isotree 0.1.0
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +7 -0
- data/CHANGELOG.md +3 -0
- data/LICENSE.txt +25 -0
- data/README.md +111 -0
- data/ext/isotree/ext.cpp +178 -0
- data/ext/isotree/extconf.rb +21 -0
- data/lib/isotree.rb +9 -0
- data/lib/isotree/isolation_forest.rb +94 -0
- data/lib/isotree/version.rb +3 -0
- data/vendor/isotree/LICENSE +25 -0
- data/vendor/isotree/README.md +167 -0
- data/vendor/isotree/src/Makevars +4 -0
- data/vendor/isotree/src/RcppExports.cpp +267 -0
- data/vendor/isotree/src/Rwrapper.cpp +762 -0
- data/vendor/isotree/src/crit.cpp +912 -0
- data/vendor/isotree/src/dealloc.cpp +66 -0
- data/vendor/isotree/src/dist.cpp +749 -0
- data/vendor/isotree/src/extended.cpp +790 -0
- data/vendor/isotree/src/fit_model.cpp +1068 -0
- data/vendor/isotree/src/helpers_iforest.cpp +309 -0
- data/vendor/isotree/src/impute.cpp +1205 -0
- data/vendor/isotree/src/isoforest.cpp +771 -0
- data/vendor/isotree/src/isotree.hpp +929 -0
- data/vendor/isotree/src/merge_models.cpp +116 -0
- data/vendor/isotree/src/mult.cpp +607 -0
- data/vendor/isotree/src/predict.cpp +849 -0
- data/vendor/isotree/src/serialize.cpp +262 -0
- data/vendor/isotree/src/utils.cpp +1574 -0
- metadata +154 -0
@@ -0,0 +1,116 @@
|
|
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. "Distance approximation using Isolation Forests." arXiv preprint arXiv:1910.12362 (2019).
|
22
|
+
* [9] Cortes, David. "Imputing missing values with unsupervised random trees." arXiv preprint arXiv:1911.06646 (2019).
|
23
|
+
*
|
24
|
+
* BSD 2-Clause License
|
25
|
+
* Copyright (c) 2019, David Cortes
|
26
|
+
* All rights reserved.
|
27
|
+
* Redistribution and use in source and binary forms, with or without
|
28
|
+
* modification, are permitted provided that the following conditions are met:
|
29
|
+
* * Redistributions of source code must retain the above copyright notice, this
|
30
|
+
* list of conditions and the following disclaimer.
|
31
|
+
* * Redistributions in binary form must reproduce the above copyright notice,
|
32
|
+
* this list of conditions and the following disclaimer in the documentation
|
33
|
+
* and/or other materials provided with the distribution.
|
34
|
+
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
35
|
+
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
36
|
+
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
37
|
+
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
38
|
+
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
39
|
+
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
40
|
+
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
41
|
+
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
42
|
+
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
43
|
+
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
44
|
+
*/
|
45
|
+
#include "isotree.hpp"
|
46
|
+
|
47
|
+
/* Append trees from one model into another
|
48
|
+
*
|
49
|
+
* Parameters
|
50
|
+
* ==========
|
51
|
+
* - model (in, out)
|
52
|
+
* Pointer to isolation forest model wich has already been fit through 'fit_iforest'.
|
53
|
+
* The trees from 'other' will be merged into this (will be at the end of vector member 'trees').
|
54
|
+
* Both 'model' and 'other' must have been fit with the same hyperparameters
|
55
|
+
* in order for this merge to work correctly - at the very least, should have
|
56
|
+
* the same 'missing_action', 'cat_split_type', 'new_cat_action'.
|
57
|
+
* Should only pass one of 'model'+'other' or 'ext_model'+'ext_other'.
|
58
|
+
* Pass NULL if this is not to be used.
|
59
|
+
* - other
|
60
|
+
* Pointer to isolation forest model which has already been fit through 'fit_iforest'.
|
61
|
+
* The trees from this object will be added into 'model' (this object will not be modified).
|
62
|
+
* Both 'model' and 'other' must have been fit with the same hyperparameters
|
63
|
+
* in order for this merge to work correctly - at the very least, should have
|
64
|
+
* the same 'missing_action', 'cat_split_type', 'new_cat_action'.
|
65
|
+
* Should only pass one of 'model'+'other' or 'ext_model'+'ext_other'.
|
66
|
+
* Pass NULL if this is not to be used.
|
67
|
+
* - ext_model (in, out)
|
68
|
+
* Pointer to extended isolation forest model which has already been fit through 'fit_iforest'.
