
545
|
+
* Whether to use 'long double' (extended precision) type for more precise calculations about
|
546
|
+
* standard deviations, means, ratios, weights, gain, and other potential aggregates. This makes
|
547
|
+
* such calculations accurate to a larger number of decimals (provided that the compiler used has
|
548
|
+
* wider long doubles than doubles) and it is highly recommended to use when the input data has
|
549
|
+
* a number of rows or columns exceeding 2^53 (an unlikely scenario), and also highly recommended
|
550
|
+
* to use when the input data has problematic scales (e.g. numbers that differ from each other by
|
551
|
+
* something like 10^-100 or columns that include values like 10^100 and 10^-100 and still need to
|
552
|
+
* be sensitive to a difference of 10^-100), but will make the calculations slower, the more so in
|
553
|
+
* platforms in which 'long double' is a software-emulated type (e.g. Power8 platforms).
|
554
|
+
* Note that some platforms (most notably windows with the msvc compiler) do not make any difference
|
555
|
+
* between 'double' and 'long double'.
|
556
|
+
* - nthreads
|
557
|
+
* Number of parallel threads to use. Note that, the more threads, the more memory will be
|
558
|
+
* allocated, even if the thread does not end up being used.
|
559
|
+
* Be aware that most of the operations are bound by memory bandwidth, which means that
|
560
|
+
* adding more threads will not result in a linear speed-up. For some types of data
|
561
|
+
* (e.g. large sparse matrices with small sample sizes), adding more threads might result
|
562
|
+
* in only a very modest speed up (e.g. 1.5x faster with 4x more threads),
|
563
|
+
* even if all threads look fully utilized.
|
564
|
+
* Ignored when not building with OpenMP support.
|
565
|
+
*
|
566
|
+
* Returns
|
567
|
+
* =======
|
568
|
+
* Will return macro 'EXIT_SUCCESS' (typically =0) upon completion.
|
569
|
+
* If the process receives an interrupt signal, will return instead
|
570
|
+
* 'EXIT_FAILURE' (typically =1). If you do not have any way of determining
|
571
|
+
* what these values correspond to, you can use the functions
|
572
|
+
* 'return_EXIT_SUCESS' and 'return_EXIT_FAILURE', which will return them
|
573
|
+
* as integers.
|
574
|
+
*/
|
575
|
+
template <class real_t, class sparse_ix>
|
576
|
+
int fit_iforest(IsoForest *model_outputs, ExtIsoForest *model_outputs_ext,
|
577
|
+
real_t numeric_data[], size_t ncols_numeric,
|
578
|
+
int categ_data[], size_t ncols_categ, int ncat[],
|
579
|
+
real_t Xc[], sparse_ix Xc_ind[], sparse_ix Xc_indptr[],
|
580
|
+
size_t ndim, size_t ntry, CoefType coef_type, bool coef_by_prop,
|
581
|
+
real_t sample_weights[], bool with_replacement, bool weight_as_sample,
|
582
|
+
size_t nrows, size_t sample_size, size_t ntrees,
|
583
|
+
size_t max_depth, size_t ncols_per_tree,
|
584
|
+
bool limit_depth, bool penalize_range, bool standardize_data,
|
585
|
+
ScoringMetric scoring_metric, bool fast_bratio,
|
586
|
+
bool standardize_dist, double tmat[],
|
587
|
+
double output_depths[], bool standardize_depth,
|
588
|
+
real_t col_weights[], bool weigh_by_kurt,
|
589
|
+
double prob_pick_by_gain_pl, double prob_pick_by_gain_avg,
|
590
|
+
double prob_pick_by_full_gain, double prob_pick_by_dens,
|
591
|
+
double prob_pick_col_by_range, double prob_pick_col_by_var,
|
592
|
+
double prob_pick_col_by_kurt,
|
593
|
+
double min_gain, MissingAction missing_action,
|
594
|
+
CategSplit cat_split_type, NewCategAction new_cat_action,
|
595
|
+
bool all_perm, Imputer *imputer, size_t min_imp_obs,
|
596
|
+
UseDepthImp depth_imp, WeighImpRows weigh_imp_rows, bool impute_at_fit,
|
597
|
+
uint64_t random_seed, bool use_long_double, int nthreads)
|
598
|
+
{
|
599
|
+
if (use_long_double && !has_long_double()) {
|
600
|
+
use_long_double = false;
|
601
|
+
fprintf(stderr, "Passed 'use_long_double=true', but library was compiled without long double support.\n");
|
602
|
+
}
|
603
|
+
#ifndef NO_LONG_DOUBLE
|
604
|
+
if (likely(!use_long_double))
|
605
|
+
#endif
|
606
|
+
return fit_iforest_internal<real_t, sparse_ix, double>(
|
607
|
+
model_outputs, model_outputs_ext,
|
608
|
+
numeric_data, ncols_numeric,
|
609
|
+
categ_data, ncols_categ, ncat,
|
610
|
+
Xc, Xc_ind, Xc_indptr,
|
611
|
+
ndim, ntry, coef_type, coef_by_prop,
|
612
|
+
sample_weights, with_replacement, weight_as_sample,
|
613
|
+
nrows, sample_size, ntrees,
|
614
|
+
max_depth, ncols_per_tree,
|
615
|
+
limit_depth, penalize_range, standardize_data,
|
616
|
+
scoring_metric, fast_bratio,
|
617
|
+
standardize_dist, tmat,
|
618
|
+
output_depths, standardize_depth,
|
619
|
+
col_weights, weigh_by_kurt,
|
620
|
+
prob_pick_by_gain_pl, prob_pick_by_gain_avg,
|
621
|
+
prob_pick_by_full_gain, prob_pick_by_dens,
|
622
|
+
prob_pick_col_by_range, prob_pick_col_by_var,
|
623
|
+
prob_pick_col_by_kurt,
|
624
|
+
min_gain, missing_action,
|
625
|
+
cat_split_type, new_cat_action,
|
626
|
+
all_perm, imputer, min_imp_obs,
|
627
|
+
depth_imp, weigh_imp_rows, impute_at_fit,
|
628
|
+
random_seed, nthreads
|
629
|
+
);
|
630
|
+
#ifndef NO_LONG_DOUBLE
|
631
|
+
else
|
632
|
+
return fit_iforest_internal<real_t, sparse_ix, long double>(
|
633
|
+
model_outputs, model_outputs_ext,
|
634
|
+
numeric_data, ncols_numeric,
|
635
|
+
categ_data, ncols_categ, ncat,
|
636
|
+
Xc, Xc_ind, Xc_indptr,
|
637
|
+
ndim, ntry, coef_type, coef_by_prop,
|
638
|
+
sample_weights, with_replacement, weight_as_sample,
|
639
|
+
nrows, sample_size, ntrees,
|
640
|
+
max_depth, ncols_per_tree,
|
641
|
+
limit_depth, penalize_range, standardize_data,
|
642
|
+
scoring_metric, fast_bratio,
|
643
|
+
standardize_dist, tmat,
|
644
|
+
output_depths, standardize_depth,
|
645
|
+
col_weights, weigh_by_kurt,
|
646
|
+
prob_pick_by_gain_pl, prob_pick_by_gain_avg,
|
647
|
+
prob_pick_by_full_gain, prob_pick_by_dens,
|
648
|
+
prob_pick_col_by_range, prob_pick_col_by_var,
|
649
|
+
prob_pick_col_by_kurt,
|
650
|
+
min_gain, missing_action,
|
651
|
+
cat_split_type, new_cat_action,
|
652
|
+
all_perm, imputer, min_imp_obs,
|
653
|
+
depth_imp, weigh_imp_rows, impute_at_fit,
|
654
|
+
random_seed, nthreads
|
655
|
+
);
|
656
|
+
#endif
|
657
|
+
}
|
658
|
+
|
659
|
+
template <class real_t, class sparse_ix, class ldouble_safe>
|
660
|
+
int fit_iforest_internal(
|
661
|
+
IsoForest *model_outputs, ExtIsoForest *model_outputs_ext,
|
662
|
+
real_t numeric_data[], size_t ncols_numeric,
|
663
|
+
int categ_data[], size_t ncols_categ, int ncat[],
|
664
|
+
real_t Xc[], sparse_ix Xc_ind[], sparse_ix Xc_indptr[],
|
665
|
+
size_t ndim, size_t ntry, CoefType coef_type, bool coef_by_prop,
|
666
|
+
real_t sample_weights[], bool with_replacement, bool weight_as_sample,
|
667
|
+
size_t nrows, size_t sample_size, size_t ntrees,
|
668
|
+
size_t max_depth, size_t ncols_per_tree,
|
669
|
+
bool limit_depth, bool penalize_range, bool standardize_data,
|
670
|
+
ScoringMetric scoring_metric, bool fast_bratio,
|
671
|
+
bool standardize_dist, double tmat[],
|
672
|
+
double output_depths[], bool standardize_depth,
|
673
|
+
real_t col_weights[], bool weigh_by_kurt,
|
674
|
+
double prob_pick_by_gain_pl, double prob_pick_by_gain_avg,
|
675
|
+
double prob_pick_by_full_gain, double prob_pick_by_dens,
|
676
|
+
double prob_pick_col_by_range, double prob_pick_col_by_var,
|
677
|
+
double prob_pick_col_by_kurt,
|
678
|
+
double min_gain, MissingAction missing_action,
|
679
|
+
CategSplit cat_split_type, NewCategAction new_cat_action,
|
680
|
+
bool all_perm, Imputer *imputer, size_t min_imp_obs,
|
681
|
+
UseDepthImp depth_imp, WeighImpRows weigh_imp_rows, bool impute_at_fit,
|
682
|
+
uint64_t random_seed, int nthreads)
|
683
|
+
{
|
684
|
+
if (
|
685
|
+
prob_pick_by_gain_avg < 0 || prob_pick_by_gain_pl < 0 ||
|
686
|
+
prob_pick_by_full_gain < 0 || prob_pick_by_dens < 0 ||
|
687
|
+
prob_pick_col_by_range < 0 ||
|
688
|
+
prob_pick_col_by_var < 0 || prob_pick_col_by_kurt < 0
|
689
|
+
) {
|
690
|
+
throw std::runtime_error("Cannot pass negative probabilities.\n");
|
691
|
+
}
|
692
|
+
if (prob_pick_col_by_range && ncols_categ)
|
693
|
+
throw std::runtime_error("'prob_pick_col_by_range' is not compatible with categorical data.\n");
|
694
|
+
if (prob_pick_by_full_gain && ncols_categ)
|
695
|
+
throw std::runtime_error("'prob_pick_by_full_gain' is not compatible with categorical data.\n");
|
696
|
+
if (prob_pick_col_by_kurt && weigh_by_kurt)
|
697
|
+
throw std::runtime_error("'weigh_by_kurt' and 'prob_pick_col_by_kurt' cannot be used together.\n");
|
698
|
+
if (ndim == 0 && model_outputs == NULL)
|
699
|
+
throw std::runtime_error("Must pass 'ndim>0' in the extended model.\n");
|
700
|
+
if (penalize_range &&
|
701
|
+
(scoring_metric == Density ||
|
702
|
+
scoring_metric == AdjDensity ||
|
703
|
+
is_boxed_metric(scoring_metric))
|
704
|
+
)
|
705
|
+
throw std::runtime_error("'penalize_range' is incompatible with density scoring.\n");
|
706
|
+
if (with_replacement) {
|
707
|
+
if (tmat != NULL)
|
708
|
+
throw std::runtime_error("Cannot calculate distance while sampling with replacement.\n");
|
709
|
+
if (output_depths != NULL)
|
710
|
+
throw std::runtime_error("Cannot make predictions at fit time when sampling with replacement.\n");
|
711
|
+
if (impute_at_fit)
|
712
|
+
throw std::runtime_error("Cannot impute at fit time when sampling with replacement.\n");
|
713
|
+
}
|
714
|
+
if (sample_size != 0 && sample_size < nrows) {
|
715
|
+
if (output_depths != NULL)
|
716
|
+
throw std::runtime_error("Cannot produce outlier scores at fit time when using sub-sampling.\n");
|
717
|
+
if (tmat != NULL)
|
718
|
+
throw std::runtime_error("Cannot calculate distances at fit time when using sub-sampling.\n");
|
719
|
+
if (impute_at_fit)
|
720
|
+
throw std::runtime_error("Cannot produce missing data imputations at fit time when using sub-sampling.\n");
|
721
|
+
}
|
722
|
+
|
723
|
+
|
724
|
+
/* TODO: this function should also accept the array as a memoryview with a
|
725
|
+
leading dimension that might not correspond to the number of columns,
|
726
|
+
so as to avoid having to make deep copies of memoryviews in python and to
|
727
|
+
allow using pointers to columns of dataframes in R and Python. */
|
728
|
+
|
729
|
+
/* calculate maximum number of categories to use later */
|
730
|
+
int max_categ = 0;
|
731
|
+
for (size_t col = 0; col < ncols_categ; col++)
|
732
|
+
max_categ = (ncat[col] > max_categ)? ncat[col] : max_categ;
|
733
|
+
|
734
|
+
bool calc_dist = tmat != NULL;
|
735
|
+
|
736
|
+
if (sample_size == 0)
|
737
|
+
sample_size = nrows;
|
738
|
+
|
739
|
+
if (model_outputs != NULL)
|
740
|
+
ntry = std::min(ntry, ncols_numeric + ncols_categ);
|
741
|
+
|
742
|
+
if (ncols_per_tree == 0)
|
743
|
+
ncols_per_tree = ncols_numeric + ncols_categ;
|
744
|
+
|
745
|
+
/* put data in structs to shorten function calls */
|
746
|
+
InputData<real_t, sparse_ix>
|
747
|
+
input_data = {numeric_data, ncols_numeric, categ_data, ncat, max_categ, ncols_categ,
|
748
|
+
nrows, ncols_numeric + ncols_categ, sample_weights,
|
749
|
+
weight_as_sample, col_weights,
|
750
|
+
Xc, Xc_ind, Xc_indptr,
|
751
|
+
0, 0, std::vector<double>(),
|
752
|
+
std::vector<char>(), 0, NULL,
|
753
|
+
(double*)NULL, (double*)NULL, (int*)NULL, std::vector<double>(),
|
754
|
+
std::vector<double>(), std::vector<double>(),
|
755
|
+
std::vector<size_t>(), std::vector<size_t>()};
|
756
|
+
ModelParams model_params = {with_replacement, sample_size, ntrees, ncols_per_tree,
|
757
|
+
limit_depth? log2ceil(sample_size) : max_depth? max_depth : (sample_size - 1),
|
758
|
+
penalize_range, standardize_data, random_seed, weigh_by_kurt,
|
759
|
+
prob_pick_by_gain_avg, prob_pick_by_gain_pl,
|
760
|
+
prob_pick_by_full_gain, prob_pick_by_dens,
|
761
|
+
prob_pick_col_by_range, prob_pick_col_by_var,
|
762
|
+
prob_pick_col_by_kurt,
|
763
|
+
min_gain, cat_split_type, new_cat_action, missing_action,
|
764
|
+
scoring_metric, fast_bratio, all_perm,
|
765
|
+
(model_outputs != NULL)? 0 : ndim, ntry,
|
766
|
