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 +1,849 @@
|
|
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
|
+
/* Predict outlier score, average depth, or terminal node numbers
|
48
|
+
*
|
49
|
+
* Parameters
|
50
|
+
* ==========
|
51
|
+
* - numeric_data[nrows * ncols_numeric]
|
52
|
+
* Pointer to numeric data for which to make predictions. Must be ordered by columns like Fortran,
|
53
|
+
* not ordered by rows like C (i.e. entries 1..n contain column 0, n+1..2n column 1, etc.),
|
54
|
+
* and the column order must be the same as in the data that was used to fit the model.
|
55
|
+
* Pass NULL if there are no dense numeric columns.
|
56
|
+
* Can only pass one of 'numeric_data', 'Xc' + 'Xc_ind' + 'Xc_indptr', 'Xr' + 'Xr_ind' + 'Xr_indptr'.
|
57
|
+
* - categ_data[nrows * ncols_categ]
|
58
|
+
* Pointer to categorical data for which to make predictions. Must be ordered by columns like Fortran,
|
59
|
+
* not ordered by rows like C (i.e. entries 1..n contain column 0, n+1..2n column 1, etc.),
|
60
|
+
* and the column order must be the same as in the data that was used to fit the model.
|
61
|
+
* Pass NULL if there are no categorical columns.
|
62
|
+
* Each category should be represented as an integer, and these integers must start at zero and
|
63
|
+
* be in consecutive order - i.e. if category '3' is present, category '2' must have also been
|
64
|
+
* present when the model was fit (note that they are not treated as being ordinal, this is just
|
65
|
+
* an encoding). Missing values should be encoded as negative numbers such as (-1). The encoding
|
66
|
+
* must be the same as was used in the data to which the model was fit.
|
67
|
+
* - Xc[nnz]
|
68
|
+
* Pointer to numeric data in sparse numeric matrix in CSC format (column-compressed).
|
69
|
+
* Pass NULL if there are no sparse numeric columns.
|
70
|
+
* Can only pass one of 'numeric_data', 'Xc' + 'Xc_ind' + 'Xc_indptr', 'Xr' + 'Xr_ind' + 'Xr_indptr'.
|
71
|
+
* - Xc_ind[nnz]
|
72
|
+
* Pointer to row indices to which each non-zero entry in 'Xc' corresponds.
|
73
|
+
* Pass NULL if there are no sparse numeric columns in CSC format.
|
74
|
+
* - Xc_indptr[ncols_categ + 1]
|
75
|
+
* Pointer to column index pointers that tell at entry [col] where does column 'col'
|
76
|
+
* start and at entry [col + 1] where does column 'col' end.
|
77
|
+
* Pass NULL if there are no sparse numeric columns in CSC format.
|
78
|
+
* - Xr[nnz]
|
79
|
+
* Pointer to numeric data in sparse numeric matrix in CSR format (row-compressed).
|
80
|
+
* Pass NULL if there are no sparse numeric columns.
|
81
|
+
* Can only pass one of 'numeric_data', 'Xc' + 'Xc_ind' + 'Xc_indptr', 'Xr' + 'Xr_ind' + 'Xr_indptr'.
|
82
|
+
* - Xr_ind[nnz]
|
83
|
+
* Pointer to column indices to which each non-zero entry in 'Xr' corresponds.
|
84
|
+
* Pass NULL if there are no sparse numeric columns in CSR format.
|
85
|
+
* - Xr_indptr[nrows + 1]
|
86
|
+
* Pointer to row index pointers that tell at entry [row] where does row 'row'
|
87
|
+
* start and at entry [row + 1] where does row 'row' end.
|
88
|
+
* Pass NULL if there are no sparse numeric columns in CSR format.
|
89
|
+
* - nrows
|
90
|
+
* Number of rows in 'numeric_data', 'Xc', 'Xr, 'categ_data'.
|
91
|
+
* - nthreads
|
92
|
+
* Number of parallel threads to use. Note that, the more threads, the more memory will be
|
93
|
+
* allocated, even if the thread does not end up being used. Ignored when not building with
|
94
|
+
* OpenMP support.
|
95
|
+
* - standardize
|
96
|
+
* Whether to standardize the average depths for each row according to their relative magnitude
|
97
|
+
* compared to the expected average, in order to obtain an outlier score. If passing 'false',
|
98
|
+
* will output the average depth instead.
|
99
|
+
* Ignored when not passing 'output_depths'.
|
100
|
+
* - model_outputs
|
101
|
+
* Pointer to fitted single-variable model object from function 'fit_iforest'. Pass NULL
|
102
|
+
* if the predictions are to be made from an extended model. Can only pass one of
|
103
|
+
* 'model_outputs' and 'model_outputs_ext'.
|
104
|
+
* - model_outputs_ext
|
105
|
+
* Pointer to fitted extended model object from function 'fit_iforest'. Pass NULL
|
106
|
+
* if the predictions are to be made from a single-variable model. Can only pass one of
|
107
|
+
* 'model_outputs' and 'model_outputs_ext'.
|
108
|
+
* - output_depths[nrows] (out)
|
109
|
+
* Pointer to array where the output average depths or outlier scores will be written into
|
110
|
+
* (the return type is control according to parameter 'standardize').