|
69
|
+
* The trees/hyperplanes from 'ext_other' will be merged into this (will be at the end of vector member 'hplanes').
|
70
|
+
* Both 'ext_model' and 'ext_other' must have been fit with the same hyperparameters
|
71
|
+
* in order for this merge to work correctly - at the very least, should have
|
72
|
+
* the same 'missing_action', 'cat_split_type', 'new_cat_action'.
|
73
|
+
* Should only pass one of 'model'+'other' or 'ext_model'+'ext_other'.
|
74
|
+
* Pass NULL if this is not to be used.
|
75
|
+
* - ext_other
|
76
|
+
* Pointer to extended isolation forest model which has already been fit through 'fit_iforest'.
|
77
|
+
* The trees/hyperplanes from this object will be added into 'ext_model' (this object will not be modified).
|
78
|
+
* Both 'ext_model' and 'ext_other' must have been fit with the same hyperparameters
|
79
|
+
* in order for this merge to work correctly - at the very least, should have
|
80
|
+
* the same 'missing_action', 'cat_split_type', 'new_cat_action'.
|
81
|
+
* Should only pass one of 'model'+'other' or 'ext_model'+'ext_other'.
|
82
|
+
* Pass NULL if this is not to be used.
|
83
|
+
* - imputer (in, out)
|
84
|
+
* Pointer to imputation object which has already been fit through 'fit_iforest' along with
|
85
|
+
* either 'model' or 'ext_model' in the same call to 'fit_iforest'.
|
86
|
+
* The imputation nodes from 'iother' will be merged into this (will be at the end of vector member 'imputer_tree').
|
87
|
+
* Hyperparameters related to imputation might differ between 'imputer' and 'iother' ('imputer' will preserve its
|
88
|
+
* hyperparameters after the merge).
|
89
|
+
* Pass NULL if this is not to be used.
|
90
|
+
* - iother
|
91
|
+
* Pointer to imputation object which has already been fit through 'fit_iforest' along with
|
92
|
+
* either 'model' or 'ext_model' in the same call to 'fit_iforest'.
|
93
|
+
* The imputation nodes from this object will be added into 'imputer' (this object will not be modified).
|
94
|
+
* Hyperparameters related to imputation might differ between 'imputer' and 'iother' ('imputer' will preserve its
|
95
|
+
* hyperparameters after the merge).
|
96
|
+
* Pass NULL if this is not to be used.
|
97
|
+
*/
|
98
|
+
void merge_models(IsoForest* model, IsoForest* other,
|
99
|
+
ExtIsoForest* ext_model, ExtIsoForest* ext_other,
|
100
|
+
Imputer* imputer, Imputer* iother)
|
101
|
+
{
|
102
|
+
if (model != NULL && other != NULL)
|
103
|
+
model->trees.insert(model->trees.end(),
|
104
|
+
other->trees.begin(),
|
105
|
+
other->trees.end());
|
106
|
+
|
107
|
+
if (ext_model != NULL && ext_other != NULL)
|
108
|
+
ext_model->hplanes.insert(ext_model->hplanes.end(),
|
109
|
+
ext_other->hplanes.begin(),
|
110
|
+
ext_other->hplanes.end());
|
111
|
+
|
112
|
+
if (imputer != NULL && iother != NULL)
|
113
|
+
imputer->imputer_tree.insert(imputer->imputer_tree.end(),
|
114
|
+
iother->imputer_tree.begin(),
|
115
|
+
iother->imputer_tree.end());
|
116
|
+
}
|
@@ -0,0 +1,607 @@
|
|
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. "Distance approximation using Isolation Forests." arXiv preprint arXiv:1910.12362 (2019).
|
22
|
+
* [9] Cortes, David. "Imputing missing values with unsupervised random trees." arXiv preprint arXiv:1911.06646 (2019).
|
23
|
+
*
|
24
|
+
* BSD 2-Clause License
|
25
|
+
* Copyright (c) 2019, David Cortes
|
26
|
+
* All rights reserved.
|
27
|
+
* Redistribution and use in source and binary forms, with or without
|
28
|
+
* modification, are permitted provided that the following conditions are met:
|
29
|
+
* * Redistributions of source code must retain the above copyright notice, this
|
30
|
+
* list of conditions and the following disclaimer.
|
31
|
+
* * Redistributions in binary form must reproduce the above copyright notice,
|
32
|
+
* this list of conditions and the following disclaimer in the documentation
|
33
|
+
* and/or other materials provided with the distribution.