+
coef_type, coef_by_prop, calc_dist, (bool)(output_depths != NULL), impute_at_fit,
|
767
|
+
depth_imp, weigh_imp_rows, min_imp_obs};
|
768
|
+
|
769
|
+
/* if calculating full gain, need to produce copies of the data in row-major order */
|
770
|
+
if (prob_pick_by_full_gain)
|
771
|
+
{
|
772
|
+
if (input_data.Xc_indptr == NULL)
|
773
|
+
colmajor_to_rowmajor(input_data.numeric_data, input_data.nrows, input_data.ncols_numeric, input_data.X_row_major);
|
774
|
+
else
|
775
|
+
colmajor_to_rowmajor(input_data.Xc, input_data.Xc_ind, input_data.Xc_indptr,
|
776
|
+
input_data.nrows, input_data.ncols_numeric,
|
777
|
+
input_data.Xr, input_data.Xr_ind, input_data.Xr_indptr);
|
778
|
+
}
|
779
|
+
|
780
|
+
/* if using weights as sampling probability, build a binary tree for faster sampling */
|
781
|
+
if (input_data.weight_as_sample && input_data.sample_weights != NULL)
|
782
|
+
{
|
783
|
+
build_btree_sampler(input_data.btree_weights_init, input_data.sample_weights,
|
784
|
+
input_data.nrows, input_data.log2_n, input_data.btree_offset);
|
785
|
+
}
|
786
|
+
|
787
|
+
/* same for column weights */
|
788
|
+
/* TODO: this should also save the kurtoses when using 'prob_pick_col_by_kurt' */
|
789
|
+
ColumnSampler<ldouble_safe> base_col_sampler;
|
790
|
+
if (
|
791
|
+
col_weights != NULL ||
|
792
|
+
(model_params.weigh_by_kurt && model_params.sample_size == input_data.nrows && !model_params.with_replacement &&
|
793
|
+
(model_params.ncols_per_tree >= input_data.ncols_tot / (model_params.ntrees * 2)))
|
794
|
+
)
|
795
|
+
{
|
796
|
+
bool avoid_col_weights = (model_outputs != NULL && model_params.ntry >= model_params.ncols_per_tree &&
|
797
|
+
model_params.prob_pick_by_gain_avg + model_params.prob_pick_by_gain_pl +
|
798
|
+
model_params.prob_pick_by_full_gain + model_params.prob_pick_by_dens >= 1)
|
799
|
+
||
|
800
|
+
(model_outputs == NULL && model_params.ndim >= model_params.ncols_per_tree)
|
801
|
+
||
|
802
|
+
(model_params.ncols_per_tree == 1);
|
803
|
+
if (!avoid_col_weights)
|
804
|
+
{
|
805
|
+
if (model_params.weigh_by_kurt && model_params.sample_size == input_data.nrows && !model_params.with_replacement)
|
806
|
+
{
|
807
|
+
RNG_engine rnd_generator(random_seed);
|
808
|
+
std::vector<double> kurt_weights = calc_kurtosis_all_data<InputData<real_t, sparse_ix>, ldouble_safe>(input_data, model_params, rnd_generator);
|
809
|
+
if (col_weights != NULL)
|
810
|
+
{
|
811
|
+
for (size_t col = 0; col < input_data.ncols_tot; col++)
|
812
|
+
{
|
813
|
+
if (kurt_weights[col] <= 0) continue;
|
814
|
+
kurt_weights[col] *= col_weights[col];
|
815
|
+
kurt_weights[col] = std::fmax(kurt_weights[col], 1e-100);
|
816
|
+
}
|
817
|
+
}
|
818
|
+
base_col_sampler.initialize(kurt_weights.data(), input_data.ncols_tot);
|
819
|
+
|
820
|
+
if (model_params.prob_pick_col_by_range || model_params.prob_pick_col_by_var)
|
821
|
+
{
|
822
|
+
input_data.all_kurtoses = std::move(kurt_weights);
|
823
|
+
}
|
824
|
+
}
|
825
|
+
|
826
|
+
else
|
827
|
+
{
|
828
|
+
base_col_sampler.initialize(input_data.col_weights, input_data.ncols_tot);
|
829
|
+
}
|
830
|
+
|
831
|
+
input_data.preinitialized_col_sampler = &base_col_sampler;
|
832
|
+
}
|
833
|
+
}
|
834
|
+
|
835
|
+
/* in some cases, all trees will need to calculate variable ranges for all columns */
|
836
|
+
/* TODO: the model might use 'leave_m_cols', or have 'prob_pick_col_by_range<1', in which
|
837
|
+
case it might not be beneficial to do this beforehand. Find out when the expected gain
|
838
|
+
from doing this here is not beneficial. */
|
839
|
+
/* TODO: move this to a different file, it doesn't belong here */
|
840
|
+
std::vector<double> variable_ranges_low;
|
841
|
+
std::vector<double> variable_ranges_high;
|
842
|
+
std::vector<int> variable_ncats;
|
843
|
+
if (
|
844
|
+
model_params.sample_size == input_data.nrows && !model_params.with_replacement &&
|
845
|
+
(model_params.ncols_per_tree >= input_data.ncols_numeric) &&
|
846
|
+
((model_params.prob_pick_col_by_range && input_data.ncols_numeric)
|
847
|
+
||
|
848
|
+
is_boxed_metric(model_params.scoring_metric))
|
849
|
+
)
|
850
|
+
{
|
851
|
+
variable_ranges_low.resize(input_data.ncols_numeric);
|
852
|
+
variable_ranges_high.resize(input_data.ncols_numeric);
|
853
|
+
|
854
|
+
std::unique_ptr<unsigned char[]> buffer_cats;
|
855
|
+
size_t adj_col;
|
856
|
+
if (is_boxed_metric(model_params.scoring_metric))
|
857
|
+
{
|
858
|
+
variable_ncats.resize(input_data.ncols_categ);
|
859
|
+
buffer_cats = std::unique_ptr<unsigned char[]>(new unsigned char[input_data.max_categ]);
|
860
|
+
}
|
861
|
+
|
862
|
+
if (base_col_sampler.col_indices.empty())
|
863
|
+
base_col_sampler.initialize(input_data.ncols_tot);
|
864
|
+
|
865
|
+
bool unsplittable;
|
866
|
+
size_t n_tried_numeric = 0;
|
867
|
+
size_t col;
|
868
|
+
base_col_sampler.prepare_full_pass();
|
869
|
+
while (base_col_sampler.sample_col(col))
|
870
|
+
{
|
871
|
+
if (col < input_data.ncols_numeric)
|
872
|
+
{
|
873
|
+
if (input_data.Xc_indptr == NULL)
|
874
|
+
{
|
875
|
+
get_range(input_data.numeric_data + nrows*col,
|
876
|
+
input_data.nrows,
|
877
|
+
model_params.missing_action,
|
878
|
+
variable_ranges_low[col],
|
879
|
+
variable_ranges_high[col],
|
880
|
+
unsplittable);
|
881
|
+
}
|
882
|
+
|
883
|
+
else
|
884
|
+
{
|
885
|
+
get_range(col, input_data.nrows,
|
886
|
+
input_data.Xc, input_data.Xc_ind, input_data.Xc_indptr,
|
887
|
+
model_params.missing_action,
|
888
|
+
variable_ranges_low[col],
|
889
|
+
variable_ranges_high[col],
|
890
|
+
unsplittable);
|
891
|
+
}
|
892
|
+
|
893
|
+
n_tried_numeric++;
|
894
|
+
|
895
|
+
if (unsplittable)
|
896
|
+
{
|
897
|
+
variable_ranges_low[col] = 0;
|
898
|
+
variable_ranges_high[col] = 0;
|
899
|
+
base_col_sampler.drop_col(col);
|
900
|
+
}
|
901
|
+
}
|
902
|
+
|
903
|
+
else
|
904
|
+
{
|
905
|
+
if (!is_boxed_metric(model_params.scoring_metric))
|
906
|
+
{
|
907
|
+
if (n_tried_numeric >= input_data.ncols_numeric)
|
908
|
+
break;
|
909
|
+
else
|
910
|
+
continue;
|
911
|
+
}
|
912
|
+
adj_col = col - input_data.ncols_numeric;
|
913
|
+
|
914
|
+
|
915
|
+
variable_ncats[adj_col] = count_ncateg_in_col(input_data.categ_data + input_data.nrows*adj_col,
|
916
|
+
input_data.nrows, input_data.ncat[adj_col],
|
917
|
+
buffer_cats.get());
|
918
|
+
if (variable_ncats[adj_col] <= 1)
|
919
|
+
base_col_sampler.drop_col(col);
|
920
|
+
}
|
921
|
+
}
|
922
|
+
|
923
|
+
input_data.preinitialized_col_sampler = &base_col_sampler;
|
924
|
+
if (input_data.ncols_numeric) {
|
925
|
+
input_data.range_low = variable_ranges_low.data();
|
926
|
+
input_data.range_high = variable_ranges_high.data();
|
927
|
+
}
|
928
|
+
if (input_data.ncols_categ) {
|
929
|
+
input_data.ncat_ = variable_ncats.data();
|
930
|
+
}
|
931
|
+
}
|
932
|
+
|
933
|
+
/* if imputing missing values on-the-fly, need to determine which are missing */
|
934
|
+
std::vector<ImputedData<sparse_ix, ldouble_safe>> impute_vec;
|
935
|
+
hashed_map<size_t, ImputedData<sparse_ix, ldouble_safe>> impute_map;
|
936
|
+
if (model_params.impute_at_fit)
|
937
|
+
check_for_missing(input_data, impute_vec, impute_map, nthreads);
|
938
|
+
|
939
|
+
/* store model data */
|
940
|
+
if (model_outputs != NULL)
|
941
|
+
{
|
942
|
+
model_outputs->trees.resize(ntrees);
|
943
|
+
model_outputs->trees.shrink_to_fit();
|
944
|
+
model_outputs->new_cat_action = new_cat_action;
|
945
|
+
model_outputs->cat_split_type = cat_split_type;
|
946
|
+
model_outputs->missing_action = missing_action;
|
947
|
+
model_outputs->scoring_metric = scoring_metric;
|
948
|
+
if (
|
949
|
+
model_outputs->scoring_metric != Density &&
|
950
|
+
model_outputs->scoring_metric != BoxedDensity &&
|
951
|
+
model_outputs->scoring_metric != BoxedDensity2 &&
|
952
|
+
model_outputs->scoring_metric != BoxedRatio
|
953
|
+
)
|
954
|
+
model_outputs->exp_avg_depth = expected_avg_depth<ldouble_safe>(sample_size);
|
955
|
+
else
|
956
|
+
model_outputs->exp_avg_depth = 1;
|
957
|
+
model_outputs->exp_avg_sep = expected_separation_depth<ldouble_safe>(model_params.sample_size);
|
958
|
+
model_outputs->orig_sample_size = input_data.nrows;
|
959
|
+
model_outputs->has_range_penalty = penalize_range;
|
960
|
+
}
|
961
|
+
|
962
|
+
else
|
963
|
+
{
|
964
|
+
model_outputs_ext->hplanes.resize(ntrees);
|
965
|
+
model_outputs_ext->hplanes.shrink_to_fit();
|
966
|
+
model_outputs_ext->new_cat_action = new_cat_action;
|
967
|
+
model_outputs_ext->cat_split_type = cat_split_type;
|
968
|
+
model_outputs_ext->missing_action = missing_action;
|
969
|
+
model_outputs_ext->scoring_metric = scoring_metric;
|
970
|
+
if (
|
971
|
+
model_outputs_ext->scoring_metric != Density &&
|
972
|
+
model_outputs_ext->scoring_metric != BoxedDensity &&
|
973
|
+
model_outputs_ext->scoring_metric != BoxedDensity2 &&
|
974
|
+
model_outputs_ext->scoring_metric != BoxedRatio
|
975
|
+
)
|
976
|
+
model_outputs_ext->exp_avg_depth = expected_avg_depth<ldouble_safe>(sample_size);
|
977
|
+
else
|
978
|
+
model_outputs_ext->exp_avg_depth = 1;
|
979
|
+
model_outputs_ext->exp_avg_sep = expected_separation_depth<ldouble_safe>(model_params.sample_size);
|
980
|
+
model_outputs_ext->orig_sample_size = input_data.nrows;
|
981
|
+
model_outputs_ext->has_range_penalty = penalize_range;
|
982
|
+
}
|
983
|
+
|
984
|
+
if (imputer != NULL)
|
985
|
+
initialize_imputer<decltype(input_data), ldouble_safe>(
|
986
|
+
*imputer, input_data, ntrees, nthreads
|
987
|
+
);
|
988
|
+
|
989
|
+
/* initialize thread-private memory */
|
990
|
+
if ((size_t)nthreads > ntrees)
|
991
|
+
nthreads = (int)ntrees;
|
992
|
+
#ifdef _OPENMP
|
993
|
+
std::vector<WorkerMemory<ImputedData<sparse_ix, ldouble_safe>, ldouble_safe, real_t>> worker_memory(nthreads);
|
994
|
+
#else
|
995
|
+
std::vector<WorkerMemory<ImputedData<sparse_ix, ldouble_safe>, ldouble_safe, real_t>> worker_memory(1);
|
996
|
+
#endif
|
997
|
+
|
998
|
+
/* Global variable that determines if the procedure receives a stop signal */
|
999
|
+
SignalSwitcher ss = SignalSwitcher();
|
1000
|
+
|
1001
|
+
/* For exception handling */
|
1002
|
+
bool threw_exception = false;
|
1003
|
+
std::exception_ptr ex = NULL;
|
1004
|
+
|
1005
|
+
/* grow trees */
|
1006
|
+
#pragma omp parallel for num_threads(nthreads) schedule(dynamic) shared(model_outputs, model_outputs_ext, worker_memory, input_data, model_params, threw_exception, ex)
|
1007
|
+
for (size_t_for tree = 0; tree < (decltype(tree))ntrees; tree++)
|
1008
|
+
{
|
1009
|
+
if (interrupt_switch || threw_exception)
|
1010
|
+
continue; /* Cannot break with OpenMP==2.0 (MSVC) */
|
1011
|
+
|
1012
|
+
try
|
1013
|
+
{
|
1014
|
+
if (
|
1015
|
+
model_params.impute_at_fit &&
|
1016
|
+
input_data.n_missing &&
|
1017
|
+
!worker_memory[omp_get_thread_num()].impute_vec.size() &&
|
1018
|
+
!worker_memory[omp_get_thread_num()].impute_map.size()
|
1019
|
+
)
|
1020
|
+
{
|
1021
|
+
#ifdef _OPENMP
|
1022
|
+
if (nthreads > 1)
|
1023
|
+
{
|
1024
|
+
worker_memory[omp_get_thread_num()].impute_vec = impute_vec;
|
1025
|
+
worker_memory[omp_get_thread_num()].impute_map = impute_map;
|
1026
|
+
}
|
1027
|
+
|
1028
|
+
else
|
1029
|
+
#endif
|
1030
|
+
{
|
1031
|
+
worker_memory[0].impute_vec = std::move(impute_vec);
|
1032
|
+
worker_memory[0].impute_map = std::move(impute_map);
|
1033
|
+
}
|
1034
|
+
}
|
1035
|
+
|
1036
|
+
fit_itree<decltype(input_data), typename std::remove_pointer<decltype(worker_memory.data())>::type, ldouble_safe>(
|
1037
|
+
(model_outputs != NULL)? &model_outputs->trees[tree] : NULL,
|
1038
|
+
(model_outputs_ext != NULL)? &model_outputs_ext->hplanes[tree] : NULL,
|
1039
|
+
worker_memory[omp_get_thread_num()],
|
1040
|
+
input_data,
|
1041
|
+
model_params,
|
1042
|
+
(imputer != NULL)? &(imputer->imputer_tree[tree]) : NULL,
|
1043
|
+
tree);
|
1044
|
+
|
1045
|
+
if ((model_outputs != NULL))
|
1046
|
+
model_outputs->trees[tree].shrink_to_fit();
|
1047
|
+
else
|
1048
|
+
model_outputs_ext->hplanes[tree].shrink_to_fit();
|
1049
|
+
}
|
1050
|
+
|
1051
|
+
catch (...)