|
111
|
+
* Must already be initialized to zeros. Must also be passed and when the desired output
|
112
|
+
* is terminal node numbers.
|
113
|
+
* - tree_num[nrows * ntrees] (out)
|
114
|
+
* Pointer to array where the output terminal node numbers will be written into.
|
115
|
+
* Note that the mapping between tree node and terminal tree node is not stored in
|
116
|
+
* the model object for efficiency reasons, so this mapping will be determined on-the-fly
|
117
|
+
* when passing this parameter, and as such, there will be some overhead regardless of
|
118
|
+
* the actual number of rows. Pass NULL if only average depths or outlier scores are desired.
|
119
|
+
*/
|
120
|
+
void predict_iforest(double numeric_data[], int categ_data[],
|
121
|
+
double Xc[], sparse_ix Xc_ind[], sparse_ix Xc_indptr[],
|
122
|
+
double Xr[], sparse_ix Xr_ind[], sparse_ix Xr_indptr[],
|
123
|
+
size_t nrows, int nthreads, bool standardize,
|
124
|
+
IsoForest *model_outputs, ExtIsoForest *model_outputs_ext,
|
125
|
+
double output_depths[], sparse_ix tree_num[])
|
126
|
+
{
|
127
|
+
/* put data in a struct for passing it in fewer lines */
|
128
|
+
PredictionData prediction_data = {numeric_data, categ_data, nrows,
|
129
|
+
Xc, Xc_ind, Xc_indptr,
|
130
|
+
Xr, Xr_ind, Xr_indptr};
|
131
|
+
|
132
|
+
if ((size_t)nthreads > nrows)
|
133
|
+
nthreads = nrows;
|
134
|
+
|
135
|
+
if (model_outputs != NULL)
|
136
|
+
{
|
137
|
+
if (
|
138
|
+
model_outputs->missing_action == Fail &&
|
139
|
+
(model_outputs->new_cat_action != Weighted || prediction_data.categ_data == NULL) &&
|
140
|
+
prediction_data.Xc == NULL && prediction_data.Xr == NULL
|
141
|
+
)
|
142
|
+
{
|
143
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) shared(nrows, model_outputs, prediction_data, output_depths, tree_num)
|
144
|
+
for (size_t_for row = 0; row < nrows; row++)
|
145
|
+
{
|
146
|
+
for (std::vector<IsoTree> &tree : model_outputs->trees)
|
147
|
+
{
|
148
|
+
traverse_itree_no_recurse(tree,
|
149
|
+
*model_outputs,
|
150
|
+
prediction_data,
|
151
|
+
output_depths[row],
|
152
|
+
(tree_num == NULL)? NULL : tree_num + nrows * (&tree - &(model_outputs->trees[0])),
|
153
|
+
(size_t) row);
|
154
|
+
}
|
155
|
+
}
|
156
|
+
}
|
157
|
+
|
158
|
+
else
|
159
|
+
{
|
160
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) shared(nrows, model_outputs, prediction_data, output_depths, tree_num)
|
161
|
+
for (size_t_for row = 0; row < nrows; row++)
|
162
|
+
{
|
163
|
+
for (std::vector<IsoTree> &tree : model_outputs->trees)
|
164
|
+
{
|
165
|
+
output_depths[row] += traverse_itree(tree,
|
166
|
+
*model_outputs,
|
167
|
+
prediction_data,
|
168
|
+
NULL, NULL, 0,
|
169
|
+
(size_t) row,
|
170
|
+
(tree_num == NULL)? NULL : tree_num + nrows * (&tree - &(model_outputs->trees[0])),
|
171
|
+
(size_t) 0);
|
172
|
+
}
|
173
|
+
}
|
174
|
+
}
|
175
|
+
}
|
176
|
+
|
177
|
+
|
178
|
+
else
|
179
|
+
{
|
180
|
+
if (
|
181
|
+
model_outputs_ext->missing_action == Fail &&
|
182
|
+
prediction_data.categ_data == NULL &&
|
183
|
+
prediction_data.Xc == NULL &&
|
184
|
+
prediction_data.Xr == NULL
|
185
|
+
)
|
186
|
+
{
|
187
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) shared(nrows, model_outputs_ext, prediction_data, output_depths, tree_num)
|
188
|
+
for (size_t_for row = 0; row < nrows; row++)
|
189
|
+
{
|
190
|
+
for (std::vector<IsoHPlane> &hplane : model_outputs_ext->hplanes)
|
191
|
+
{
|
192
|
+
traverse_hplane_fast(hplane,
|
193
|
+
*model_outputs_ext,
|
194
|
+
prediction_data,
|
195
|
+
output_depths[row],
|
196
|
+
(tree_num == NULL)? NULL : tree_num + nrows * (&hplane - &(model_outputs_ext->hplanes[0])),
|
197
|
+
(size_t) row);
|
198
|
+
}
|
199
|
+
}
|
200
|
+
}
|
201
|
+
|
202
|
+
else
|
203
|
+
{
|
204
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) shared(nrows, model_outputs_ext, prediction_data, output_depths, tree_num)
|
205
|
+
for (size_t_for row = 0; row < nrows; row++)
|
206
|
+
{