|
34
|
+
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
35
|
+
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
36
|
+
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
37
|
+
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
38
|
+
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
39
|
+
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
40
|
+
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
41
|
+
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
42
|
+
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
43
|
+
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
44
|
+
*/
|
45
|
+
#include "isotree.hpp"
|
46
|
+
|
47
|
+
/* for regular numerical */
|
48
|
+
void calc_mean_and_sd(size_t ix_arr[], size_t st, size_t end, double *restrict x,
|
49
|
+
MissingAction missing_action, double &x_sd, double &x_mean)
|
50
|
+
{
|
51
|
+
long double m = 0;
|
52
|
+
long double s = 0;
|
53
|
+
long double m_prev = 0;
|
54
|
+
|
55
|
+
if (missing_action == Fail)
|
56
|
+
{
|
57
|
+
for (size_t row = st; row <= end; row++)
|
58
|
+
{
|
59
|
+
m += (x[ix_arr[row]] - m) / (long double)(row - st + 1);
|
60
|
+
s += (x[ix_arr[row]] - m) * (x[ix_arr[row]] - m_prev);
|
61
|
+
m_prev = m;
|
62
|
+
}
|
63
|
+
|
64
|
+
x_mean = m;
|
65
|
+
x_sd = sqrtl(s / (long double)(end - st + 1));
|
66
|
+
}
|
67
|
+
|
68
|
+
else
|
69
|
+
{
|
70
|
+
size_t cnt = 0;
|
71
|
+
for (size_t row = st; row <= end; row++)
|
72
|
+
{
|
73
|
+
if (!is_na_or_inf(x[ix_arr[row]]))
|
74
|
+
{
|
75
|
+
cnt++;
|
76
|
+
m += (x[ix_arr[row]] - m) / (long double)cnt;
|
77
|
+
s += (x[ix_arr[row]] - m) * (x[ix_arr[row]] - m_prev);
|
78
|
+
m_prev = m;
|
79
|
+
}
|
80
|
+
}
|
81
|
+
|
82
|
+
x_mean = m;
|
83
|
+
x_sd = sqrtl(s / (long double)cnt);
|
84
|
+
}
|
85
|
+
}
|
86
|
+
|
87
|
+
/* for sparse numerical */
|
88
|
+
void calc_mean_and_sd(size_t ix_arr[], size_t st, size_t end, size_t col_num,
|
89
|
+
double *restrict Xc, sparse_ix Xc_ind[], sparse_ix Xc_indptr[],
|
90
|
+
double &x_sd, double &x_mean)
|
91
|
+
{
|
92
|
+
/* ix_arr must be already sorted beforehand */
|
93
|
+
if (Xc_indptr[col_num] == Xc_indptr[col_num + 1])
|
94
|
+
{
|
95
|
+
x_sd = 0;
|
96
|
+
x_mean = 0;
|
97
|
+
return;
|
98
|
+
}
|
99
|
+
size_t st_col = Xc_indptr[col_num];
|
100
|
+
size_t end_col = Xc_indptr[col_num + 1] - 1;
|
101
|
+
size_t curr_pos = st_col;
|
102
|
+
size_t ind_end_col = (size_t) Xc_ind[end_col];
|
103
|
+
size_t *ptr_st = std::lower_bound(ix_arr + st, ix_arr + end + 1, (size_t)Xc_ind[st_col]);
|
104
|
+
|
105
|
+
size_t cnt = end - st + 1;
|
106
|
+
long double sum = 0;
|
107
|
+
long double sum_sq = 0;
|
108
|
+
|
109
|
+
/* Note: this function will discard NAs regardless of chosen action. If reaching the point of calling
|
110
|
+
this function, chances are that the performance gain of not checking for them will not be important */
|
111
|
+
|
112
|
+
for (size_t *row = ptr_st;
|
113
|
+
row != ix_arr + end + 1 && curr_pos != end_col + 1 && ind_end_col >= *row;
|
114
|
+
)
|
115
|
+
{
|
116
|
+
if (Xc_ind[curr_pos] == *row)
|
117
|
+
{
|
118
|
+
if (is_na_or_inf(Xc[curr_pos]))
|
119
|
+
{
|
120
|
+
cnt--;
|
121
|
+
}
|
122
|