|
1052
|
+
{
|
1053
|
+
#pragma omp critical
|
1054
|
+
{
|
1055
|
+
if (!threw_exception)
|
1056
|
+
{
|
1057
|
+
threw_exception = true;
|
1058
|
+
ex = std::current_exception();
|
1059
|
+
}
|
1060
|
+
}
|
1061
|
+
}
|
1062
|
+
}
|
1063
|
+
|
1064
|
+
/* check if the procedure got interrupted */
|
1065
|
+
check_interrupt_switch(ss);
|
1066
|
+
#if defined(DONT_THROW_ON_INTERRUPT)
|
1067
|
+
if (interrupt_switch) return EXIT_FAILURE;
|
1068
|
+
#endif
|
1069
|
+
|
1070
|
+
/* check if some exception was thrown */
|
1071
|
+
if (threw_exception)
|
1072
|
+
std::rethrow_exception(ex);
|
1073
|
+
|
1074
|
+
if ((model_outputs != NULL))
|
1075
|
+
model_outputs->trees.shrink_to_fit();
|
1076
|
+
else
|
1077
|
+
model_outputs_ext->hplanes.shrink_to_fit();
|
1078
|
+
|
1079
|
+
/* if calculating similarity/distance, now need to reduce and average */
|
1080
|
+
if (calc_dist)
|
1081
|
+
gather_sim_result< PredictionData<real_t, sparse_ix>, InputData<real_t, sparse_ix> >
|
1082
|
+
(NULL, &worker_memory,
|
1083
|
+
NULL, &input_data,
|
1084
|
+
model_outputs, model_outputs_ext,
|
1085
|
+
tmat, NULL, 0,
|
1086
|
+
model_params.ntrees, false,
|
1087
|
+
standardize_dist, false, nthreads);
|
1088
|
+
|
1089
|
+
check_interrupt_switch(ss);
|
1090
|
+
#if defined(DONT_THROW_ON_INTERRUPT)
|
1091
|
+
if (interrupt_switch) return EXIT_FAILURE;
|
1092
|
+
#endif
|
1093
|
+
|
1094
|
+
/* same for depths */
|
1095
|
+
if (output_depths != NULL)
|
1096
|
+
{
|
1097
|
+
#ifdef _OPENMP
|
1098
|
+
if (nthreads > 1)
|
1099
|
+
{
|
1100
|
+
for (auto &w : worker_memory)
|
1101
|
+
{
|
1102
|
+
if (w.row_depths.size())
|
1103
|
+
{
|
1104
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) shared(input_data, output_depths, w, worker_memory)
|
1105
|
+
for (size_t_for row = 0; row < (decltype(row))input_data.nrows; row++)
|
1106
|
+
output_depths[row] += w.row_depths[row];
|
1107
|
+
}
|
1108
|
+
}
|
1109
|
+
}
|
1110
|
+
else
|
1111
|
+
#endif
|
1112
|
+
{
|
1113
|
+
std::copy(worker_memory[0].row_depths.begin(), worker_memory[0].row_depths.end(), output_depths);
|
1114
|
+
}
|
1115
|
+
|
1116
|
+
if (standardize_depth)
|
1117
|
+
{
|
1118
|
+
double depth_divisor = (double)ntrees * ((model_outputs != NULL)?
|
1119
|
+
model_outputs->exp_avg_depth : model_outputs_ext->exp_avg_depth);
|
1120
|
+
for (size_t row = 0; row < nrows; row++)
|
1121
|
+
output_depths[row] = std::exp2( - output_depths[row] / depth_divisor );
|
1122
|
+
}
|
1123
|
+
|
1124
|
+
else
|
1125
|
+
{
|
1126
|
+
double ntrees_dbl = (double) ntrees;
|
1127
|
+
for (size_t row = 0; row < nrows; row++)
|
1128
|
+
output_depths[row] /= ntrees_dbl;
|
1129
|
+
}
|
1130
|
+
}
|
1131
|
+
|
1132
|
+
check_interrupt_switch(ss);
|
1133
|
+
#if defined(DONT_THROW_ON_INTERRUPT)
|
1134
|
+
if (interrupt_switch) return EXIT_FAILURE;
|
1135
|
+
#endif
|
1136
|
+
|
1137
|
+
/* if imputing missing values, now need to reduce and write final values */
|
1138
|
+
if (model_params.impute_at_fit)
|
1139
|
+
{
|
1140
|
+
#ifdef _OPENMP
|
1141
|
+
if (nthreads > 1)
|
1142
|
+
{
|
1143
|
+
for (auto &w : worker_memory)
|
1144
|
+
combine_tree_imputations(w, impute_vec, impute_map, input_data.has_missing, nthreads);
|
1145
|
+
}
|
1146
|
+
|
1147
|
+
else
|
1148
|
+
#endif
|
1149
|
+
{
|
1150
|
+
impute_vec = std::move(worker_memory[0].impute_vec);
|
1151
|
+
impute_map = std::move(worker_memory[0].impute_map);
|
1152
|
+
}
|
1153
|
+
|
1154
|
+
apply_imputation_results(impute_vec, impute_map, *imputer, input_data, nthreads);
|
1155
|
+
}
|
1156
|
+
|
1157
|
+
check_interrupt_switch(ss);
|
1158
|
+
#if defined(DONT_THROW_ON_INTERRUPT)
|
1159
|
+
if (interrupt_switch) return EXIT_FAILURE;
|
1160
|
+
#endif
|
1161
|
+
|
1162
|
+
return EXIT_SUCCESS;
|
1163
|
+
}
|
1164
|
+
|
1165
|
+
|
1166
|
+
/* Add additional trees to already-fitted isolation forest model
|
1167
|
+
*
|
1168
|
+
* Parameters
|
1169
|
+
* ==========
|
1170
|
+
* - model_outputs
|
1171
|
+
* Pointer to fitted single-variable model object from function 'fit_iforest'. Pass NULL
|
1172
|
+
* if the trees are are to be added to an extended model. Can only pass one of
|
1173
|
+
* 'model_outputs' and 'model_outputs_ext'. Note that this function is not thread-safe,
|
1174
|
+
* so it cannot be run in parallel for the same model object.
|
1175
|
+
* - model_outputs_ext
|
1176
|
+
* Pointer to fitted extended model object from function 'fit_iforest'. Pass NULL
|
1177
|
+
* if the trees are are to be added to an single-variable model. Can only pass one of
|
1178
|
+
* 'model_outputs' and 'model_outputs_ext'. Note that this function is not thread-safe,
|
1179
|
+
* so it cannot be run in parallel for the same model object.
|
1180
|
+
* - numeric_data[nrows * ncols_numeric]
|
1181
|
+
* Pointer to numeric data to which to fit this additional tree. Must be ordered by columns like Fortran,
|
1182
|
+
* not ordered by rows like C (i.e. entries 1..n contain column 0, n+1..2n column 1, etc.).
|
1183
|
+
* Pass NULL if there are no dense numeric columns.
|
1184
|
+
* Can only pass one of 'numeric_data' or 'Xc' + 'Xc_ind' + 'Xc_indptr'.
|
1185
|
+
* If the model from 'fit_iforest' was fit to numeric data, must pass numeric data with the same number
|
1186
|
+
* of columns, either as dense or as sparse arrays.
|
1187
|
+
* - ncols_numeric
|
1188
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Cannot be changed from
|
1189
|
+
* what was originally passed to 'fit_iforest'.
|
1190
|
+
* - categ_data[nrows * ncols_categ]
|
1191
|
+
* Pointer to categorical data to which to fit this additional tree. Must be ordered by columns like Fortran,
|
1192
|
+
* not ordered by rows like C (i.e. entries 1..n contain column 0, n+1..2n column 1, etc.).
|
1193
|
+
* Pass NULL if there are no categorical columns. The encoding must be the same as was used
|
1194
|
+
* in the data to which the model was fit.
|
1195
|
+
* Each category should be represented as an integer, and these integers must start at zero and
|
1196
|
+
* be in consecutive order - i.e. if category '3' is present, category '2' must have also been
|
1197
|
+
* present when the model was fit (note that they are not treated as being ordinal, this is just
|
1198
|
+
* an encoding). Missing values should be encoded as negative numbers such as (-1). The encoding
|
1199
|
+
* must be the same as was used in the data to which the model was fit.
|
1200
|
+
* If the model from 'fit_iforest' was fit to categorical data, must pass categorical data with the same number
|
1201
|
+
* of columns and the same category encoding.
|
1202
|
+
* - ncols_categ
|
1203
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Cannot be changed from
|
1204
|
+
* what was originally passed to 'fit_iforest'.
|
1205
|
+
* - ncat[ncols_categ]
|
1206
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). May contain new categories,
|
1207
|
+
* but should keep the same encodings that were used for previous categories.
|
1208
|
+
* - Xc[nnz]
|
1209
|
+
* Pointer to numeric data in sparse numeric matrix in CSC format (column-compressed).
|
1210
|
+
* Pass NULL if there are no sparse numeric columns.
|
1211
|
+
* Can only pass one of 'numeric_data' or 'Xc' + 'Xc_ind' + 'Xc_indptr'.
|
1212
|
+
* - Xc_ind[nnz]
|
1213
|
+
* Pointer to row indices to which each non-zero entry in 'Xc' corresponds.
|
1214
|
+
* Must be in sorted order, otherwise results will be incorrect.
|
1215
|
+
* Pass NULL if there are no sparse numeric columns.
|
1216
|
+
* - Xc_indptr[ncols_numeric + 1]
|
1217
|
+
* Pointer to column index pointers that tell at entry [col] where does column 'col'
|
1218
|
+
* start and at entry [col + 1] where does column 'col' end.
|
1219
|
+
* Pass NULL if there are no sparse numeric columns.
|
1220
|
+
* - ndim
|
1221
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Cannot be changed from
|
1222
|
+
* what was originally passed to 'fit_iforest'.
|
1223
|
+
* - ntry
|
1224
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1225
|
+
* what was originally passed to 'fit_iforest'.
|
1226
|
+
* - coef_type
|
1227
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1228
|
+
* what was originally passed to 'fit_iforest'.
|
1229
|
+
* - sample_weights
|
1230
|
+
* Weights for the rows when adding this tree, either as sampling importances when using
|
1231
|
+
* sub-samples for each tree (i.e. passing weight '2' makes a row twice as likely to be included
|
1232
|
+
* in a random sub-sample), or as density measurement (i.e. passing weight '2' is the same as if
|
1233
|
+
* the row appeared twice, thus it's less of an outlier) - how this is taken is determined
|
1234
|
+
* through parameter 'weight_as_sample' that was passed to 'fit_iforest.
|
1235
|
+
* Pass NULL if the rows all have uniform weights.
|
1236
|
+
* - nrows
|
1237
|
+
* Number of rows in 'numeric_data', 'Xc', 'categ_data'.
|
1238
|
+
* - max_depth
|
1239
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1240
|
+
* what was originally passed to 'fit_iforest'.
|
1241
|
+
* - ncols_per_tree
|
1242
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1243
|
+
* what was originally passed to 'fit_iforest'.
|
1244
|
+
* - limit_depth
|
1245
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1246
|
+
* what was originally passed to 'fit_iforest'.
|
1247
|
+
* - penalize_range
|
1248
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1249
|
+
* what was originally passed to 'fit_iforest'.
|
1250
|
+
* - standardize_data
|
1251
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1252
|
+
* what was originally passed to 'fit_iforest'.
|
1253
|
+
* - fast_bratio
|
1254
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1255
|
+
* what was originally passed to 'fit_iforest'.
|
1256
|
+
* - col_weights
|
1257
|
+
* Sampling weights for each column, assuming all the numeric columns come before the categorical columns.