|
207
|
+
for (std::vector<IsoHPlane> &hplane : model_outputs_ext->hplanes)
|
208
|
+
{
|
209
|
+
traverse_hplane(hplane,
|
210
|
+
*model_outputs_ext,
|
211
|
+
prediction_data,
|
212
|
+
output_depths[row],
|
213
|
+
NULL, NULL,
|
214
|
+
(tree_num == NULL)? NULL : tree_num + nrows * (&hplane - &(model_outputs_ext->hplanes[0])),
|
215
|
+
(size_t) row);
|
216
|
+
}
|
217
|
+
}
|
218
|
+
}
|
219
|
+
}
|
220
|
+
|
221
|
+
/* translate sum-of-depths to outlier score */
|
222
|
+
double ntrees, depth_divisor;
|
223
|
+
if (model_outputs != NULL)
|
224
|
+
{
|
225
|
+
ntrees = (double) model_outputs->trees.size();
|
226
|
+
depth_divisor = ntrees * (model_outputs->exp_avg_depth);
|
227
|
+
}
|
228
|
+
|
229
|
+
else
|
230
|
+
{
|
231
|
+
ntrees = (double) model_outputs_ext->hplanes.size();
|
232
|
+
depth_divisor = ntrees * (model_outputs_ext->exp_avg_depth);
|
233
|
+
}
|
234
|
+
|
235
|
+
if (standardize)
|
236
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) shared(nrows, output_depths, depth_divisor)
|
237
|
+
for (size_t_for row = 0; row < nrows; row++)
|
238
|
+
output_depths[row] = exp2( - output_depths[row] / depth_divisor );
|
239
|
+
else
|
240
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) shared(nrows, output_depths, ntrees)
|
241
|
+
for (size_t_for row = 0; row < nrows; row++)
|
242
|
+
output_depths[row] /= ntrees;
|
243
|
+
|
244
|
+
|
245
|
+
/* re-map tree numbers to start at zero (if predicting tree numbers) */
|
246
|
+
/* Note: usually this type of 'prediction' is not required,
|
247
|
+
thus this mapping is not stored in the model objects so as to
|
248
|
+
save memory */
|
249
|
+
if (tree_num != NULL)
|
250
|
+
remap_terminal_trees(model_outputs, model_outputs_ext,
|
251
|
+
prediction_data, tree_num, nthreads);
|
252
|
+
}
|
253
|
+
|
254
|
+
|
255
|
+
void traverse_itree_no_recurse(std::vector<IsoTree> &tree,
|
256
|
+
IsoForest &model_outputs,
|
257
|
+
PredictionData &prediction_data,
|
258
|
+
double &output_depth,
|
259
|
+
sparse_ix *restrict tree_num,
|
260
|
+
size_t row)
|
261
|
+
{
|
262
|
+
size_t curr_lev = 0;
|
263
|
+
double xval;
|
264
|
+
while (true)
|
265
|
+
{
|
266
|
+
if (tree[curr_lev].score > 0)
|
267
|
+
{
|
268
|
+
output_depth += tree[curr_lev].score;
|
269
|
+
if (tree_num != NULL)
|
270
|
+
tree_num[row] = curr_lev;
|
271
|
+
break;
|
272
|
+
}
|
273
|
+
|
274
|
+
else
|
275
|
+
{
|
276
|
+
switch(tree[curr_lev].col_type)
|
277
|
+
{
|
278
|
+
case Numeric:
|
279
|
+
{
|
280
|
+
xval = prediction_data.numeric_data[row + tree[curr_lev].col_num * prediction_data.nrows];
|
281
|
+
curr_lev = (xval <= tree[curr_lev].num_split)?
|
282
|
+
tree[curr_lev].tree_left : tree[curr_lev].tree_right;
|
283
|
+
output_depth += (xval < tree[curr_lev].range_low) || (xval > tree[curr_lev].range_high);
|
284
|
+
break;
|
285
|
+
}
|
286
|
+
|
287
|
+
case Categorical:
|
288
|
+
{
|
289
|
+
switch(model_outputs.cat_split_type)
|
290
|
+
{
|
291
|
+
case SubSet:
|
292
|
+
{
|
293
|
+
|
294
|
+
if (!tree[curr_lev].cat_split.size()) /* this is for binary columns */
|
295
|
+
{
|
296
|
+
if (prediction_data.categ_data[row + tree[curr_lev].col_num * prediction_data.nrows] <= 1)
|
297
|
+
{
|
298
|
+
curr_lev = (
|
299
|
+
prediction_data.categ_data[row + tree[curr_lev].col_num * prediction_data.nrows]
|
300
|
+
== 0
|
301
|
+
)?
|
302
|
+
tree[curr_lev].tree_left : tree[curr_lev].tree_right;
|
303
|
+
}
|
304
|
+
|
305
|
+
else /* can only work with 'Smallest' + no NAs if reaching this point */
|
306
|
+
{
|
307
|
+
curr_lev = (tree[curr_lev].pct_tree_left < .5)? tree[curr_lev].tree_left : tree[curr_lev].tree_right;
|
308
|
+
}
|
309
|
+
}
|
310
|
+
|
311
|
+
else
|
312
|
+
{
|
313
|
+
|
314
|
+
switch(model_outputs.new_cat_action)
|
315
|
+
{
|
316
|
+
case Random:
|
317
|
+
{
|
318
|
+
curr_lev = (tree[curr_lev].cat_split[
|
319
|
+
prediction_data.categ_data[row + tree[curr_lev].col_num * prediction_data.nrows]
|
320
|
+
]
|
321
|
+
)?