+
|
123
|
+
else
|
124
|
+
{
|
125
|
+
sum += Xc[curr_pos];
|
126
|
+
sum_sq += square(Xc[curr_pos]);
|
127
|
+
}
|
128
|
+
|
129
|
+
if (row == ix_arr + end || curr_pos == end_col) break;
|
130
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
131
|
+
}
|
132
|
+
|
133
|
+
else
|
134
|
+
{
|
135
|
+
if (Xc_ind[curr_pos] > *row)
|
136
|
+
row = std::lower_bound(row + 1, ix_arr + end + 1, Xc_ind[curr_pos]);
|
137
|
+
else
|
138
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *row) - Xc_ind;
|
139
|
+
}
|
140
|
+
}
|
141
|
+
|
142
|
+
x_mean = sum / (long double) cnt;
|
143
|
+
x_sd = calc_sd_raw(cnt, sum, sum_sq);
|
144
|
+
}
|
145
|
+
|
146
|
+
|
147
|
+
/* Note about these functions: they write into an array that does not need to match to 'ix_arr',
|
148
|
+
and instead, the index that is stored in ix_arr[n] will have the value in res[n] */
|
149
|
+
|
150
|
+
/* for regular numerical */
|
151
|
+
void add_linear_comb(size_t ix_arr[], size_t st, size_t end, double *restrict res,
|
152
|
+
double *restrict x, double &coef, double x_sd, double x_mean, double &fill_val,
|
153
|
+
MissingAction missing_action, double *restrict buffer_arr,
|
154
|
+
size_t *restrict buffer_NAs, bool first_run)
|
155
|
+
{
|
156
|
+
/* TODO: here don't need the buffer for NAs */
|
157
|
+
|
158
|
+
if (first_run)
|
159
|
+
coef /= x_sd;
|
160
|
+
|
161
|
+
size_t cnt = 0;
|
162
|
+
size_t cnt_NA = 0;
|
163
|
+
double *restrict res_write = res - st;
|
164
|
+
|
165
|
+
if (missing_action == Fail)
|
166
|
+
{
|
167
|
+
for (size_t row = st; row <= end; row++)
|
168
|
+
res_write[row] += (x[ix_arr[row]] - x_mean) * coef;
|
169
|
+
}
|
170
|
+
|
171
|
+
else
|
172
|
+
{
|
173
|
+
if (first_run)
|
174
|
+
{
|
175
|
+
for (size_t row = st; row <= end; row++)
|
176
|
+
{
|
177
|
+
if (!is_na_or_inf(x[ix_arr[row]]))
|
178
|
+
{
|
179
|
+
res_write[row] += (x[ix_arr[row]] - x_mean) * coef;
|
180
|
+
buffer_arr[cnt++] = x[ix_arr[row]];
|
181
|
+
}
|
182
|
+
|
183
|
+
else
|
184
|
+
{
|
185
|
+
buffer_NAs[cnt_NA++] = row;
|
186
|
+
}
|
187
|
+
|
188
|
+
}
|
189
|
+
}
|
190
|
+
|
191
|
+
else
|
192
|
+
{
|
193
|
+
for (size_t row = st; row <= end; row++)
|
194
|
+
{
|
195
|
+
res_write[row] += (is_na_or_inf(x[ix_arr[row]]))? fill_val : ( (x[ix_arr[row]]-x_mean) * coef );
|
196
|
+
}
|
197
|
+
return;
|
198
|
+
}
|
199
|
+
|
200
|
+
size_t mid_ceil = cnt / 2;
|
201
|
+
std::partial_sort(buffer_arr, buffer_arr + mid_ceil + 1, buffer_arr + cnt);
|
202
|
+
|
203
|
+
if ((cnt % 2) == 0)
|
204
|
+
fill_val = (buffer_arr[mid_ceil - 1] + buffer_arr[mid_ceil]) / 2.0;
|
205
|
+
else
|
206
|
+
fill_val = buffer_arr[mid_ceil];
|
207
|
+
|
208
|
+
fill_val = (fill_val - x_mean) * coef;
|
209
|
+
if (cnt_NA)
|
210
|
+
{
|
211
|
+
for (size_t row = 0; row < cnt_NA; row++)
|
212
|
+
res_write[buffer_NAs[row]] += fill_val;
|
213
|
+
}
|
214
|
+
|
215
|
+
}
|
216
|
+
}
|
217
|
+
|
218
|
+
/* for sparse numerical */
|
219
|
+
void add_linear_comb(size_t *restrict ix_arr, size_t st, size_t end, size_t col_num, double *restrict res,
|
220
|
+
double *restrict Xc, sparse_ix *restrict Xc_ind, sparse_ix *restrict Xc_indptr,
|
221
|
+