|
1258
|
+
* Ignored when picking columns by deterministic criterion.
|
1259
|
+
* If passing NULL, each column will have a uniform weight. If used along with kurtosis weights, the
|
1260
|
+
* effect is multiplicative.
|
1261
|
+
* - weigh_by_kurt
|
1262
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1263
|
+
* what was originally passed to 'fit_iforest'.
|
1264
|
+
* - prob_pick_by_gain_pl
|
1265
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1266
|
+
* what was originally passed to 'fit_iforest'.
|
1267
|
+
* - prob_pick_by_gain_avg
|
1268
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1269
|
+
* what was originally passed to 'fit_iforest'.
|
1270
|
+
* - prob_pick_by_full_gain
|
1271
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1272
|
+
* what was originally passed to 'fit_iforest'.
|
1273
|
+
* - prob_pick_by_dens
|
1274
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1275
|
+
* what was originally passed to 'fit_iforest'.
|
1276
|
+
* - prob_pick_col_by_range
|
1277
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1278
|
+
* what was originally passed to 'fit_iforest'.
|
1279
|
+
* - prob_pick_col_by_var
|
1280
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1281
|
+
* what was originally passed to 'fit_iforest'.
|
1282
|
+
* - prob_pick_col_by_kurt
|
1283
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1284
|
+
* what was originally passed to 'fit_iforest'.
|
1285
|
+
* - min_gain
|
1286
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1287
|
+
* what was originally passed to 'fit_iforest'.
|
1288
|
+
* - missing_action
|
1289
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Cannot be changed from
|
1290
|
+
* what was originally passed to 'fit_iforest'.
|
1291
|
+
* - cat_split_type
|
1292
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Cannot be changed from
|
1293
|
+
* what was originally passed to 'fit_iforest'.
|
1294
|
+
* - new_cat_action
|
1295
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Cannot be changed from
|
1296
|
+
* what was originally passed to 'fit_iforest'.
|
1297
|
+
* - depth_imp
|
1298
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Cannot be changed from
|
1299
|
+
* what was originally passed to 'fit_iforest'.
|
1300
|
+
* - weigh_imp_rows
|
1301
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Cannot be changed from
|
1302
|
+
* what was originally passed to 'fit_iforest'.
|
1303
|
+
* - all_perm
|
1304
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1305
|
+
* what was originally passed to 'fit_iforest'.
|
1306
|
+
* - coef_by_prop
|
1307
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1308
|
+
* what was originally passed to 'fit_iforest'.
|
1309
|
+
* - imputer
|
1310
|
+
* Pointer to already-allocated imputer object, as it was output from function 'fit_model' while
|
1311
|
+
* producing either 'model_outputs' or 'model_outputs_ext'.
|
1312
|
+
* Pass NULL if the model was built without imputer.
|
1313
|
+
* - min_imp_obs
|
1314
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1315
|
+
* what was originally passed to 'fit_iforest'.
|
1316
|
+
* - indexer
|
1317
|
+
* Indexer object associated to the model object ('model_outputs' or 'model_outputs_ext'), which will
|
1318
|
+
* be updated with the new tree to add.
|
1319
|
+
* If 'indexer' has reference points, these must be passed again here in order to index them.
|
1320
|
+
* Pass NULL if the model has no associated indexer.
|
1321
|
+
* - ref_numeric_data[nref * ncols_numeric]
|
1322
|
+
* Pointer to numeric data for reference points. May be ordered by rows
|
1323
|
+
* (i.e. entries 1..n contain row 0, n+1..2n row 1, etc.) - a.k.a. row-major - or by
|
1324
|
+
* columns (i.e. entries 1..n contain column 0, n+1..2n column 1, etc.) - a.k.a. column-major
|
1325
|
+
* (see parameter 'ref_is_col_major').
|
1326
|
+
* Pass NULL if there are no dense numeric columns or no reference points.
|
1327
|
+
* Can only pass one of 'ref_numeric_data' or 'ref_Xc' + 'ref_Xc_ind' + 'ref_Xc_indptr'.
|
1328
|
+
* If 'indexer' is passed, it has reference points, and the data to which the model was fit had
|
1329
|
+
* numeric columns, then numeric data for reference points must be passed (in either dense or sparse format).
|
1330
|
+
* - ref_categ_data[nref * ncols_categ]
|
1331
|
+
* Pointer to categorical data for reference points. May be ordered by rows
|
1332
|
+
* (i.e. entries 1..n contain row 0, n+1..2n row 1, etc.) - a.k.a. row-major - or by
|
1333
|
+
* columns (i.e. entries 1..n contain column 0, n+1..2n column 1, etc.) - a.k.a. column-major
|
1334
|
+
* (see parameter 'ref_is_col_major').
|
1335
|
+
* Pass NULL if there are no categorical columns or no reference points.
|
1336
|
+
* If 'indexer' is passed, it has reference points, and the data to which the model was fit had
|
1337
|
+
* categorical columns, then 'ref_categ_data' must be passed.
|
1338
|
+
* - ref_is_col_major
|
1339
|
+
* Whether 'ref_numeric_data' and/or 'ref_categ_data' are in column-major order. If numeric data is
|
1340
|
+
* passed in sparse format, categorical data must be passed in column-major format. If passing dense
|
1341
|
+
* data, row-major format is preferred as it will be faster. If the data is passed in row-major format,
|
1342
|
+
* must also pass 'ref_ld_numeric' and/or 'ref_ld_categ'.
|
1343
|
+
* If both 'ref_numeric_data' and 'ref_categ_data' are passed, they must have the same orientation
|
1344
|
+
* (row-major or column-major).
|
1345
|
+
* - ref_ld_numeric
|
1346
|
+
* Leading dimension of the array 'ref_numeric_data', if it is passed in row-major format.
|
1347
|
+
* Typically, this corresponds to the number of columns, but may be larger (the array will
|
1348
|
+
* be accessed assuming that row 'n' starts at 'ref_numeric_data + n*ref_ld_numeric'). If passing
|
1349
|
+
* 'ref_numeric_data' in column-major order, this is ignored and will be assumed that the
|
1350
|
+
* leading dimension corresponds to the number of rows. This is ignored when passing numeric
|
1351
|
+
* data in sparse format.
|
1352
|
+
* - ref_ld_categ
|
1353
|
+
* Leading dimension of the array 'ref_categ_data', if it is passed in row-major format.
|
1354
|
+
* Typically, this corresponds to the number of columns, but may be larger (the array will
|
1355
|
+
* be accessed assuming that row 'n' starts at 'ref_categ_data + n*ref_ld_categ'). If passing
|
1356
|
+
* 'ref_categ_data' in column-major order, this is ignored and will be assumed that the
|
1357
|
+
* leading dimension corresponds to the number of rows.
|
1358
|
+
* - ref_Xc[ref_nnz]
|
1359
|
+
* Pointer to numeric data for reference points in sparse numeric matrix in CSC format (column-compressed).
|
1360
|
+
* Pass NULL if there are no sparse numeric columns for reference points or no reference points.
|
1361
|
+
* Can only pass one of 'ref_numeric_data' or 'ref_Xc' + 'ref_Xc_ind' + 'ref_Xc_indptr'.
|
1362
|
+
* - ref_Xc_ind[ref_nnz]
|
1363
|
+
* Pointer to row indices to which each non-zero entry in 'ref_Xc' corresponds.
|
1364
|
+
* Must be in sorted order, otherwise results will be incorrect.
|
1365
|
+
* Pass NULL if there are no sparse numeric columns in CSC format for reference points or no reference points.
|
1366
|
+
* - ref_Xc_indptr[ref_nnz]
|
1367
|
+
* Pointer to column index pointers that tell at entry [col] where does column 'col'
|
1368
|
+
* start and at entry [col + 1] where does column 'col' end.
|
1369
|
+
* Pass NULL if there are no sparse numeric columns in CSC format for reference points or no reference points.
|
1370
|
+
* - random_seed
|
1371
|
+
* Seed that will be used to generate random numbers used by the model.
|
1372
|
+
* - use_long_double
|
1373
|
+
* Same parameter as for 'fit_iforest' (see the documentation in there for details). Can be changed from
|
1374
|
+
* what was originally passed to 'fit_iforest'.
|
1375
|
+
*/
|
1376
|
+
template <class real_t, class sparse_ix>
|
1377
|
+
int add_tree(IsoForest *model_outputs, ExtIsoForest *model_outputs_ext,
|
1378
|
+
real_t numeric_data[], size_t ncols_numeric,
|
1379
|
+
int categ_data[], size_t ncols_categ, int ncat[],
|
1380
|
+
real_t Xc[], sparse_ix Xc_ind[], sparse_ix Xc_indptr[],
|
1381
|
+
size_t ndim, size_t ntry, CoefType coef_type, bool coef_by_prop,
|
1382
|
+
real_t sample_weights[], size_t nrows,
|
1383
|
+
size_t max_depth, size_t ncols_per_tree,
|
1384
|
+
bool limit_depth, bool penalize_range, bool standardize_data,
|
1385
|
+
bool fast_bratio,
|
1386
|
+
real_t col_weights[], bool weigh_by_kurt,
|
1387
|
+
double prob_pick_by_gain_pl, double prob_pick_by_gain_avg,
|
1388
|
+
double prob_pick_by_full_gain, double prob_pick_by_dens,
|
1389
|
+
double prob_pick_col_by_range, double prob_pick_col_by_var,
|
1390
|
+
double prob_pick_col_by_kurt,
|
1391
|
+
double min_gain, MissingAction missing_action,
|
1392
|
+
CategSplit cat_split_type, NewCategAction new_cat_action,
|
1393
|
+
UseDepthImp depth_imp, WeighImpRows weigh_imp_rows,
|
1394
|
+
bool all_perm, Imputer *imputer, size_t min_imp_obs,
|
1395
|
+
TreesIndexer *indexer,
|
1396
|
+
real_t ref_numeric_data[], int ref_categ_data[],
|
1397
|
+
bool ref_is_col_major, size_t ref_ld_numeric, size_t ref_ld_categ,
|
1398
|
+
real_t ref_Xc[], sparse_ix ref_Xc_ind[], sparse_ix ref_Xc_indptr[],
|
1399
|
+
uint64_t random_seed, bool use_long_double)
|
1400
|
+
{
|
1401
|
+
if (use_long_double && !has_long_double()) {
|
1402
|
+
use_long_double = false;
|
1403
|
+
fprintf(stderr, "Passed 'use_long_double=true', but library was compiled without long double support.\n");
|
1404
|
+
}
|
1405
|
+
#ifndef NO_LONG_DOUBLE
|
1406
|
+
if (likely(!use_long_double))
|
1407
|
+
#endif
|
1408
|
+
return add_tree_internal<real_t, sparse_ix, double>(
|
1409
|
+
model_outputs, model_outputs_ext,
|
1410
|
+
numeric_data, ncols_numeric,
|
1411
|
+
categ_data, ncols_categ, ncat,
|
1412
|
+
Xc, Xc_ind, Xc_indptr,
|
1413
|
+
ndim, ntry, coef_type, coef_by_prop,
|
1414
|
+
sample_weights, nrows,
|
1415
|
+
max_depth, ncols_per_tree,
|
1416
|
+
limit_depth, penalize_range, standardize_data,
|
1417
|
+
fast_bratio,
|
1418
|
+
col_weights, weigh_by_kurt,
|
1419
|
+
prob_pick_by_gain_pl, prob_pick_by_gain_avg,
|
1420
|
+
prob_pick_by_full_gain, prob_pick_by_dens,
|
1421
|
+
prob_pick_col_by_range, prob_pick_col_by_var,
|
1422
|
+
prob_pick_col_by_kurt,
|
1423
|
+
min_gain, missing_action,
|
1424
|
+
cat_split_type, new_cat_action,
|
1425
|
+
depth_imp, weigh_imp_rows,
|
1426
|
+
all_perm, imputer, min_imp_obs,
|
1427
|
+
indexer,
|
1428
|
+
ref_numeric_data, ref_categ_data,
|
1429
|
+
ref_is_col_major, ref_ld_numeric, ref_ld_categ,
|
1430
|
+
ref_Xc, ref_Xc_ind, ref_Xc_indptr,
|
1431
|
+
random_seed
|
1432
|
+
);
|
1433
|
+
#ifndef NO_LONG_DOUBLE
|
1434
|
+
else
|
1435
|
+
return add_tree_internal<real_t, sparse_ix, long double>(
|
1436
|
+
model_outputs, model_outputs_ext,
|
1437
|
+
numeric_data, ncols_numeric,
|
1438
|
+
categ_data, ncols_categ, ncat,
|
1439
|
+
Xc, Xc_ind, Xc_indptr,
|
1440
|
+
ndim, ntry, coef_type, coef_by_prop,
|
1441
|
+
sample_weights, nrows,
|
1442
|
+
max_depth, ncols_per_tree,
|
1443
|
+
limit_depth, penalize_range, standardize_data,
|
1444
|
+
fast_bratio,
|
1445
|
+
col_weights, weigh_by_kurt,
|
1446
|
+
prob_pick_by_gain_pl, prob_pick_by_gain_avg,
|
1447
|
+
prob_pick_by_full_gain, prob_pick_by_dens,
|
1448
|
+
prob_pick_col_by_range, prob_pick_col_by_var,
|
1449
|
+
prob_pick_col_by_kurt,
|
1450
|
+
min_gain, missing_action,