|
322
|
+
tree[curr_lev].tree_left : tree[curr_lev].tree_right;
|
323
|
+
break;
|
324
|
+
}
|
325
|
+
|
326
|
+
case Smallest:
|
327
|
+
{
|
328
|
+
if (
|
329
|
+
prediction_data.categ_data[row + tree[curr_lev].col_num * prediction_data.nrows]
|
330
|
+
>= (int)tree[curr_lev].cat_split.size()
|
331
|
+
)
|
332
|
+
{
|
333
|
+
curr_lev = (tree[curr_lev].pct_tree_left < .5)? tree[curr_lev].tree_left : tree[curr_lev].tree_right;
|
334
|
+
}
|
335
|
+
|
336
|
+
else
|
337
|
+
{
|
338
|
+
curr_lev = (tree[curr_lev].cat_split[
|
339
|
+
prediction_data.categ_data[row + tree[curr_lev].col_num * prediction_data.nrows]
|
340
|
+
]
|
341
|
+
)?
|
342
|
+
tree[curr_lev].tree_left : tree[curr_lev].tree_right;
|
343
|
+
}
|
344
|
+
break;
|
345
|
+
}
|
346
|
+
}
|
347
|
+
}
|
348
|
+
break;
|
349
|
+
}
|
350
|
+
|
351
|
+
case SingleCateg:
|
352
|
+
{
|
353
|
+
curr_lev = (
|
354
|
+
prediction_data.categ_data[row + tree[curr_lev].col_num * prediction_data.nrows]
|
355
|
+
==
|
356
|
+
tree[curr_lev].chosen_cat
|
357
|
+
)?
|
358
|
+
tree[curr_lev].tree_left : tree[curr_lev].tree_right;
|
359
|
+
break;
|
360
|
+
}
|
361
|
+
}
|
362
|
+
break;
|
363
|
+
}
|
364
|
+
}
|
365
|
+
}
|
366
|
+
}
|
367
|
+
}
|
368
|
+
|
369
|
+
|
370
|
+
double traverse_itree(std::vector<IsoTree> &tree,
|
371
|
+
IsoForest &model_outputs,
|
372
|
+
PredictionData &prediction_data,
|
373
|
+
std::vector<ImputeNode> *impute_nodes, /* only when imputing missing */
|
374
|
+
ImputedData *imputed_data, /* only when imputing missing */
|
375
|
+
double curr_weight, /* only when imputing missing */
|
376
|
+
size_t row,
|
377
|
+
sparse_ix *restrict tree_num,
|
378
|
+
size_t curr_lev)
|
379
|
+
{
|
380
|
+
double xval;
|
381
|
+
double range_penalty = 0;
|
382
|
+
|
383
|
+
sparse_ix *row_st = NULL, *row_end = NULL;
|
384
|
+
if (prediction_data.Xr != NULL)
|
385
|
+
{
|
386
|
+
row_st = prediction_data.Xr_ind + prediction_data.Xr_indptr[row];
|
387
|
+
row_end = prediction_data.Xr_ind + prediction_data.Xr_indptr[row + 1];
|
388
|
+
}
|
389
|
+
|
390
|
+
while (true)
|
391
|
+
{
|
392
|
+
if (tree[curr_lev].score >= 0.)
|
393
|
+
{
|
394
|
+
if (tree_num != NULL)
|
395
|
+
tree_num[row] = curr_lev;
|
396
|
+
if (imputed_data != NULL)
|
397
|
+
add_from_impute_node((*impute_nodes)[curr_lev], *imputed_data, curr_weight);
|
398
|
+
|
399
|
+
return tree[curr_lev].score + range_penalty;
|
400
|
+
}
|
401
|
+
|
402
|
+
else
|
403
|
+
{
|
404
|
+
switch(tree[curr_lev].col_type)
|
405
|
+
{
|
406
|
+
case Numeric:
|
407
|
+
{
|
408
|
+
|
409
|
+
if (prediction_data.Xc == NULL && prediction_data.Xr == NULL)
|
410
|
+
xval = prediction_data.numeric_data[row + tree[curr_lev].col_num * prediction_data.nrows];
|
411
|
+
else if (row_st != NULL)
|
412
|
+
xval = extract_spR(prediction_data, row_st, row_end, tree[curr_lev].col_num);
|
413
|
+
else
|
414
|
+
xval = extract_spC(prediction_data, row, tree[curr_lev].col_num);
|
415
|
+
|
416
|
+
if (isnan(xval))
|
417
|
+
{
|
418
|
+
switch(model_outputs.missing_action)
|
419
|
+
{
|
420
|
+
case Divide:
|
421
|
+
{
|
422
|
+
return
|
423
|
+
tree[curr_lev].pct_tree_left
|
424
|
+
* traverse_itree(tree, model_outputs, prediction_data,
|
425
|
+
impute_nodes, imputed_data, curr_weight * tree[curr_lev].pct_tree_left,
|
426
|
+
row, NULL, tree[curr_lev].tree_left)
|
427
|
+
+ (1 - tree[curr_lev].pct_tree_left)
|
428
|
+
* traverse_itree(tree, model_outputs, prediction_data,
|
429
|
+
impute_nodes, imputed_data, curr_weight * (1 - tree[curr_lev].pct_tree_left),
|
430
|
+
row, NULL, tree[curr_lev].tree_right)
|
431
|
+
+ range_penalty;
|
432
|
+
}
|
433
|
+
|
434
|
+
case Impute:
|
435
|
+
{
|
436
|
+
curr_lev = (tree[curr_lev].pct_tree_left >= .5)?