double &coef, double x_sd, double x_mean, double &fill_val, MissingAction missing_action,
|
222
|
+
double *restrict buffer_arr, size_t *restrict buffer_NAs, bool first_run)
|
223
|
+
{
|
224
|
+
/* ix_arr must be already sorted beforehand */
|
225
|
+
|
226
|
+
/* if it's all zeros, no need to do anything, but this is not supposed
|
227
|
+
to happen while fitting because the range is determined before calling this */
|
228
|
+
if (
|
229
|
+
Xc_indptr[col_num] == Xc_indptr[col_num + 1] ||
|
230
|
+
Xc_ind[Xc_indptr[col_num]] > ix_arr[end] ||
|
231
|
+
Xc_ind[Xc_indptr[col_num + 1] - 1] < ix_arr[st]
|
232
|
+
)
|
233
|
+
{
|
234
|
+
if (first_run)
|
235
|
+
{
|
236
|
+
coef /= x_sd;
|
237
|
+
if (missing_action != Fail)
|
238
|
+
fill_val = 0;
|
239
|
+
}
|
240
|
+
|
241
|
+
double *restrict res_write = res - st;
|
242
|
+
double offset = x_mean * coef;
|
243
|
+
for (size_t row = st; row <= end; row++)
|
244
|
+
res_write[row] -= offset;
|
245
|
+
|
246
|
+
return;
|
247
|
+
}
|
248
|
+
|
249
|
+
size_t st_col = Xc_indptr[col_num];
|
250
|
+
size_t end_col = Xc_indptr[col_num + 1] - 1;
|
251
|
+
size_t curr_pos = st_col;
|
252
|
+
size_t *ptr_st = std::lower_bound(ix_arr + st, ix_arr + end + 1, (size_t)Xc_ind[st_col]);
|
253
|
+
|
254
|
+
size_t cnt_non_NA = 0; /* when NAs need to be imputed */
|
255
|
+
size_t cnt_NA = 0; /* when NAs need to be imputed */
|
256
|
+
size_t n_sample = end - st + 1;
|
257
|
+
size_t *ix_arr_plus_st = ix_arr + st;
|
258
|
+
|
259
|
+
if (first_run)
|
260
|
+
coef /= x_sd;
|
261
|
+
|
262
|
+
double *restrict res_write = res - st;
|
263
|
+
double offset = x_mean * coef;
|
264
|
+
for (size_t row = st; row <= end; row++)
|
265
|
+
res_write[row] -= offset;
|
266
|
+
|
267
|
+
size_t ind_end_col = Xc_ind[end_col];
|
268
|
+
size_t nmatches = 0;
|
269
|
+
|
270
|
+
if (missing_action != Fail)
|
271
|
+
{
|
272
|
+
if (first_run)
|
273
|
+
{
|
274
|
+
for (size_t *row = ptr_st;
|
275
|
+
row != ix_arr + end + 1 && curr_pos != end_col + 1 && ind_end_col >= *row;
|
276
|
+
)
|
277
|
+
{
|
278
|
+
if (Xc_ind[curr_pos] == *row)
|
279
|
+
{
|
280
|
+
if (is_na_or_inf(Xc[curr_pos]))
|
281
|
+
{
|
282
|
+
buffer_NAs[cnt_NA++] = row - ix_arr_plus_st;
|
283
|
+
}
|
284
|
+
|
285
|
+
else
|
286
|
+
{
|
287
|
+
buffer_arr[cnt_non_NA++] = Xc[curr_pos];
|
288
|
+
res[row - ix_arr_plus_st] += Xc[curr_pos] * coef;
|
289
|
+
}
|
290
|
+
|
291
|
+
nmatches++;
|
292
|
+
if (row == ix_arr + end || curr_pos == end_col) break;
|
293
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
294
|
+
}
|
295
|
+
|
296
|
+
else
|
297
|
+
{
|
298
|
+
if (Xc_ind[curr_pos] > *row)
|
299
|
+
row = std::lower_bound(row + 1, ix_arr + end + 1, Xc_ind[curr_pos]);
|
300
|
+
else
|
301
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *row) - Xc_ind;
|
302
|
+
}
|
303
|
+
}
|
304
|
+
}
|
305
|
+
|
306
|
+
else
|
307
|
+
{
|
308
|
+
/* when impute value for missing has already been determined */
|
309
|
+
for (size_t *row = ptr_st;
|
310
|
+
row != ix_arr + end + 1 && curr_pos != end_col + 1 && ind_end_col >= *row;
|
311
|
+
)
|
312
|
+
{
|
313
|
+
if (Xc_ind[curr_pos] == *row)
|
314
|
+
{
|
315
|
+
res[row - ix_arr_plus_st] += is_na_or_inf(Xc[curr_pos])?