|
1451
|
+
cat_split_type, new_cat_action,
|
1452
|
+
depth_imp, weigh_imp_rows,
|
1453
|
+
all_perm, imputer, min_imp_obs,
|
1454
|
+
indexer,
|
1455
|
+
ref_numeric_data, ref_categ_data,
|
1456
|
+
ref_is_col_major, ref_ld_numeric, ref_ld_categ,
|
1457
|
+
ref_Xc, ref_Xc_ind, ref_Xc_indptr,
|
1458
|
+
random_seed
|
1459
|
+
);
|
1460
|
+
#endif
|
1461
|
+
}
|
1462
|
+
|
1463
|
+
template <class real_t, class sparse_ix, class ldouble_safe>
|
1464
|
+
int add_tree_internal(
|
1465
|
+
IsoForest *model_outputs, ExtIsoForest *model_outputs_ext,
|
1466
|
+
real_t numeric_data[], size_t ncols_numeric,
|
1467
|
+
int categ_data[], size_t ncols_categ, int ncat[],
|
1468
|
+
real_t Xc[], sparse_ix Xc_ind[], sparse_ix Xc_indptr[],
|
1469
|
+
size_t ndim, size_t ntry, CoefType coef_type, bool coef_by_prop,
|
1470
|
+
real_t sample_weights[], size_t nrows,
|
1471
|
+
size_t max_depth, size_t ncols_per_tree,
|
1472
|
+
bool limit_depth, bool penalize_range, bool standardize_data,
|
1473
|
+
bool fast_bratio,
|
1474
|
+
real_t col_weights[], bool weigh_by_kurt,
|
1475
|
+
double prob_pick_by_gain_pl, double prob_pick_by_gain_avg,
|
1476
|
+
double prob_pick_by_full_gain, double prob_pick_by_dens,
|
1477
|
+
double prob_pick_col_by_range, double prob_pick_col_by_var,
|
1478
|
+
double prob_pick_col_by_kurt,
|
1479
|
+
double min_gain, MissingAction missing_action,
|
1480
|
+
CategSplit cat_split_type, NewCategAction new_cat_action,
|
1481
|
+
UseDepthImp depth_imp, WeighImpRows weigh_imp_rows,
|
1482
|
+
bool all_perm, Imputer *imputer, size_t min_imp_obs,
|
1483
|
+
TreesIndexer *indexer,
|
1484
|
+
real_t ref_numeric_data[], int ref_categ_data[],
|
1485
|
+
bool ref_is_col_major, size_t ref_ld_numeric, size_t ref_ld_categ,
|
1486
|
+
real_t ref_Xc[], sparse_ix ref_Xc_ind[], sparse_ix ref_Xc_indptr[],
|
1487
|
+
uint64_t random_seed)
|
1488
|
+
{
|
1489
|
+
if (
|
1490
|
+
prob_pick_by_gain_avg < 0 || prob_pick_by_gain_pl < 0 ||
|
1491
|
+
prob_pick_by_full_gain < 0 || prob_pick_by_dens < 0 ||
|
1492
|
+
prob_pick_col_by_range < 0 ||
|
1493
|
+
prob_pick_col_by_var < 0 || prob_pick_col_by_kurt < 0
|
1494
|
+
) {
|
1495
|
+
throw std::runtime_error("Cannot pass negative probabilities.\n");
|
1496
|
+
}
|
1497
|
+
if (prob_pick_col_by_range && ncols_categ)
|
1498
|
+
throw std::runtime_error("'prob_pick_col_by_range' is not compatible with categorical data.\n");
|
1499
|
+
if (prob_pick_by_full_gain && ncols_categ)
|
1500
|
+
throw std::runtime_error("'prob_pick_by_full_gain' is not compatible with categorical data.\n");
|
1501
|
+
if (prob_pick_col_by_kurt && weigh_by_kurt)
|
1502
|
+
throw std::runtime_error("'weigh_by_kurt' and 'prob_pick_col_by_kurt' cannot be used together.\n");
|
1503
|
+
if (ndim == 0 && model_outputs == NULL)
|
1504
|
+
throw std::runtime_error("Must pass 'ndim>0' in the extended model.\n");
|
1505
|
+
if (indexer != NULL && !indexer->indices.empty() && !indexer->indices.front().reference_points.empty()) {
|
1506
|
+
if (ref_numeric_data == NULL && ref_categ_data == NULL && ref_Xc_indptr == NULL)
|
1507
|
+
throw std::runtime_error("'indexer' has reference points. Those points must be passed to index them in the new tree to add.\n");
|
1508
|
+
}
|
1509
|
+
|
1510
|
+
std::vector<ImputeNode> *impute_nodes = NULL;
|
1511
|
+
|
1512
|
+
int max_categ = 0;
|
1513
|
+
for (size_t col = 0; col < ncols_categ; col++)
|
1514
|
+
max_categ = (ncat[col] > max_categ)? ncat[col] : max_categ;
|
1515
|
+
|
1516
|
+
if (model_outputs != NULL)
|
1517
|
+
ntry = std::min(ntry, ncols_numeric + ncols_categ);
|
1518
|
+
|
1519
|
+
if (ncols_per_tree == 0)
|
1520
|
+
ncols_per_tree = ncols_numeric + ncols_categ;
|
1521
|
+
|
1522
|
+
if (indexer != NULL && indexer->indices.empty())
|
1523
|
+
indexer = NULL;
|
1524
|
+
|
1525
|
+
InputData<real_t, sparse_ix>
|
1526
|
+
input_data = {numeric_data, ncols_numeric, categ_data, ncat, max_categ, ncols_categ,
|
1527
|
+
nrows, ncols_numeric + ncols_categ, sample_weights,
|
1528
|
+
false, col_weights,
|
1529
|
+
Xc, Xc_ind, Xc_indptr,
|
1530
|
+
0, 0, std::vector<double>(),
|
1531
|
+
std::vector<char>(), 0, NULL,
|
1532
|
+
(double*)NULL, (double*)NULL, (int*)NULL, std::vector<double>(),
|
1533
|
+
std::vector<double>(), std::vector<double>(),
|
1534
|
+
std::vector<size_t>(), std::vector<size_t>()};
|
1535
|
+
ModelParams model_params = {false, nrows, (size_t)1, ncols_per_tree,
|
1536
|
+
max_depth? max_depth : (nrows - 1),
|
1537
|
+
penalize_range, standardize_data, random_seed, weigh_by_kurt,
|
1538
|
+
prob_pick_by_gain_avg, prob_pick_by_gain_pl,
|
1539
|
+
prob_pick_by_full_gain, prob_pick_by_dens,
|
1540
|
+
prob_pick_col_by_range, prob_pick_col_by_var,
|
1541
|
+
prob_pick_col_by_kurt,
|
1542
|
+
min_gain, cat_split_type, new_cat_action, missing_action,
|
1543
|
+
(model_outputs != NULL)? model_outputs->scoring_metric : model_outputs_ext->scoring_metric,
|
1544
|
+
fast_bratio, all_perm,
|
1545
|
+
(model_outputs != NULL)? 0 : ndim, ntry,
|
1546
|
+
coef_type, coef_by_prop, false, false, false, depth_imp, weigh_imp_rows, min_imp_obs};
|
1547
|
+
|
1548
|
+
if (prob_pick_by_full_gain)
|
1549
|
+
{
|
1550
|
+
if (input_data.Xc_indptr == NULL)
|
1551
|
+
colmajor_to_rowmajor(input_data.numeric_data, input_data.nrows, input_data.ncols_numeric, input_data.X_row_major);
|
1552
|
+
else
|
1553
|
+
colmajor_to_rowmajor(input_data.Xc, input_data.Xc_ind, input_data.Xc_indptr,
|
1554
|
+
input_data.nrows, input_data.ncols_numeric,
|
1555
|
+
input_data.Xr, input_data.Xr_ind, input_data.Xr_indptr);
|
1556
|
+
}
|
1557
|
+
|
1558
|
+
std::unique_ptr<WorkerMemory<ImputedData<sparse_ix, ldouble_safe>, ldouble_safe, real_t>> workspace(
|
1559
|
+
new WorkerMemory<ImputedData<sparse_ix, ldouble_safe>, ldouble_safe, real_t>()
|
1560
|
+
);
|
1561
|
+
|
1562
|
+
size_t last_tree;
|
1563
|
+
bool added_tree = false;
|
1564
|
+
try
|
1565
|
+
{
|
1566
|
+
if (model_outputs != NULL)
|
1567
|
+
{
|
1568
|
+
last_tree = model_outputs->trees.size();
|
1569
|
+
model_outputs->trees.emplace_back();
|
1570
|
+
added_tree = true;
|
1571
|
+
}
|
1572
|
+
|
1573
|
+
else
|
1574
|
+
{
|
1575
|
+
last_tree = model_outputs_ext->hplanes.size();
|
1576
|
+
model_outputs_ext->hplanes.emplace_back();
|
1577
|
+
added_tree = true;
|
1578
|
+
}
|
1579
|
+
|
1580
|
+
if (imputer != NULL)
|
1581
|
+
{
|
1582
|
+
imputer->imputer_tree.emplace_back();
|
1583
|
+
impute_nodes = &(imputer->imputer_tree.back());
|
1584
|
+
}
|
1585
|
+
|
1586
|
+
if (indexer != NULL)
|
1587
|
+
{
|
1588
|
+
indexer->indices.emplace_back();
|
1589
|
+
}
|
1590
|
+
|
1591
|
+
SignalSwitcher ss = SignalSwitcher();
|
1592
|
+
check_interrupt_switch(ss);
|
1593
|
+
|
1594
|
+
fit_itree<decltype(input_data), typename std::remove_pointer<decltype(workspace.get())>::type, ldouble_safe>(
|
1595
|
+
(model_outputs != NULL)? &model_outputs->trees.back() : NULL,
|
1596
|
+
(model_outputs_ext != NULL)? &model_outputs_ext->hplanes.back() : NULL,
|
1597
|
+
*workspace,
|
1598
|
+
input_data,
|
1599
|
+
model_params,
|
1600
|
+
impute_nodes,
|
1601
|
+
last_tree);
|
1602
|
+
|
1603
|
+
check_interrupt_switch(ss);
|
1604
|
+
|
1605
|
+
if (model_outputs != NULL) {
|
1606
|
+
model_outputs->trees.back().shrink_to_fit();
|
1607
|
+
model_outputs->has_range_penalty = model_outputs->has_range_penalty || penalize_range;
|
1608
|
+
}
|
1609
|
+
else {
|
1610
|
+
model_outputs_ext->hplanes.back().shrink_to_fit();
|
1611
|
+
model_outputs_ext->has_range_penalty = model_outputs_ext->has_range_penalty || penalize_range;
|
1612
|
+
}
|
1613
|
+
|
1614
|
+
if (imputer != NULL)
|
1615
|
+
imputer->imputer_tree.back().shrink_to_fit();
|
1616
|
+
|
1617
|
+
if (indexer != NULL)
|
1618
|
+
{
|
1619
|
+
if (model_outputs != NULL)
|
1620
|
+
build_terminal_node_mappings_single_tree(indexer->indices.back().terminal_node_mappings,
|
1621
|
+
indexer->indices.back().n_terminal,
|
1622
|
+
model_outputs->trees.back());
|
1623
|
+
else
|
1624
|
+
build_terminal_node_mappings_single_tree(indexer->indices.back().terminal_node_mappings,
|
1625
|
+
indexer->indices.back().n_terminal,
|
1626
|
+
model_outputs_ext->hplanes.back());
|
1627
|
+
|
1628
|
+
check_interrupt_switch(ss);
|
1629
|
+
|
1630
|
+
|
1631
|
+
if (!indexer->indices.front().node_distances.empty())
|
1632
|
+
{
|
1633
|
+
std::vector<size_t> temp;
|
1634
|
+
temp.reserve(indexer->indices.back().n_terminal);
|
1635
|
+
if (model_outputs != NULL) {
|
1636
|
+
build_dindex(
|
1637
|
+
temp,
|
1638
|
+
indexer->indices.back().terminal_node_mappings,
|
1639
|
+
indexer->indices.back().node_distances,
|
1640
|
+
indexer->indices.back().node_depths,
|
1641
|
+
indexer->indices.back().n_terminal,
|
1642
|
+
model_outputs->trees.back()
|
1643
|
+
);
|
1644
|
+
}
|
1645
|
+
else {
|
1646
|
+
build_dindex(
|
1647
|
+
temp,
|
1648
|
+
indexer->indices.back().terminal_node_mappings,
|
1649
|
+
indexer->indices.back().node_distances,
|
1650
|
+
indexer->indices.back().node_depths,
|
1651
|
+
indexer->indices.back().n_terminal,
|
1652
|
+
model_outputs_ext->hplanes.back()
|
1653
|
+
);
|
1654
|
+
}
|
1655
|
+
}
|
1656
|
+
|
1657
|
+
check_interrupt_switch(ss);
|
1658
|
+
if (!indexer->indices.front().reference_points.empty())
|
1659
|
+
{
|
1660
|
+
size_t n_ref = indexer->indices.front().reference_points.size();
|
1661
|
+
std::vector<sparse_ix> terminal_indices(n_ref);
|
1662
|
+
std::unique_ptr<double[]> ignored(new double[n_ref]);
|
1663
|
+
if (model_outputs != NULL)
|
1664
|
+
{
|
1665
|
+
IsoForest single_tree_model;
|
1666
|
+
single_tree_model.new_cat_action = model_outputs->new_cat_action;
|
1667
|
+
single_tree_model.cat_split_type = model_outputs->cat_split_type;
|
1668
|
+
single_tree_model.missing_action = model_outputs->missing_action;
|
1669
|
+
single_tree_model.trees.push_back(model_outputs->trees.back());
|
1670
|
+
|
1671
|
+
predict_iforest(ref_numeric_data, ref_categ_data,
|
1672
|
+
ref_is_col_major, ref_ld_numeric, ref_ld_categ,
|
1673
|
+
ref_Xc, ref_Xc_ind, ref_Xc_indptr,
|
1674
|
+
(real_t*)NULL, (sparse_ix*)NULL, (sparse_ix*)NULL,
|
1675
|
+
n_ref, 1, false,
|
1676
|
+
&single_tree_model, (ExtIsoForest*)NULL,
|
1677
|
+
ignored.get(), terminal_indices.data(),
|
1678
|
+
(double*)NULL,
|
1679
|
+
indexer);
|
1680
|
+
}
|
1681
|
+
|
1682
|
+
else
|
1683
|
+
{
|
1684
|
+
ExtIsoForest single_tree_model;
|
1685
|
+
single_tree_model.new_cat_action = model_outputs_ext->new_cat_action;
|
1686
|
+
single_tree_model.cat_split_type = model_outputs_ext->cat_split_type;
|
1687
|
+
single_tree_model.missing_action = model_outputs_ext->missing_action;
|
1688
|
+
single_tree_model.hplanes.push_back(model_outputs_ext->hplanes.back());
|
1689
|
+
|
1690
|
+
predict_iforest(ref_numeric_data, ref_categ_data,
|
1691
|
+
ref_is_col_major, ref_ld_numeric, ref_ld_categ,
|
1692
|
+
ref_Xc, ref_Xc_ind, ref_Xc_indptr,
|
1693
|
+
(real_t*)NULL, (sparse_ix*)NULL, (sparse_ix*)NULL,
|
1694
|
+
n_ref, 1, false,
|
1695
|
+
(IsoForest*)NULL, &single_tree_model,
|
1696
|
+
ignored.get(), terminal_indices.data(),
|
1697
|
+
(double*)NULL,
|
1698
|
+
indexer);
|
1699
|
+
}
|
1700
|
+
|
1701
|
+
ignored.reset();
|
1702
|
+
indexer->indices.back().reference_points.assign(terminal_indices.begin(), terminal_indices.end());
|
1703
|
+
indexer->indices.back().reference_points.shrink_to_fit();
|
1704
|
+
build_ref_node(indexer->indices.back());
|
1705
|
+
}
|
1706
|
+
|
1707
|
+
check_interrupt_switch(ss);
|
1708
|
+
}
|
1709
|
+
}
|
1710
|
+
|
1711
|
+
catch (...)