|
437
|
+
tree[curr_lev].tree_left : tree[curr_lev].tree_right;
|
438
|
+
break;
|
439
|
+
}
|
440
|
+
|
441
|
+
case Fail:
|
442
|
+
{
|
443
|
+
return NAN;
|
444
|
+
}
|
445
|
+
}
|
446
|
+
}
|
447
|
+
|
448
|
+
else
|
449
|
+
{
|
450
|
+
curr_lev = (xval <=tree[curr_lev].num_split)?
|
451
|
+
tree[curr_lev].tree_left : tree[curr_lev].tree_right;
|
452
|
+
range_penalty += (xval < tree[curr_lev].range_low) || (xval > tree[curr_lev].range_high);
|
453
|
+
}
|
454
|
+
break;
|
455
|
+
}
|
456
|
+
|
457
|
+
case Categorical:
|
458
|
+
{
|
459
|
+
|
460
|
+
if (prediction_data.categ_data[row + tree[curr_lev].col_num * prediction_data.nrows] < 0)
|
461
|
+
{
|
462
|
+
switch(model_outputs.missing_action)
|
463
|
+
{
|
464
|
+
case Divide:
|
465
|
+
{
|
466
|
+
return
|
467
|
+
tree[curr_lev].pct_tree_left
|
468
|
+
* traverse_itree(tree, model_outputs, prediction_data,
|
469
|
+
impute_nodes, imputed_data, curr_weight * tree[curr_lev].pct_tree_left,
|
470
|
+
row, NULL, tree[curr_lev].tree_left)
|
471
|
+
+ (1 - tree[curr_lev].pct_tree_left)
|
472
|
+
* traverse_itree(tree, model_outputs, prediction_data,
|
473
|
+
impute_nodes, imputed_data, curr_weight * (1 - tree[curr_lev].pct_tree_left),
|
474
|
+
row, NULL, tree[curr_lev].tree_right)
|
475
|
+
+ range_penalty;
|
476
|
+
}
|
477
|
+
|
478
|
+
case Impute:
|
479
|
+
{
|
480
|
+
curr_lev = (tree[curr_lev].pct_tree_left >= .5)?
|
481
|
+
tree[curr_lev].tree_left : tree[curr_lev].tree_right;
|
482
|
+
break;
|
483
|
+
}
|
484
|
+
|
485
|
+
case Fail:
|
486
|
+
{
|
487
|
+
return NAN;
|
488
|
+
}
|
489
|
+
}
|
490
|
+
}
|
491
|
+
|
492
|
+
else
|
493
|
+
{
|
494
|
+
switch(model_outputs.cat_split_type)
|
495
|
+
{
|
496
|
+
case SingleCateg:
|
497
|
+
{
|
498
|
+
curr_lev = (
|
499
|
+
prediction_data.categ_data[row + tree[curr_lev].col_num * prediction_data.nrows]
|
500
|
+
==
|
501
|
+
tree[curr_lev].chosen_cat
|
502
|
+
)?
|
503
|
+
tree[curr_lev].tree_left : tree[curr_lev].tree_right;
|
504
|
+
break;
|
505
|
+
}
|
506
|
+
|
507
|
+
case SubSet:
|
508
|
+
{
|
509
|
+
|
510
|
+
if (!tree[curr_lev].cat_split.size())
|
511
|
+
{
|
512
|
+
if (prediction_data.categ_data[row + tree[curr_lev].col_num * prediction_data.nrows] <= 1)
|
513
|
+
{
|
514
|
+
curr_lev = (
|
515
|
+
prediction_data.categ_data[row + tree[curr_lev].col_num * prediction_data.nrows]
|
516
|
+
== 0
|
517
|
+
)?