|
316
|
+
(fill_val + offset) : (Xc[curr_pos] * coef);
|
317
|
+
if (row == ix_arr + end) break;
|
318
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
319
|
+
}
|
320
|
+
|
321
|
+
else
|
322
|
+
{
|
323
|
+
if (Xc_ind[curr_pos] > *row)
|
324
|
+
row = std::lower_bound(row + 1, ix_arr + end + 1, Xc_ind[curr_pos]);
|
325
|
+
else
|
326
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *row) - Xc_ind;
|
327
|
+
}
|
328
|
+
}
|
329
|
+
|
330
|
+
return;
|
331
|
+
}
|
332
|
+
|
333
|
+
|
334
|
+
/* Determine imputation value */
|
335
|
+
std::sort(buffer_arr, buffer_arr + cnt_non_NA);
|
336
|
+
size_t mid_ceil = (n_sample - cnt_NA) / 2;
|
337
|
+
size_t nzeros = (end - st + 1) - nmatches;
|
338
|
+
if (nzeros > mid_ceil && buffer_arr[0] > 0)
|
339
|
+
{
|
340
|
+
fill_val = 0;
|
341
|
+
return;
|
342
|
+
}
|
343
|
+
|
344
|
+
else
|
345
|
+
{
|
346
|
+
size_t n_neg = (buffer_arr[0] > 0)?
|
347
|
+
0 : ((buffer_arr[cnt_non_NA - 1] < 0)?
|
348
|
+
cnt_non_NA : std::lower_bound(buffer_arr, buffer_arr + cnt_non_NA, (double)0) - buffer_arr);
|
349
|
+
|
350
|
+
|
351
|
+
if (n_neg < (mid_ceil-1) && n_neg + nzeros > mid_ceil)
|
352
|
+
{
|
353
|
+
fill_val = 0;
|
354
|
+
return;
|
355
|
+
}
|
356
|
+
|
357
|
+
else
|
358
|
+
{
|
359
|
+
/* if the sample size is odd, take the middle, otherwise take a simple average */
|
360
|
+
if (((n_sample - cnt_NA) % 2) != 0)
|
361
|
+
{
|
362
|
+
if (mid_ceil < n_neg)
|
363
|
+
fill_val = buffer_arr[mid_ceil];
|
364
|
+
else if (mid_ceil < n_neg + nzeros)
|
365
|
+
fill_val = 0;
|
366
|
+
else
|
367
|
+
fill_val = buffer_arr[mid_ceil - nzeros];
|
368
|
+
}
|
369
|
+
|
370
|
+
else
|
371
|
+
{
|
372
|
+
if (mid_ceil < n_neg)
|
373
|
+
{
|
374
|
+
fill_val = (buffer_arr[mid_ceil - 1] + buffer_arr[mid_ceil]) / 2;
|
375
|
+
}
|
376
|
+
|
377
|
+
else if (mid_ceil < n_neg + nzeros)
|
378
|
+
{
|
379
|
+
if (mid_ceil == n_neg)
|
380
|
+
fill_val = buffer_arr[mid_ceil - 1] / 2;
|
381
|
+
else
|
382
|
+
fill_val = 0;
|
383
|
+
}
|
384
|
+
|
385
|
+
else
|
386
|
+
{
|
387
|
+
if (mid_ceil == n_neg + nzeros && nzeros > 0)
|
388
|
+
fill_val = buffer_arr[n_neg] / 2;
|
389
|
+
else
|
390
|
+
fill_val = (buffer_arr[mid_ceil - nzeros - 1] + buffer_arr[mid_ceil - nzeros]) / 2; /* WRONG!!!! */
|
391
|
+
}
|
392
|
+
}
|
393
|
+
|
394
|
+
/* fill missing if any */
|
395
|
+
fill_val *= coef;
|
396
|
+
if (cnt_NA && fill_val)
|
397
|
+
for (size_t ix = 0; ix < cnt_NA; ix++)
|
398
|
+
res[buffer_NAs[ix]] += fill_val;
|
399
|
+
|
400
|
+
/* next time, it will need to have the offset added */
|
401
|
+
fill_val -= offset;
|
402
|
+
}
|
403
|
+
}
|
404
|
+
}
|
405
|
+
|
406
|
+
else /* no NAs */
|
407
|
+
{
|
408
|
+
for (size_t *row = ptr_st;
|
409
|
+
row != ix_arr + end + 1 && curr_pos != end_col + 1 && ind_end_col >= *row;
|
410
|
+
)
|
411
|
+
{
|
412
|
+
if (Xc_ind[curr_pos] == *row)
|
413
|
+
{
|
414
|
+
res[row - ix_arr_plus_st] += Xc[curr_pos] * coef;
|
415
|
+
if (row == ix_arr + end || curr_pos == end_col) break;
|
416