|
1712
|
+
{
|
1713
|
+
if (added_tree)
|
1714
|
+
{
|
1715
|
+
if (model_outputs != NULL)
|
1716
|
+
model_outputs->trees.pop_back();
|
1717
|
+
else
|
1718
|
+
model_outputs_ext->hplanes.pop_back();
|
1719
|
+
if (imputer != NULL) {
|
1720
|
+
if (model_outputs != NULL)
|
1721
|
+
imputer->imputer_tree.resize(model_outputs->trees.size());
|
1722
|
+
else
|
1723
|
+
imputer->imputer_tree.resize(model_outputs_ext->hplanes.size());
|
1724
|
+
}
|
1725
|
+
if (indexer != NULL) {
|
1726
|
+
if (model_outputs != NULL)
|
1727
|
+
indexer->indices.resize(model_outputs->trees.size());
|
1728
|
+
else
|
1729
|
+
indexer->indices.resize(model_outputs_ext->hplanes.size());
|
1730
|
+
}
|
1731
|
+
}
|
1732
|
+
throw;
|
1733
|
+
}
|
1734
|
+
|
1735
|
+
return EXIT_SUCCESS;
|
1736
|
+
}
|
1737
|
+
|
1738
|
+
template <class InputData, class WorkerMemory, class ldouble_safe>
|
1739
|
+
void fit_itree(std::vector<IsoTree> *tree_root,
|
1740
|
+
std::vector<IsoHPlane> *hplane_root,
|
1741
|
+
WorkerMemory &workspace,
|
1742
|
+
InputData &input_data,
|
1743
|
+
ModelParams &model_params,
|
1744
|
+
std::vector<ImputeNode> *impute_nodes,
|
1745
|
+
size_t tree_num)
|
1746
|
+
{
|
1747
|
+
/* initialize array for depths if called for */
|
1748
|
+
if (workspace.ix_arr.empty() && model_params.calc_depth)
|
1749
|
+
workspace.row_depths.resize(input_data.nrows, 0);
|
1750
|
+
|
1751
|
+
/* choose random sample of rows */
|
1752
|
+
if (workspace.ix_arr.empty()) workspace.ix_arr.resize(model_params.sample_size);
|
1753
|
+
if (input_data.log2_n > 0)
|
1754
|
+
workspace.btree_weights.assign(input_data.btree_weights_init.begin(),
|
1755
|
+
input_data.btree_weights_init.end());
|
1756
|
+
workspace.rnd_generator.seed(model_params.random_seed + tree_num);
|
1757
|
+
workspace.rbin = UniformUnitInterval(0, 1);
|
1758
|
+
sample_random_rows<typename std::remove_pointer<decltype(input_data.numeric_data)>::type, ldouble_safe>(
|
1759
|
+
workspace.ix_arr, input_data.nrows, model_params.with_replacement,
|
1760
|
+
workspace.rnd_generator, workspace.ix_all,
|
1761
|
+
(input_data.weight_as_sample)? input_data.sample_weights : NULL,
|
1762
|
+
workspace.btree_weights, input_data.log2_n, input_data.btree_offset,
|
1763
|
+
workspace.is_repeated);
|
1764
|
+
workspace.st = 0;
|
1765
|
+
workspace.end = model_params.sample_size - 1;
|
1766
|
+
|
1767
|
+
/* in some cases, it's not possible to use column weights even if they are given,
|
1768
|
+
because every single column will always need to be checked or end up being used. */
|
1769
|
+
bool avoid_col_weights = (tree_root != NULL && model_params.ntry >= model_params.ncols_per_tree &&
|
1770
|
+
model_params.prob_pick_by_gain_avg + model_params.prob_pick_by_gain_pl +
|
1771
|
+
model_params.prob_pick_by_full_gain + model_params.prob_pick_by_dens >= 1)
|
1772
|
+
||
|
1773
|
+
(tree_root == NULL && model_params.ndim >= model_params.ncols_per_tree)
|
1774
|
+
||
|
1775
|
+
(model_params.ncols_per_tree == 1);
|
1776
|
+
if (input_data.preinitialized_col_sampler == NULL)
|
1777
|
+
{
|
1778
|
+
if (input_data.col_weights != NULL && !avoid_col_weights && !model_params.weigh_by_kurt)
|
1779
|
+
workspace.col_sampler.initialize(input_data.col_weights, input_data.ncols_tot);
|
1780
|
+
}
|
1781
|
+
|
1782
|
+
|
1783
|
+
/* set expected tree size and add root node */
|
1784
|
+
{
|
1785
|
+
size_t exp_nodes = mult2(model_params.sample_size);
|
1786
|
+
if (model_params.sample_size >= div2(SIZE_MAX))
|
1787
|
+
exp_nodes = SIZE_MAX;
|
1788
|
+
else if (model_params.max_depth <= (size_t)30)
|
1789
|
+
exp_nodes = std::min(exp_nodes, pow2(model_params.max_depth));
|
1790
|
+
if (tree_root != NULL)
|
1791
|
+
{
|
1792
|
+
tree_root->reserve(exp_nodes);
|
1793
|
+
tree_root->emplace_back();
|
1794
|
+
}
|
1795
|
+
else
|
1796
|
+
{
|
1797
|
+
hplane_root->reserve(exp_nodes);
|
1798
|
+
hplane_root->emplace_back();
|
1799
|
+
}
|
1800
|
+
if (impute_nodes != NULL)
|
1801
|
+
{
|
1802
|
+
impute_nodes->reserve(exp_nodes);
|
1803
|
+
impute_nodes->emplace_back((size_t) 0);
|
1804
|
+
}
|
1805
|
+
}
|
1806
|
+
|
1807
|
+
/* initialize array with candidate categories if not already done */
|
1808
|
+
if (workspace.categs.empty())
|
1809
|
+
workspace.categs.resize(input_data.max_categ);
|
1810
|
+
|
1811
|
+
/* initialize array with per-node column weights if needed */
|
1812
|
+
if ((model_params.prob_pick_col_by_range ||
|
1813
|
+
model_params.prob_pick_col_by_var ||
|
1814
|
+
model_params.prob_pick_col_by_kurt) && workspace.node_col_weights.empty())
|
1815
|
+
{
|
1816
|
+
workspace.node_col_weights.resize(input_data.ncols_tot);
|
1817
|
+
if (tree_root != NULL || model_params.standardize_data || model_params.missing_action != Fail)
|
1818
|
+
{
|
1819
|
+
workspace.saved_stat1.resize(input_data.ncols_numeric);
|
1820
|
+
workspace.saved_stat2.resize(input_data.ncols_numeric);
|
1821
|
+
}
|
1822
|
+
}
|
1823
|
+
|
1824
|
+
/* IMPORTANT!!!!!
|
1825
|
+
The standard library implementation is likely going to use the Box-Muller method
|
1826
|
+
for normal sampling, which has some state memory in the **distribution object itself**
|
1827
|
+
in addition to the state memory from the RNG engine. DO NOT avoid re-generating this
|
1828
|
+
object on each tree, despite being inefficient, because then it can cause seed
|
1829
|
+
irreproducibility when the number of splitting dimensions is odd and the number
|
1830
|
+
of threads is more than 1. This is a very hard issue to debug since everything
|
1831
|
+
works fine depending on the order in which trees are assigned to threads.
|
1832
|
+
DO NOT PUT THESE LINES BELOW THE NEXT IF. */
|
1833
|
+
if (hplane_root != NULL)
|
1834
|
+
{
|
1835
|
+
if (input_data.ncols_categ || model_params.coef_type == Normal)
|
1836
|
+
workspace.coef_norm = StandardNormalDistr(0, 1);
|
1837
|
+
if (model_params.coef_type == Uniform)
|
1838
|
+
workspace.coef_unif = UniformMinusOneToOne(-1, 1);
|
1839
|
+
}
|
1840
|
+
|
1841
|
+
/* for the extended model, initialize extra vectors and objects */
|
1842
|
+
if (hplane_root != NULL && workspace.comb_val.empty())
|
1843
|
+
{
|
1844
|
+
workspace.comb_val.resize(model_params.sample_size);
|
1845
|
+
workspace.col_take.resize(model_params.ndim);
|
1846
|
+
workspace.col_take_type.resize(model_params.ndim);
|
1847
|
+
|
1848
|
+
if (input_data.ncols_numeric)
|
1849
|
+
{
|
1850
|
+
workspace.ext_offset.resize(input_data.ncols_tot);
|
1851
|
+
workspace.ext_coef.resize(input_data.ncols_tot);
|
1852
|
+
workspace.ext_mean.resize(input_data.ncols_tot);
|
1853
|
+
}
|
1854
|
+
|
1855
|
+
if (input_data.ncols_categ)
|
1856
|
+
{
|
1857
|
+
workspace.ext_fill_new.resize(input_data.max_categ);
|
1858
|
+
switch(model_params.cat_split_type)
|
1859
|
+
{
|
1860
|
+
case SingleCateg:
|
1861
|
+
{
|
1862
|
+
workspace.chosen_cat.resize(input_data.max_categ);
|
1863
|
+
break;
|
1864
|
+
}
|
1865
|
+
|
1866
|
+
case SubSet:
|
1867
|
+
{
|
1868
|
+
workspace.ext_cat_coef.resize(input_data.ncols_tot);
|
1869
|
+
for (std::vector<double> &v : workspace.ext_cat_coef)
|
1870
|
+
v.resize(input_data.max_categ);
|
1871
|
+
break;
|
1872
|
+
}
|
1873
|
+
}
|
1874
|
+
}
|
1875
|
+
|
1876
|
+
workspace.ext_fill_val.resize(input_data.ncols_tot);
|
1877
|
+
|
1878
|
+
}
|
1879
|
+
|
1880
|
+
/* If there are density weights, need to standardize them to sum up to
|
1881
|
+
the sample size here. Note that weights for missing values with 'Divide'
|
1882
|
+
are only initialized on-demand later on. */
|
1883
|
+
workspace.changed_weights = false;
|
1884
|
+
if (hplane_root == NULL) workspace.weights_map.clear();
|
1885
|
+
|
1886
|
+
ldouble_safe weight_scaling = 0;
|
1887
|
+
if (input_data.sample_weights != NULL && !input_data.weight_as_sample)
|
1888
|
+
{
|
1889
|
+
workspace.changed_weights = true;
|
1890
|
+
|
1891
|
+
/* For the extended model, if there is no sub-sampling, these weights will remain
|
1892
|
+
constant throughout and do not need to be re-generated. */
|
1893
|
+
if (!( hplane_root != NULL &&
|
1894
|
+
(!workspace.weights_map.empty() || !workspace.weights_arr.empty()) &&
|
1895
|
+
model_params.sample_size == input_data.nrows && !model_params.with_replacement
|
1896
|
+
)
|
1897
|
+
)
|
1898
|
+
{
|
1899
|
+
workspace.weights_map.clear();
|
1900
|
+
|
1901
|
+
/* if the sub-sample size is small relative to the full sample size, use a mapping */
|
1902
|
+
if (input_data.Xc_indptr != NULL && model_params.sample_size < input_data.nrows / 50)
|
1903
|
+
{
|
1904
|
+
for (const size_t ix : workspace.ix_arr)
|
1905
|
+
weight_scaling += input_data.sample_weights[ix];
|
1906
|
+
weight_scaling = (ldouble_safe)model_params.sample_size / weight_scaling;
|
1907
|
+
workspace.weights_map.reserve(workspace.ix_arr.size());
|
1908
|
+
for (const size_t ix : workspace.ix_arr)
|
1909
|
+
workspace.weights_map[ix] = input_data.sample_weights[ix] * weight_scaling;
|
1910
|
+
}
|
1911
|
+
|
1912
|
+
/* if the sub-sample size is large, fill a full array matching to the sample size */
|
1913
|
+
else
|
1914
|
+
{
|
1915
|
+
if (workspace.weights_arr.empty())
|
1916
|
+
{
|
1917
|
+
workspace.weights_arr.assign(input_data.sample_weights, input_data.sample_weights + input_data.nrows);
|
1918
|
+
weight_scaling = std::accumulate(workspace.ix_arr.begin(),
|
1919
|
+
workspace.ix_arr.end(),
|
1920
|
+
(ldouble_safe)0,
|
1921
|
+
[&input_data](const ldouble_safe a, const size_t b){return a + (ldouble_safe)input_data.sample_weights[b];}
|
1922
|
+
);
|
1923
|
+
weight_scaling = (ldouble_safe)model_params.sample_size / weight_scaling;
|
1924
|
+
for (double &w : workspace.weights_arr)
|
1925
|
+
w *= weight_scaling;
|
1926
|
+
}
|
1927
|
+
|
1928
|
+
else
|
1929
|
+
{
|
1930
|
+
for (const size_t ix : workspace.ix_arr)
|
1931
|
+
{
|
1932
|
+
weight_scaling += input_data.sample_weights[ix];
|
1933
|
+
workspace.weights_arr[ix] = input_data.sample_weights[ix];
|
1934
|
+
}
|
1935
|
+
weight_scaling = (ldouble_safe)model_params.sample_size / weight_scaling;
|
1936
|
+
for (double &w : workspace.weights_arr)
|
1937
|
+
w *= weight_scaling;
|
1938
|
+
}
|
1939
|
+
}
|
1940
|
+
}
|
1941
|
+
}
|
1942
|
+
|
1943
|
+
/* if producing distance/similarity, also need to initialize the triangular matrix */
|
1944
|
+
if (model_params.calc_dist && workspace.tmat_sep.empty())
|
1945
|
+
workspace.tmat_sep.resize((input_data.nrows * (input_data.nrows - 1)) / 2, 0);
|
1946
|
+
|
1947
|
+