|
518
|
+
tree[curr_lev].tree_left : tree[curr_lev].tree_right;
|
519
|
+
}
|
520
|
+
|
521
|
+
else
|
522
|
+
{
|
523
|
+
switch(model_outputs.new_cat_action)
|
524
|
+
{
|
525
|
+
case Smallest:
|
526
|
+
{
|
527
|
+
curr_lev = (tree[curr_lev].pct_tree_left < .5)? tree[curr_lev].tree_left : tree[curr_lev].tree_right;
|
528
|
+
break;
|
529
|
+
}
|
530
|
+
|
531
|
+
case Weighted:
|
532
|
+
{
|
533
|
+
return
|
534
|
+
tree[curr_lev].pct_tree_left
|
535
|
+
* traverse_itree(tree, model_outputs, prediction_data,
|
536
|
+
impute_nodes, imputed_data, curr_weight * tree[curr_lev].pct_tree_left,
|
537
|
+
row, NULL, tree[curr_lev].tree_left)
|
538
|
+
+ (1 - tree[curr_lev].pct_tree_left)
|
539
|
+
* traverse_itree(tree, model_outputs, prediction_data,
|
540
|
+
impute_nodes, imputed_data, curr_weight * (1 - tree[curr_lev].pct_tree_left),
|
541
|
+
row, NULL, tree[curr_lev].tree_right)
|
542
|
+
+ range_penalty;
|
543
|
+
}
|
544
|
+
}
|
545
|
+
}
|
546
|
+
}
|
547
|
+
|
548
|
+
else
|
549
|
+
{
|
550
|
+
switch(model_outputs.new_cat_action)
|
551
|
+
{
|
552
|
+
case Random:
|
553
|
+
{
|
554
|
+
curr_lev = (tree[curr_lev].cat_split[
|
555
|
+
prediction_data.categ_data[row + tree[curr_lev].col_num * prediction_data.nrows]
|
556
|
+
]
|
557
|
+
)?
|
558
|
+
tree[curr_lev].tree_left : tree[curr_lev].tree_right;
|
559
|
+
break;
|
560
|
+
}
|
561
|
+
|
562
|
+
case Smallest:
|
563
|
+
{
|
564
|
+
if (
|
565
|
+
prediction_data.categ_data[row + tree[curr_lev].col_num * prediction_data.nrows]
|
566
|
+
>= (int)tree[curr_lev].cat_split.size()
|
567
|
+
)
|
568
|
+
{
|
569
|
+
curr_lev = (tree[curr_lev].pct_tree_left < .5)? tree[curr_lev].tree_left : tree[curr_lev].tree_right;
|
570
|
+
}
|
571
|
+
|
572
|
+
else
|
573
|
+
{
|
574
|
+
curr_lev = (tree[curr_lev].cat_split[
|
575
|
+
prediction_data.categ_data[row + tree[curr_lev].col_num * prediction_data.nrows]
|
576
|
+
]
|
577
|
+
)?
|
578
|
+
tree[curr_lev].tree_left : tree[curr_lev].tree_right;
|
579
|
+
}
|
580
|
+
break;
|
581
|
+
}
|
582
|
+
|
583
|
+
case Weighted:
|
584
|
+
{
|
585
|
+
if (
|
586
|
+
prediction_data.categ_data[row + tree[curr_lev].col_num * prediction_data.nrows]
|
587
|
+
>= (int)tree[curr_lev].cat_split.size()
|
588
|
+
||
|
589
|
+
tree[curr_lev].cat_split[
|
590
|
+
prediction_data.categ_data[row + tree[curr_lev].col_num * prediction_data.nrows]
|
591
|
+
]
|
592
|
+
== (-1)
|
593
|
+
)
|
594
|
+
{
|
595
|
+
return
|
596
|
+
tree[curr_lev].pct_tree_left
|
597
|
+
* traverse_itree(tree, model_outputs, prediction_data,
|
598
|
+
impute_nodes, imputed_data, curr_weight * tree[curr_lev].pct_tree_left,
|
599
|
+
row, NULL, tree[curr_lev].tree_left)
|
600
|
+
+ (1 - tree[curr_lev].pct_tree_left)
|
601
|
+
* traverse_itree(tree, model_outputs, prediction_data,
|
602
|
+
impute_nodes, imputed_data, curr_weight * (1 - tree[curr_lev].pct_tree_left),
|
603
|
+
row, NULL, tree[curr_lev].tree_right)
|
604
|
+
+ range_penalty;
|
605
|
+
}
|
606
|
+
|
607
|
+
else
|
608
|
+
{
|
609
|
+
curr_lev = (tree[curr_lev].cat_split[
|
610
|
+
prediction_data.categ_data[row + tree[curr_lev].col_num * prediction_data.nrows]
|
611
|
+
]
|
612
|
+
)?
|
613
|
+
tree[curr_lev].tree_left : tree[curr_lev].tree_right;
|
614
|
+
}
|
615
|
+
break;
|
616
|
+
}
|
617
|
+
}
|
618
|
+
}
|
619
|
+
break;
|
620
|
+
}
|
621
|
+
}
|
622
|
+
}
|
623
|
+
break;
|
624
|
+
}
|
625
|
+
}
|
626
|
+
}
|
627
|
+
}
|
628
|
+
}
|
629
|
+
|
630
|
+
/* this is a simpler version for situations in which there is
|
631
|
+
only numeric data in dense arrays and no missing values */
|
632
|
+
void traverse_hplane_fast(std::vector<IsoHPlane> &hplane,
|
633
|
+
ExtIsoForest &model_outputs,
|
634
|
+
PredictionData &prediction_data,
|
635
|
+
double &output_depth,
|
636
|
+
sparse_ix *restrict tree_num,
|
637
|
+
size_t row)
|
638
|
+
{
|
639
|
+
size_t curr_lev = 0;
|
640
|
+
double hval;
|
641
|
+
|
642
|
+
while(true)
|
643
|
+
{
|
644
|
+
if (hplane[curr_lev].score > 0)
|
645
|
+
{
|
646
|
+
output_depth += hplane[curr_lev].score;
|
647
|
+
if (tree_num != NULL)
|
648
|
+
tree_num[row] = curr_lev;
|
649
|
+
return;
|
650
|
+
}
|
651
|
+
|
652
|
+
else
|
653
|
+
{
|
654
|
+
hval = 0;
|
655
|
+
for (size_t col = 0; col < hplane[curr_lev].col_num.size(); col++)
|
656
|
+
hval += (prediction_data.numeric_data[row + hplane[curr_lev].col_num[col] * prediction_data.nrows]
|
657
|
+
- hplane[curr_lev].mean[col]) * hplane[curr_lev].coef[col];
|
658
|
+
}
|
659
|
+
|
660
|
+
output_depth += (hval < hplane[curr_lev].range_low) ||
|
661
|
+
(hval > hplane[curr_lev].range_high);
|
662
|
+
curr_lev = (hval <= hplane[curr_lev].split_point)?