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *(++row)) - Xc_ind;
|
417
|
+
}
|
418
|
+
|
419
|
+
else
|
420
|
+
{
|
421
|
+
if (Xc_ind[curr_pos] > *row)
|
422
|
+
row = std::lower_bound(row + 1, ix_arr + end + 1, Xc_ind[curr_pos]);
|
423
|
+
else
|
424
|
+
curr_pos = std::lower_bound(Xc_ind + curr_pos + 1, Xc_ind + end_col + 1, *row) - Xc_ind;
|
425
|
+
}
|
426
|
+
}
|
427
|
+
}
|
428
|
+
}
|
429
|
+
|
430
|
+
/* for categoricals */
|
431
|
+
void add_linear_comb(size_t *restrict ix_arr, size_t st, size_t end, double *restrict res,
|
432
|
+
int x[], int ncat, double *restrict cat_coef, double single_cat_coef, int chosen_cat,
|
433
|
+
double &fill_val, double &fill_new, size_t *restrict buffer_cnt, size_t *restrict buffer_pos,
|
434
|
+
NewCategAction new_cat_action, MissingAction missing_action, CategSplit cat_split_type, bool first_run)
|
435
|
+
{
|
436
|
+
double *restrict res_write = res - st;
|
437
|
+
switch(cat_split_type)
|
438
|
+
{
|
439
|
+
case SingleCateg:
|
440
|
+
{
|
441
|
+
/* in this case there's no need to make-up an impute value for new categories, only for NAs */
|
442
|
+
switch(missing_action)
|
443
|
+
{
|
444
|
+
case Fail:
|
445
|
+
{
|
446
|
+
for (size_t row = st; row <= end; row++)
|
447
|
+
res_write[row] += (x[ix_arr[row]] == chosen_cat)? single_cat_coef : 0;
|
448
|
+
return;
|
449
|
+
}
|
450
|
+
|
451
|
+
case Impute:
|
452
|
+
{
|
453
|
+
size_t cnt_NA = 0;
|
454
|
+
size_t cnt_this = 0;
|
455
|
+
size_t cnt = end - st + 1;
|
456
|
+
if (first_run)
|
457
|
+
{
|
458
|
+
for (size_t row = st; row <= end; row++)
|
459
|
+
{
|
460
|
+
if (x[ix_arr[row]] < 0)
|
461
|
+
{
|
462
|
+
cnt_NA++;
|
463
|
+
}
|
464
|
+
|
465
|
+
else if (x[ix_arr[row]] == chosen_cat)
|
466
|
+
{
|
467
|
+
cnt_this++;
|
468
|
+
res_write[row] += single_cat_coef;
|
469
|
+
}
|
470
|
+
}
|
471
|
+
}
|
472
|
+
|
473
|
+
else
|
474
|
+
{
|
475
|
+
for (size_t row = st; row <= end; row++)
|
476
|
+
res_write[row] += (x[ix_arr[row]] < 0)? fill_val : ((x[ix_arr[row]] == chosen_cat)? single_cat_coef : 0);
|
477
|
+
return;
|
478
|
+
}
|
479
|
+
|
480
|
+
fill_val = (cnt_this > (cnt - cnt_NA - cnt_this))? single_cat_coef : 0;
|
481
|
+
if (cnt_NA)
|
482
|
+
{
|
483
|
+
for (size_t row = st; row <= end; row++)
|
484
|
+
if (x[ix_arr[row]] < 0)
|
485
|
+
res_write[row] += fill_val;
|
486
|
+
}
|
487
|
+
return;
|
488
|
+
}
|
489
|
+
}
|
490
|
+
}
|
491
|
+
|
492
|
+
case SubSet:
|
493
|
+
{
|
494
|
+
/* in this case, since the splits are by more than 1 variable, it's not possible to
|
495
|
+
divide missing/new categoricals by assigning weights, so they have to be imputed
|
496
|
+
in both cases, unless using random weights for the new ones, in which case they won't
|
497
|
+
need to be imputed for new, but sill need it for NA */
|
498
|
+
|
499
|
+
if (new_cat_action == Random && missing_action == Fail)
|
500
|
+
{
|
501
|
+
for (size_t row = st; row <= end; row++)
|
502
|
+
res_write[row] += cat_coef[x[ix_arr[row]]];
|
503
|
+
return;
|
504
|
+
}
|
505
|
+
|
506
|
+