/* make space for buffers if not already allocated */
|
1948
|
+
if (
|
1949
|
+
(model_params.prob_pick_by_gain_avg > 0 ||
|
1950
|
+
model_params.prob_pick_by_gain_pl > 0 ||
|
1951
|
+
model_params.prob_pick_by_full_gain > 0 ||
|
1952
|
+
model_params.prob_pick_by_dens > 0 ||
|
1953
|
+
model_params.prob_pick_col_by_range > 0 ||
|
1954
|
+
model_params.prob_pick_col_by_var > 0 ||
|
1955
|
+
model_params.prob_pick_col_by_kurt > 0 ||
|
1956
|
+
model_params.weigh_by_kurt || hplane_root != NULL)
|
1957
|
+
&&
|
1958
|
+
(workspace.buffer_dbl.empty() && workspace.buffer_szt.empty() && workspace.buffer_chr.empty())
|
1959
|
+
)
|
1960
|
+
{
|
1961
|
+
size_t min_size_dbl = 0;
|
1962
|
+
size_t min_size_szt = 0;
|
1963
|
+
size_t min_size_chr = 0;
|
1964
|
+
|
1965
|
+
bool gain = model_params.prob_pick_by_gain_avg > 0 ||
|
1966
|
+
model_params.prob_pick_by_gain_pl > 0 ||
|
1967
|
+
model_params.prob_pick_by_full_gain > 0 ||
|
1968
|
+
model_params.prob_pick_by_dens > 0;
|
1969
|
+
|
1970
|
+
if (input_data.ncols_categ)
|
1971
|
+
{
|
1972
|
+
min_size_szt = (size_t)2 * (size_t)input_data.max_categ;
|
1973
|
+
min_size_dbl = input_data.max_categ + 1;
|
1974
|
+
if (gain && model_params.cat_split_type == SubSet)
|
1975
|
+
min_size_chr = input_data.max_categ;
|
1976
|
+
}
|
1977
|
+
|
1978
|
+
if (input_data.Xc_indptr != NULL && gain)
|
1979
|
+
{
|
1980
|
+
min_size_szt = std::max(min_size_szt, model_params.sample_size);
|
1981
|
+
min_size_dbl = std::max(min_size_dbl, model_params.sample_size);
|
1982
|
+
}
|
1983
|
+
|
1984
|
+
/* TODO: revisit if this covers all the cases */
|
1985
|
+
if (model_params.ntry > 1 || gain)
|
1986
|
+
{
|
1987
|
+
min_size_dbl = std::max(min_size_dbl, model_params.sample_size);
|
1988
|
+
if (model_params.ndim < 2 && input_data.Xc_indptr != NULL)
|
1989
|
+
min_size_dbl = std::max(min_size_dbl, (size_t)2*model_params.sample_size);
|
1990
|
+
}
|
1991
|
+
|
1992
|
+
/* for sampled column choices */
|
1993
|
+
if (model_params.prob_pick_col_by_var)
|
1994
|
+
{
|
1995
|
+
if (input_data.ncols_categ) {
|
1996
|
+
min_size_szt = std::max(min_size_szt, (size_t)input_data.max_categ + 1);
|
1997
|
+
min_size_dbl = std::max(min_size_dbl, (size_t)input_data.max_categ + 1);
|
1998
|
+
}
|
1999
|
+
}
|
2000
|
+
|
2001
|
+
if (model_params.prob_pick_col_by_kurt)
|
2002
|
+
{
|
2003
|
+
if (input_data.ncols_categ) {
|
2004
|
+
min_size_szt = std::max(min_size_szt, (size_t)input_data.max_categ + 1);
|
2005
|
+
min_size_dbl = std::max(min_size_dbl, (size_t)input_data.max_categ);
|
2006
|
+
}
|
2007
|
+
|
2008
|
+
}
|
2009
|
+
|
2010
|
+
/* for the extended model */
|
2011
|
+
if (hplane_root != NULL)
|
2012
|
+
{
|
2013
|
+
min_size_dbl = std::max(min_size_dbl, pow2(log2ceil(input_data.ncols_tot) + 1));
|
2014
|
+
if (model_params.missing_action != Fail)
|
2015
|
+
{
|
2016
|
+
min_size_szt = std::max(min_size_szt, model_params.sample_size);
|
2017
|
+
min_size_dbl = std::max(min_size_dbl, model_params.sample_size);
|
2018
|
+
}
|
2019
|
+
|
2020
|
+
if (input_data.ncols_categ && model_params.cat_split_type == SubSet)
|
2021
|
+
{
|
2022
|
+
min_size_szt = std::max(min_size_szt, (size_t)2 * (size_t)input_data.max_categ + (size_t)1);
|
2023
|
+
min_size_dbl = std::max(min_size_dbl, (size_t)input_data.max_categ);
|
2024
|
+
}
|
2025
|
+
|
2026
|
+
if (model_params.weigh_by_kurt)
|
2027
|
+
min_size_szt = std::max(min_size_szt, input_data.ncols_tot);
|
2028
|
+
|
2029
|
+
if (gain && (!workspace.weights_arr.empty() || !workspace.weights_map.empty()))
|
2030
|
+
{
|
2031
|
+
workspace.sample_weights.resize(model_params.sample_size);
|
2032
|
+
min_size_szt = std::max(min_size_szt, model_params.sample_size);
|
2033
|
+
}
|
2034
|
+
}
|
2035
|
+
|
2036
|
+
/* now resize */
|
2037
|
+
if (workspace.buffer_dbl.size() < min_size_dbl)
|
2038
|
+
workspace.buffer_dbl.resize(min_size_dbl);
|
2039
|
+
|
2040
|
+
if (workspace.buffer_szt.size() < min_size_szt)
|
2041
|
+
workspace.buffer_szt.resize(min_size_szt);
|
2042
|
+
|
2043
|
+
if (workspace.buffer_chr.size() < min_size_chr)
|
2044
|
+
workspace.buffer_chr.resize(min_size_chr);
|
2045
|
+
|
2046
|
+
/* for guided column choice, need to also remember the best split so far */
|
2047
|
+
if (
|
2048
|
+
model_params.cat_split_type == SubSet &&
|
2049
|
+
(
|
2050
|
+
model_params.prob_pick_by_gain_avg ||
|
2051
|
+
model_params.prob_pick_by_gain_pl ||
|
2052
|
+
model_params.prob_pick_by_full_gain ||
|
2053
|
+
model_params.prob_pick_by_dens
|
2054
|
+
)
|
2055
|
+
)
|
2056
|
+
{
|
2057
|
+
workspace.this_split_categ.resize(input_data.max_categ);
|
2058
|
+
}
|
2059
|
+
|
2060
|
+
}
|
2061
|
+
|
2062
|
+
/* Other potentially necessary buffers */
|
2063
|
+
if (
|
2064
|
+
tree_root != NULL && model_params.missing_action == Impute &&
|
2065
|
+
(model_params.prob_pick_by_gain_avg || model_params.prob_pick_by_gain_pl ||
|
2066
|
+
model_params.prob_pick_by_full_gain || model_params.prob_pick_by_dens) &&
|
2067
|
+
input_data.Xc_indptr == NULL && input_data.ncols_numeric && workspace.imputed_x_buffer.empty()
|
2068
|
+
)
|
2069
|
+
{
|
2070
|
+
workspace.imputed_x_buffer.resize(input_data.nrows);
|
2071
|
+
}
|
2072
|
+
|
2073
|
+
if (model_params.prob_pick_by_full_gain && workspace.col_indices.empty())
|
2074
|
+
workspace.col_indices.resize(model_params.ncols_per_tree);
|
2075
|
+
|
2076
|
+
if (
|
2077
|
+
(model_params.prob_pick_col_by_range || model_params.prob_pick_col_by_var) &&
|
2078
|
+
model_params.weigh_by_kurt &&
|
2079
|
+
model_params.sample_size == input_data.nrows && !model_params.with_replacement &&
|
2080
|
+
(model_params.ncols_per_tree == input_data.ncols_tot) &&
|
2081
|
+
!input_data.all_kurtoses.empty()
|
2082
|
+
) {
|
2083
|
+
workspace.tree_kurtoses = input_data.all_kurtoses.data();
|
2084
|
+
}
|
2085
|
+
else {
|
2086
|
+
workspace.tree_kurtoses = NULL;
|
2087
|
+
}
|
2088
|
+
|
2089
|
+
/* weigh columns by kurtosis in the sample if required */
|
2090
|
+
/* TODO: this one could probably be refactored to use the function in the helpers */
|
2091
|
+
std::vector<double> kurt_weights;
|
2092
|
+
bool avoid_leave_m_cols = false;
|
2093
|
+
if (
|
2094
|
+
model_params.weigh_by_kurt &&
|
2095
|
+
!avoid_col_weights &&
|
2096
|
+
(input_data.preinitialized_col_sampler == NULL
|
2097
|
+
||
|
2098
|
+
((model_params.prob_pick_col_by_range || model_params.prob_pick_col_by_var) && workspace.tree_kurtoses == NULL))
|
2099
|
+
)
|
2100
|
+
{
|
2101
|
+
kurt_weights.resize(input_data.ncols_numeric + input_data.ncols_categ, 0.);
|
2102
|
+
|
2103
|
+
if (model_params.ncols_per_tree >= input_data.ncols_tot)
|
2104
|
+
{
|
2105
|
+
|
2106
|
+
if (input_data.Xc_indptr == NULL)
|
2107
|
+
{
|
2108
|
+
|
2109
|
+
for (size_t col = 0; col < input_data.ncols_numeric; col++)
|
2110
|
+
{
|
2111
|
+
if (workspace.weights_arr.empty() && workspace.weights_map.empty())
|
2112
|
+
kurt_weights[col] = calc_kurtosis<typename std::remove_pointer<decltype(input_data.numeric_data)>::type, ldouble_safe>(
|
2113
|
+
workspace.ix_arr.data(), workspace.st, workspace.end,
|
2114
|
+
input_data.numeric_data + col * input_data.nrows,
|
2115
|
+
model_params.missing_action);
|
2116
|
+
else if (!workspace.weights_arr.empty())
|
2117
|
+
kurt_weights[col] = calc_kurtosis_weighted<typename std::remove_pointer<decltype(input_data.numeric_data)>::type, decltype(workspace.weights_arr), ldouble_safe>(
|
2118
|
+
workspace.ix_arr.data(), workspace.st, workspace.end,
|
2119
|
+
input_data.numeric_data + col * input_data.nrows,
|
2120
|
+
model_params.missing_action, workspace.weights_arr);
|
2121
|
+
else
|
2122
|
+
kurt_weights[col] = calc_kurtosis_weighted<typename std::remove_pointer<decltype(input_data.numeric_data)>::type,
|
2123
|
+
decltype(workspace.weights_map), ldouble_safe>(
|
2124
|
+
workspace.ix_arr.data(), workspace.st, workspace.end,
|
2125
|
+
input_data.numeric_data + col * input_data.nrows,
|
2126
|
+
model_params.missing_action, workspace.weights_map);
|
2127
|
+
}
|
2128
|
+
}
|
2129
|
+
|
2130
|
+
else
|
2131
|
+
{
|
2132
|
+
std::sort(workspace.ix_arr.begin(), workspace.ix_arr.end());
|
2133
|
+
for (size_t col = 0; col < input_data.ncols_numeric; col++)
|
2134
|
+
{
|
2135
|
+
if (workspace.weights_arr.empty() && workspace.weights_map.empty())
|
2136
|
+
kurt_weights[col] = calc_kurtosis<typename std::remove_pointer<decltype(input_data.Xc)>::type,
|
2137
|
+
typename std::remove_pointer<decltype(input_data.Xc_indptr)>::type,
|
2138
|
+
ldouble_safe>(
|
2139
|
+
workspace.ix_arr.data(), workspace.st, workspace.end, col,
|
2140
|
+
input_data.Xc, input_data.Xc_ind, input_data.Xc_indptr,
|
2141
|
+
model_params.missing_action);
|
2142
|
+
else if (!workspace.weights_arr.empty())
|
2143
|
+
kurt_weights[col] = calc_kurtosis_weighted<typename std::remove_pointer<decltype(input_data.Xc)>::type,
|
2144
|
+
typename std::remove_pointer<decltype(input_data.Xc_indptr)>::type,
|
2145
|
+
decltype(workspace.weights_arr), ldouble_safe>(
|
2146
|
+
workspace.ix_arr.data(), workspace.st, workspace.end, col,
|
2147
|
+
input_data.Xc, input_data.Xc_ind, input_data.Xc_indptr,
|
2148
|
+
model_params.missing_action, workspace.weights_arr);
|
2149
|
+
else
|
2150
|
+
kurt_weights[col] = calc_kurtosis_weighted<typename std::remove_pointer<decltype(input_data.Xc)>::type,
|
2151
|
+
typename std::remove_pointer<decltype(input_data.Xc_indptr)>::type,
|
2152
|
+
decltype(workspace.weights_map), ldouble_safe>(
|
2153
|
+
workspace.ix_arr.data(), workspace.st, workspace.end, col,
|
2154
|
+
input_data.Xc, input_data.Xc_ind, input_data.Xc_indptr,
|
2155
|
+
model_params.missing_action, workspace.weights_map);
|
2156
|
+
}
|
2157
|
+
}
|
2158
|
+
|
2159
|
+
for (size_t col = 0; col < input_data.ncols_categ; col++)
|
2160
|
+
{
|
2161
|
+
if (workspace.weights_arr.empty() && workspace.weights_map.empty())
|
2162
|
+
kurt_weights[col + input_data.ncols_numeric] =
|
2163
|
+
calc_kurtosis<ldouble_safe>(
|
2164
|
+
workspace.ix_arr.data(), workspace.st, workspace.end,
|
2165
|
+
input_data.categ_data + col * input_data.nrows, input_data.ncat[col],
|
2166
|
+
workspace.buffer_szt.data(), workspace.buffer_dbl.data(),
|
2167
|
+
model_params.missing_action, model_params.cat_split_type, workspace.rnd_generator);
|
2168
|
+
else if (!workspace.weights_arr.empty())
|
2169
|
+
kurt_weights[col + input_data.ncols_numeric] =
|
2170
|
+
calc_kurtosis_weighted<decltype(workspace.weights_arr), ldouble_safe>(