|
663
|
+
hplane[curr_lev].hplane_left : hplane[curr_lev].hplane_right;
|
664
|
+
}
|
665
|
+
}
|
666
|
+
|
667
|
+
/* this is the full version that works with potentially missing values, sparse matrices, and categoricals */
|
668
|
+
void traverse_hplane(std::vector<IsoHPlane> &hplane,
|
669
|
+
ExtIsoForest &model_outputs,
|
670
|
+
PredictionData &prediction_data,
|
671
|
+
double &output_depth,
|
672
|
+
std::vector<ImputeNode> *impute_nodes, /* only when imputing missing */
|
673
|
+
ImputedData *imputed_data, /* only when imputing missing */
|
674
|
+
sparse_ix *restrict tree_num,
|
675
|
+
size_t row)
|
676
|
+
{
|
677
|
+
size_t curr_lev = 0;
|
678
|
+
double xval;
|
679
|
+
int cval;
|
680
|
+
double hval;
|
681
|
+
|
682
|
+
size_t ncols_numeric, ncols_categ;
|
683
|
+
|
684
|
+
sparse_ix *row_st = NULL, *row_end = NULL;
|
685
|
+
if (prediction_data.Xr != NULL)
|
686
|
+
{
|
687
|
+
row_st = prediction_data.Xr_ind + prediction_data.Xr_indptr[row];
|
688
|
+
row_end = prediction_data.Xr_ind + prediction_data.Xr_indptr[row + 1];
|
689
|
+
}
|
690
|
+
|
691
|
+
while(true)
|
692
|
+
{
|
693
|
+
if (hplane[curr_lev].score > 0)
|
694
|
+
{
|
695
|
+
output_depth += hplane[curr_lev].score;
|
696
|
+
if (tree_num != NULL)
|
697
|
+
tree_num[row] = curr_lev;
|
698
|
+
if (imputed_data != NULL)
|
699
|
+
{
|
700
|
+
add_from_impute_node((*impute_nodes)[curr_lev], *imputed_data, (double)1);
|
701
|
+
}
|
702
|
+
return;
|
703
|
+
}
|
704
|
+
|
705
|
+
else
|
706
|
+
{
|
707
|
+
hval = 0;
|
708
|
+
ncols_numeric = 0; ncols_categ = 0;
|
709
|
+
for (size_t col = 0; col < hplane[curr_lev].col_num.size(); col++)
|
710
|
+
{
|
711
|
+
switch(hplane[curr_lev].col_type[col])
|
712
|
+
{
|
713
|
+
case Numeric:
|
714
|
+
{
|
715
|
+
if (prediction_data.Xc == NULL && prediction_data.Xr == NULL)
|
716
|
+
xval = prediction_data.numeric_data[row + hplane[curr_lev].col_num[col] * prediction_data.nrows];
|
717
|
+
else if (row_st != NULL)
|
718
|
+
xval = extract_spR(prediction_data, row_st, row_end, hplane[curr_lev].col_num[col]);
|
719
|
+
else
|
720
|
+
xval = extract_spC(prediction_data, row, hplane[curr_lev].col_num[col]);
|
721
|
+
|
722
|
+
if (is_na_or_inf(xval))
|
723
|
+
{
|
724
|
+
if (model_outputs.missing_action != Fail)
|
725
|
+
{
|
726
|
+
hval += hplane[curr_lev].fill_val[col];
|
727
|
+
}
|
728
|
+
|
729
|
+
else
|
730
|
+
{
|
731
|
+
output_depth = NAN;
|
732
|
+
return;
|
733
|
+
}
|
734
|
+
}
|
735
|
+
|
736
|
+
else
|
737
|
+
{
|
738
|
+
hval += (xval - hplane[curr_lev].mean[ncols_numeric]) * hplane[curr_lev].coef[ncols_numeric];
|
739
|
+
}
|
740
|
+
|
741
|
+
ncols_numeric++;
|
742
|
+
break;
|
743
|
+
}
|
744
|
+
|
745
|
+
case Categorical:
|
746
|
+
{
|
747
|
+
cval = prediction_data.categ_data[row + hplane[curr_lev].col_num[col] * prediction_data.nrows];
|
748
|
+
if (cval < 0)
|
749
|
+
{
|
750
|
+
if (model_outputs.missing_action != Fail)
|
751
|
+
{
|
752
|
+
hval += hplane[curr_lev].fill_val[col];
|
753
|
+
}
|
754
|
+
|
755
|
+
else
|
756
|
+
{
|
757
|
+
output_depth = NAN;
|
758
|
+
return;
|
759
|
+
}
|
760
|
+
}
|
761
|
+
|
762
|
+
else
|
763
|
+
{
|
764
|
+
switch(model_outputs.cat_split_type)
|
765
|
+
{
|
766
|
+
case SingleCateg:
|
767
|
+
{
|
768
|
+
hval += (cval == hplane[curr_lev].chosen_cat[ncols_categ])? hplane[curr_lev].fill_new[ncols_categ] : 0;
|
769
|
+
break;
|
770
|
+
}
|
771
|
+
|
772
|
+
case SubSet:
|
773
|
+
{
|
774
|
+
if (cval >= (int)hplane[curr_lev].cat_coef[ncols_categ].size())
|
775
|
+
hval += hplane[curr_lev].fill_new[ncols_categ];
|
776
|
+
else
|
777
|
+
hval += hplane[curr_lev].cat_coef[ncols_categ][cval];
|
778
|
+
break;
|
779
|
+
}
|
780
|
+
}
|
781
|
+
}
|
782
|
+
|
783
|
+
ncols_categ++;
|
784
|
+
break;
|
785
|
+
}
|
786
|
+
}
|
787
|
+
|
788
|
+
}
|
789
|
+
|
790
|
+
output_depth += (hval < hplane[curr_lev].range_low) ||
|
791
|
+
(hval > hplane[curr_lev].range_high);
|
792
|
+
curr_lev = (hval <= hplane[curr_lev].split_point)?
|
793
|
+
hplane[curr_lev].hplane_left : hplane[curr_lev].hplane_right;
|
794
|
+
}
|
795
|
+
}
|
796
|
+
}
|
797
|
+
|
798
|
+
double extract_spC(PredictionData &prediction_data, size_t row, size_t col_num)
|
799
|
+
{
|
800
|
+
sparse_ix *search_res = std::lower_bound(prediction_data.Xc_ind + prediction_data.Xc_indptr[col_num],
|
801
|
+
prediction_data.Xc_ind + prediction_data.Xc_indptr[col_num + 1],
|
802
|
+
(sparse_ix) row);
|
803
|
+
if (
|
804
|
+
search_res == (prediction_data.Xc_ind + prediction_data.Xc_indptr[col_num + 1])
|
805
|
+
||
|
806
|
+
*search_res != row
|
807
|
+
)
|
808
|
+
return 0;
|
809
|
+
else
|
810
|
+
return prediction_data.Xc[search_res - prediction_data.Xc_ind];
|
811
|
+
}
|
812
|
+
|
813
|
+
double extract_spR(PredictionData &prediction_data, sparse_ix *row_st, sparse_ix *row_end, size_t col_num)
|
814
|
+
{
|
815
|
+
sparse_ix *search_res = std::lower_bound(row_st, row_end, (sparse_ix) col_num);
|
816
|
+
if (search_res == row_end || *search_res != (sparse_ix)col_num)
|
817
|
+
return 0;
|
818
|
+
else
|
819
|
+
return prediction_data.Xr[search_res - prediction_data.Xr_ind];
|
820
|
+
}
|
821
|
+
|
822
|
+
void get_num_nodes(IsoForest &model_outputs, sparse_ix *restrict n_nodes, sparse_ix *restrict n_terminal, int nthreads)
|
823
|
+
{
|
824
|
+
std::fill(n_terminal, n_terminal + model_outputs.trees.size(), 0);
|
825
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) shared(model_outputs, n_nodes, n_terminal)
|
826
|
+
for (size_t_for tree = 0; tree < model_outputs.trees.size(); tree++)
|
827
|
+
{
|
828
|
+
n_nodes[tree] = model_outputs.trees[tree].size();
|
829
|
+
for (IsoTree &node : model_outputs.trees[tree])
|
830
|
+
{
|
831
|
+
n_terminal[tree] += (node.score > 0);
|
832
|
+
}
|
833
|
+
}
|
834
|
+
}
|
835
|
+
|
836
|
+
void get_num_nodes(ExtIsoForest &model_outputs, sparse_ix *restrict n_nodes, sparse_ix *restrict n_terminal, int nthreads)
|
837
|
+
{
|
838
|
+
std::fill(n_terminal, n_terminal + model_outputs.hplanes.size(), 0);
|
839
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) shared(model_outputs, n_nodes, n_terminal)
|
840
|
+
for (size_t_for hplane = 0; hplane < model_outputs.hplanes.size(); hplane++)
|
841
|
+
{
|
842
|
+
n_nodes[hplane] = model_outputs.hplanes[hplane].size();
|
843
|
+
for (IsoHPlane &node : model_outputs.hplanes[hplane])
|
844
|
+
{
|
845
|
+
n_terminal[hplane] += (node.score > 0);
|
846
|
+
}
|
847
|
+
}
|
848
|
+
}
|
849
|
+
|