if (!first_run)
|
507
|
+
{
|
508
|
+
if (missing_action == Fail)
|
509
|
+
{
|
510
|
+
for (size_t row = st; row <= end; row++)
|
511
|
+
res_write[row] += (x[ix_arr[row]] >= ncat)? fill_new : cat_coef[x[ix_arr[row]]];
|
512
|
+
}
|
513
|
+
|
514
|
+
else
|
515
|
+
{
|
516
|
+
for (size_t row = st; row <= end; row++)
|
517
|
+
res_write[row] += (x[ix_arr[row]] < 0)? fill_val : ((x[ix_arr[row]] >= ncat)? fill_new : cat_coef[x[ix_arr[row]]]);
|
518
|
+
}
|
519
|
+
return;
|
520
|
+
}
|
521
|
+
|
522
|
+
std::fill(buffer_cnt, buffer_cnt + ncat + 1, 0);
|
523
|
+
switch(missing_action)
|
524
|
+
{
|
525
|
+
case Fail:
|
526
|
+
{
|
527
|
+
for (size_t row = st; row <= end; row++)
|
528
|
+
{
|
529
|
+
buffer_cnt[x[ix_arr[row]]]++;
|
530
|
+
res_write[row] += cat_coef[x[ix_arr[row]]];
|
531
|
+
}
|
532
|
+
break;
|
533
|
+
}
|
534
|
+
|
535
|
+
default:
|
536
|
+
{
|
537
|
+
for (size_t row = st; row <= end; row++)
|
538
|
+
{
|
539
|
+
if (x[ix_arr[row]] >= 0)
|
540
|
+
{
|
541
|
+
buffer_cnt[x[ix_arr[row]]]++;
|
542
|
+
res_write[row] += cat_coef[x[ix_arr[row]]];
|
543
|
+
}
|
544
|
+
|
545
|
+
else
|
546
|
+
{
|
547
|
+
buffer_cnt[ncat]++;
|
548
|
+
}
|
549
|
+
|
550
|
+
}
|
551
|
+
break;
|
552
|
+
}
|
553
|
+
}
|
554
|
+
|
555
|
+
switch(new_cat_action)
|
556
|
+
{
|
557
|
+
case Smallest:
|
558
|
+
{
|
559
|
+
size_t smallest = SIZE_MAX;
|
560
|
+
int cat_smallest;
|
561
|
+
for (int cat = 0; cat < ncat; cat++)
|
562
|
+
{
|
563
|
+
if (buffer_cnt[cat] > 0 && buffer_cnt[cat] < smallest)
|
564
|
+
{
|
565
|
+
smallest = buffer_cnt[cat];
|
566
|
+
cat_smallest = cat;
|
567
|
+
}
|
568
|
+
}
|
569
|
+
fill_new = cat_coef[cat_smallest];
|
570
|
+
if (missing_action == Fail) break;
|
571
|
+
}
|
572
|
+
|
573
|
+
default:
|
574
|
+
{
|
575
|
+
/* Determine imputation value as the category in sorted order that gives 50% + 1 */
|
576
|
+
long double cnt_l = (long double)((end - st + 1) - buffer_cnt[ncat]);
|
577
|
+
std::iota(buffer_pos, buffer_pos + ncat, (size_t)0);
|
578
|
+
std::sort(buffer_pos, buffer_pos + ncat, [&cat_coef](const size_t a, const size_t b){return cat_coef[a] < cat_coef[b];});
|
579
|
+
|
580
|
+
double cumprob = 0;
|
581
|
+
int cat;
|
582
|
+
for (cat = 0; cat < ncat; cat++)
|
583
|
+
{
|
584
|
+
cumprob += (long double)buffer_cnt[buffer_pos[cat]] / cnt_l;
|
585
|
+
if (cumprob >= .5) break;
|
586
|
+
}
|
587
|
+
// cat = std::min(cat, ncat); /* in case it picks the last one */
|
588
|
+
fill_val = cat_coef[buffer_pos[cat]];
|
589
|
+
if (new_cat_action != Smallest)
|
590
|
+
fill_new = fill_val;
|
591
|
+
|
592
|
+
if (buffer_cnt[ncat] > 0) /* NAs */
|
593
|
+
for (size_t row = st; row <= end; row++)
|
594
|
+
if (x[ix_arr[row]] < 0)
|
595
|
+
res_write[row] += fill_val;
|
596
|
+
}
|
597
|
+
}
|
598
|
+
|
599
|
+
/* now fill unseen categories */
|
600
|
+
if (new_cat_action != Random)
|
601
|
+
for (int cat = 0; cat < ncat; cat++)
|
602
|
+
if (!buffer_cnt[cat])
|
603
|
+
cat_coef[cat] = fill_new;
|
604
|
+
|
605
|
+
}
|
606
|
+
}
|
607
|
+
}
|