|
2171
|
+
workspace.ix_arr.data(), workspace.st, workspace.end,
|
2172
|
+
input_data.categ_data + col * input_data.nrows, input_data.ncat[col],
|
2173
|
+
workspace.buffer_dbl.data(),
|
2174
|
+
model_params.missing_action, model_params.cat_split_type, workspace.rnd_generator,
|
2175
|
+
workspace.weights_arr);
|
2176
|
+
else
|
2177
|
+
kurt_weights[col + input_data.ncols_numeric] =
|
2178
|
+
calc_kurtosis_weighted<decltype(workspace.weights_map), ldouble_safe>(
|
2179
|
+
workspace.ix_arr.data(), workspace.st, workspace.end,
|
2180
|
+
input_data.categ_data + col * input_data.nrows, input_data.ncat[col],
|
2181
|
+
workspace.buffer_dbl.data(),
|
2182
|
+
model_params.missing_action, model_params.cat_split_type, workspace.rnd_generator,
|
2183
|
+
workspace.weights_map);
|
2184
|
+
}
|
2185
|
+
|
2186
|
+
for (auto &w : kurt_weights) w = (w == -HUGE_VAL)? 0. : std::fmax(1e-8, -1. + w);
|
2187
|
+
if (input_data.col_weights != NULL)
|
2188
|
+
{
|
2189
|
+
for (size_t col = 0; col < input_data.ncols_tot; col++)
|
2190
|
+
{
|
2191
|
+
if (kurt_weights[col] <= 0) continue;
|
2192
|
+
kurt_weights[col] *= input_data.col_weights[col];
|
2193
|
+
kurt_weights[col] = std::fmax(kurt_weights[col], 1e-100);
|
2194
|
+
}
|
2195
|
+
}
|
2196
|
+
workspace.col_sampler.initialize(kurt_weights.data(), kurt_weights.size());
|
2197
|
+
}
|
2198
|
+
|
2199
|
+
|
2200
|
+
|
2201
|
+
else
|
2202
|
+
{
|
2203
|
+
std::vector<size_t> cols_take(model_params.ncols_per_tree);
|
2204
|
+
std::vector<size_t> buffer1;
|
2205
|
+
std::vector<bool> buffer2;
|
2206
|
+
sample_random_rows<double, double>(
|
2207
|
+
cols_take, input_data.ncols_tot, false,
|
2208
|
+
workspace.rnd_generator, buffer1,
|
2209
|
+
(double*)NULL, kurt_weights, /* <- will not get used */
|
2210
|
+
(size_t)0, (size_t)0, buffer2);
|
2211
|
+
|
2212
|
+
if (
|
2213
|
+
model_params.sample_size == input_data.nrows &&
|
2214
|
+
!model_params.with_replacement &&
|
2215
|
+
!input_data.all_kurtoses.empty()
|
2216
|
+
)
|
2217
|
+
{
|
2218
|
+
for (size_t col : cols_take)
|
2219
|
+
kurt_weights[col] = input_data.all_kurtoses[col];
|
2220
|
+
goto skip_kurt_calculations;
|
2221
|
+
}
|
2222
|
+
|
2223
|
+
if (input_data.Xc_indptr != NULL)
|
2224
|
+
std::sort(workspace.ix_arr.begin(), workspace.ix_arr.end());
|
2225
|
+
|
2226
|
+
for (size_t col : cols_take)
|
2227
|
+
{
|
2228
|
+
if (col < input_data.ncols_numeric)
|
2229
|
+
{
|
2230
|
+
if (input_data.Xc_indptr == NULL)
|
2231
|
+
{
|
2232
|
+
if (workspace.weights_arr.empty() && workspace.weights_map.empty())
|
2233
|
+
kurt_weights[col] = calc_kurtosis<typename std::remove_pointer<decltype(input_data.numeric_data)>::type, ldouble_safe>(
|
2234
|
+
workspace.ix_arr.data(), workspace.st, workspace.end,
|
2235
|
+
input_data.numeric_data + col * input_data.nrows,
|
2236
|
+
model_params.missing_action);
|
2237
|
+
else if (!workspace.weights_arr.empty())
|
2238
|
+
kurt_weights[col] = calc_kurtosis_weighted<typename std::remove_pointer<decltype(input_data.numeric_data)>::type,
|
2239
|
+
decltype(workspace.weights_arr), ldouble_safe>(
|
2240
|
+
workspace.ix_arr.data(), workspace.st, workspace.end,
|
2241
|
+
input_data.numeric_data + col * input_data.nrows,
|
2242
|
+
model_params.missing_action, workspace.weights_arr);
|
2243
|
+
else
|
2244
|
+
kurt_weights[col] = calc_kurtosis_weighted<typename std::remove_pointer<decltype(input_data.numeric_data)>::type,
|
2245
|
+
decltype(workspace.weights_map), ldouble_safe>(
|
2246
|
+
workspace.ix_arr.data(), workspace.st, workspace.end,
|
2247
|
+
input_data.numeric_data + col * input_data.nrows,
|
2248
|
+
model_params.missing_action, workspace.weights_map);
|
2249
|
+
}
|
2250
|
+
|
2251
|
+
else
|
2252
|
+
{
|
2253
|
+
if (workspace.weights_arr.empty() && workspace.weights_map.empty())
|
2254
|
+
kurt_weights[col] = calc_kurtosis<typename std::remove_pointer<decltype(input_data.Xc)>::type,
|
2255
|
+
typename std::remove_pointer<decltype(input_data.Xc_indptr)>::type,
|
2256
|
+
ldouble_safe>(
|
2257
|
+
workspace.ix_arr.data(), workspace.st, workspace.end, col,
|
2258
|
+
input_data.Xc, input_data.Xc_ind, input_data.Xc_indptr,
|
2259
|
+
model_params.missing_action);
|
2260
|
+
else if (!workspace.weights_arr.empty())
|
2261
|
+
kurt_weights[col] = calc_kurtosis_weighted<typename std::remove_pointer<decltype(input_data.Xc)>::type,
|
2262
|
+
typename std::remove_pointer<decltype(input_data.Xc_indptr)>::type,
|
2263
|
+
decltype(workspace.weights_arr), ldouble_safe>(
|
2264
|
+
workspace.ix_arr.data(), workspace.st, workspace.end, col,
|
2265
|
+
input_data.Xc, input_data.Xc_ind, input_data.Xc_indptr,
|
2266
|
+
model_params.missing_action, workspace.weights_arr);
|
2267
|
+
else
|
2268
|
+
kurt_weights[col] = calc_kurtosis_weighted<typename std::remove_pointer<decltype(input_data.Xc)>::type,
|
2269
|
+
typename std::remove_pointer<decltype(input_data.Xc_indptr)>::type,
|
2270
|
+
decltype(workspace.weights_map), ldouble_safe>(
|
2271
|
+
workspace.ix_arr.data(), workspace.st, workspace.end, col,
|
2272
|
+
input_data.Xc, input_data.Xc_ind, input_data.Xc_indptr,
|
2273
|
+
model_params.missing_action, workspace.weights_map);
|
2274
|
+
}
|
2275
|
+
}
|
2276
|
+
|
2277
|
+
else
|
2278
|
+
{
|
2279
|
+
if (workspace.weights_arr.empty() && workspace.weights_map.empty())
|
2280
|
+
kurt_weights[col] =
|
2281
|
+
calc_kurtosis<ldouble_safe>(
|
2282
|
+
workspace.ix_arr.data(), workspace.st, workspace.end,
|
2283
|
+
input_data.categ_data + (col - input_data.ncols_numeric) * input_data.nrows,
|
2284
|
+
input_data.ncat[col - input_data.ncols_numeric],
|
2285
|
+
workspace.buffer_szt.data(), workspace.buffer_dbl.data(),
|
2286
|
+
model_params.missing_action, model_params.cat_split_type, workspace.rnd_generator);
|
2287
|
+
else if (!workspace.weights_arr.empty())
|
2288
|
+
kurt_weights[col] =
|
2289
|
+
calc_kurtosis_weighted<decltype(workspace.weights_arr), ldouble_safe>(
|
2290
|
+
workspace.ix_arr.data(), workspace.st, workspace.end,
|
2291
|
+
input_data.categ_data + (col - input_data.ncols_numeric) * input_data.nrows,
|
2292
|
+
input_data.ncat[col - input_data.ncols_numeric],
|
2293
|
+
workspace.buffer_dbl.data(),
|
2294
|
+
model_params.missing_action, model_params.cat_split_type, workspace.rnd_generator,
|
2295
|
+
workspace.weights_arr);
|
2296
|
+
else
|
2297
|
+
kurt_weights[col] =
|
2298
|
+
calc_kurtosis_weighted<decltype(workspace.weights_map), ldouble_safe>(
|
2299
|
+
workspace.ix_arr.data(), workspace.st, workspace.end,
|
2300
|
+
input_data.categ_data + (col - input_data.ncols_numeric) * input_data.nrows,
|
2301
|
+
input_data.ncat[col - input_data.ncols_numeric],
|
2302
|
+
workspace.buffer_dbl.data(),
|
2303
|
+
model_params.missing_action, model_params.cat_split_type, workspace.rnd_generator,
|
2304
|
+
workspace.weights_map);
|
2305
|
+
}
|
2306
|
+
|
2307
|
+
/* Note to self: don't move this to outside of the braces, as it needs to assign a weight
|
2308
|
+
of zero to the columns that were not selected, thus it should only do this clipping
|
2309
|
+
for columns that are chosen. */
|
2310
|
+
if (kurt_weights[col] == -HUGE_VAL)
|
2311
|
+
{
|
2312
|
+
kurt_weights[col] = 0;
|
2313
|
+
}
|
2314
|
+
|
2315
|
+
else
|
2316
|
+
{
|
2317
|
+
kurt_weights[col] = std::fmax(1e-8, -1. + kurt_weights[col]);
|
2318
|
+
if (input_data.col_weights != NULL)
|
2319
|
+
{
|
2320
|
+
kurt_weights[col] *= input_data.col_weights[col];
|
2321
|
+
kurt_weights[col] = std::fmax(kurt_weights[col], 1e-100);
|
2322
|
+
}
|
2323
|
+
}
|
2324
|
+
}
|
2325
|
+
|
2326
|
+
skip_kurt_calculations:
|
2327
|
+
workspace.col_sampler.initialize(kurt_weights.data(), kurt_weights.size());
|
2328
|
+
avoid_leave_m_cols = true;
|
2329
|
+
}
|
2330
|
+
|
2331
|
+
if (model_params.prob_pick_col_by_range || model_params.prob_pick_col_by_var)
|
2332
|
+
{
|
2333
|
+
workspace.tree_kurtoses = kurt_weights.data();
|
2334
|
+
}
|
2335
|
+
}
|
2336
|
+
|
2337
|
+
bool col_sampler_is_fresh = true;
|
2338
|
+
if (input_data.preinitialized_col_sampler == NULL) {
|
2339
|
+
workspace.col_sampler.initialize(input_data.ncols_tot);
|
2340
|
+
}
|
2341
|
+
else {
|
2342
|
+
workspace.col_sampler = *((ColumnSampler<ldouble_safe>*)input_data.preinitialized_col_sampler);
|
2343
|
+
col_sampler_is_fresh = false;
|
2344
|
+
}
|
2345
|
+
/* TODO: this can be done more efficiently when sub-sampling columns */
|
2346
|
+
if (!avoid_leave_m_cols)
|
2347
|
+
workspace.col_sampler.leave_m_cols(model_params.ncols_per_tree, workspace.rnd_generator);
|
2348
|
+
if (model_params.ncols_per_tree < input_data.ncols_tot) col_sampler_is_fresh = false;
|
2349
|
+
workspace.try_all = false;
|
2350
|
+
if (hplane_root != NULL && model_params.ndim >= input_data.ncols_tot)
|
2351
|
+
workspace.try_all = true;
|
2352
|
+
|
2353
|
+
if (model_params.scoring_metric != Depth && !is_boxed_metric(model_params.scoring_metric))
|
2354
|
+
{
|
2355
|
+
workspace.density_calculator.initialize(model_params.max_depth,
|
2356
|
+
input_data.ncols_categ? input_data.max_categ : 0,
|
2357
|
+
tree_root != NULL && input_data.ncols_categ,
|
2358
|
+
model_params.scoring_metric);
|
2359
|
+
}
|
2360
|
+
|
2361
|
+
else if (is_boxed_metric(model_params.scoring_metric))
|
2362
|
+
{
|
2363
|
+
if (tree_root != NULL)
|
2364
|
+
workspace.density_calculator.initialize_bdens(input_data,
|
2365
|
+
model_params,
|
2366
|
+
workspace.ix_arr,
|
2367
|
+
workspace.col_sampler);
|
2368
|
+
else
|
2369
|
+
workspace.density_calculator.initialize_bdens_ext(input_data,
|
2370
|
+
model_params,
|
2371
|
+
workspace.ix_arr,
|
2372
|
+
workspace.col_sampler,
|
2373
|
+
col_sampler_is_fresh);
|
2374
|
+
}
|
2375
|
+
|
2376
|
+
if (tree_root != NULL)
|
2377
|
+
{
|
2378
|
+
split_itree_recursive<InputData, WorkerMemory, ldouble_safe>(
|
2379
|
+
*tree_root,
|
2380
|
+
workspace,
|
2381
|
+
input_data,
|
2382
|
+
model_params,
|
2383
|
+
impute_nodes,
|
2384
|
+
0);
|
2385
|
+
}
|
2386
|
+
|
2387
|
+
else
|
2388
|
+
{
|
2389
|
+
split_hplane_recursive<InputData, WorkerMemory, ldouble_safe>(
|
2390
|
+
*hplane_root,
|
2391
|
+
workspace,
|
2392
|
+
input_data,
|
2393
|
+
model_params,
|
2394
|
+
impute_nodes,
|
2395
|
+
0);
|
2396
|
+
}
|
2397
|
+
|
2398
|
+
/* if producing imputation structs, only need to keep the ones for terminal nodes */
|
2399
|
+
if (impute_nodes != NULL)
|
2400
|
+
drop_nonterminal_imp_node(*impute_nodes, tree_root, hplane_root);
|
2401
|
+
}
|