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 +1,309 @@
|
|
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
|
+
void decide_column(size_t ncols_numeric, size_t ncols_categ, size_t &col_chosen, ColType &col_type,
|
48
|
+
RNG_engine &rnd_generator, std::uniform_int_distribution<size_t> &runif,
|
49
|
+
std::discrete_distribution<size_t> &col_sampler)
|
50
|
+
{
|
51
|
+
if (!col_sampler.max())
|
52
|
+
col_chosen = runif(rnd_generator);
|
53
|
+
else
|
54
|
+
col_chosen = col_sampler(rnd_generator);
|
55
|
+
|
56
|
+
if (col_chosen >= ncols_numeric)
|
57
|
+
{
|
58
|
+
col_chosen -= ncols_numeric;
|
59
|
+
col_type = Categorical;
|
60
|
+
}
|
61
|
+
|
62
|
+
else { col_type = Numeric; }
|
63
|
+
}
|
64
|
+
|
65
|
+
void add_unsplittable_col(WorkerMemory &workspace, IsoTree &tree, InputData &input_data)
|
66
|
+
{
|
67
|
+
if (tree.col_type == Numeric)
|
68
|
+
workspace.cols_possible[tree.col_num] = false;
|
69
|
+
else
|
70
|
+
workspace.cols_possible[tree.col_num + input_data.ncols_numeric] = false;
|
71
|
+
}
|
72
|
+
|
73
|
+
void add_unsplittable_col(WorkerMemory &workspace, InputData &input_data)
|
74
|
+
{
|
75
|
+
if (workspace.col_type == Numeric)
|
76
|
+
workspace.cols_possible[workspace.col_chosen] = false;
|
77
|
+
else
|
78
|
+
workspace.cols_possible[workspace.col_chosen + input_data.ncols_numeric] = false;
|
79
|
+
}
|
80
|
+
|
81
|
+
bool check_is_not_unsplittable_col(WorkerMemory &workspace, IsoTree &tree, InputData &input_data)
|
82
|
+
{
|
83
|
+
if (tree.col_type == Numeric)
|
84
|
+
return workspace.cols_possible[tree.col_num];
|
85
|
+
else
|
86
|
+
return workspace.cols_possible[tree.col_num + input_data.ncols_numeric];
|
87
|
+
}
|
88
|
+
|
89
|
+
/* for use in regular model */
|
90
|
+
void get_split_range(WorkerMemory &workspace, InputData &input_data, ModelParams &model_params, IsoTree &tree)
|
91
|
+
{
|
92
|
+
if (tree.col_type == Numeric)
|
93
|
+
{
|
94
|
+
if (input_data.Xc == NULL)
|
95
|
+
get_range(workspace.ix_arr.data(), input_data.numeric_data + input_data.nrows * tree.col_num,
|
96
|
+
workspace.st, workspace.end, model_params.missing_action,
|
97
|
+
workspace.xmin, workspace.xmax, workspace.unsplittable);
|
98
|
+
else
|
99
|
+
get_range(workspace.ix_arr.data(), workspace.st, workspace.end, tree.col_num,
|
100
|
+
input_data.Xc, input_data.Xc_ind, input_data.Xc_indptr,
|
101
|
+
model_params.missing_action, workspace.xmin, workspace.xmax, workspace.unsplittable);
|
102
|
+
}
|
103
|
+
|
104
|
+
else
|
105
|
+
{
|
106
|
+
get_categs(workspace.ix_arr.data(), input_data.categ_data + input_data.nrows * tree.col_num,
|
107
|
+
workspace.st, workspace.end, input_data.ncat[tree.col_num],
|
108
|
+
model_params.missing_action, workspace.categs.data(), workspace.npresent, workspace.unsplittable);
|
109
|
+
}
|
110
|
+
}
|
111
|
+
|
112
|
+
/* for use in extended model */
|
113
|
+
void get_split_range(WorkerMemory &workspace, InputData &input_data, ModelParams &model_params)
|
114
|
+
{
|
115
|
+
if (workspace.col_type == Numeric)
|
116
|
+
{
|
117
|
+
if (input_data.Xc == NULL)
|
118
|
+
get_range(workspace.ix_arr.data(), input_data.numeric_data + input_data.nrows * workspace.col_chosen,
|
119
|
+
workspace.st, workspace.end, model_params.missing_action,
|
120
|
+
workspace.xmin, workspace.xmax, workspace.unsplittable);
|
121
|
+
else
|
122
|
+
get_range(workspace.ix_arr.data(), workspace.st, workspace.end, workspace.col_chosen,
|
123
|
+
input_data.Xc, input_data.Xc_ind, input_data.Xc_indptr,
|
124
|
+
model_params.missing_action, workspace.xmin, workspace.xmax, workspace.unsplittable);
|
125
|
+
}
|
126
|
+
|
127
|
+
else
|
128
|
+
{
|
129
|
+
get_categs(workspace.ix_arr.data(), input_data.categ_data + input_data.nrows * workspace.col_chosen,
|
130
|
+
workspace.st, workspace.end, input_data.ncat[workspace.col_chosen],
|
131
|
+
model_params.missing_action, workspace.categs.data(), workspace.npresent, workspace.unsplittable);
|
132
|
+
}
|
133
|
+
}
|
134
|
+
|
135
|
+
int choose_cat_from_present(WorkerMemory &workspace, InputData &input_data, size_t col_num)
|
136
|
+
{
|
137
|
+
int chosen_cat = std::uniform_int_distribution<int>
|
138
|
+
(0, workspace.npresent - 1)
|
139
|
+
(workspace.rnd_generator);
|
140
|
+
workspace.ncat_tried = 0;
|
141
|
+
for (int cat = 0; cat < input_data.ncat[col_num]; cat++)
|
142
|
+
{
|
143
|
+
if (workspace.categs[cat] > 0)
|
144
|
+
{
|
145
|
+
if (workspace.ncat_tried == chosen_cat)
|
146
|
+
return cat;
|
147
|
+
else
|
148
|
+
workspace.ncat_tried++;
|
149
|
+
}
|
150
|
+
}
|
151
|
+
|
152
|
+
return -1; /* this will never be reached, but CRAN complains otherwise */
|
153
|
+
}
|
154
|
+
|
155
|
+
void update_col_sampler(WorkerMemory &workspace, InputData &input_data)
|
156
|
+
{
|
157
|
+
if (!workspace.col_sampler.max())
|
158
|
+
return;
|
159
|
+
|
160
|
+
std::vector<double> col_weights = workspace.col_sampler.probabilities();
|
161
|
+
for (size_t col = 0; col < input_data.ncols_numeric; col++)
|
162
|
+
if (!workspace.cols_possible[col])
|
163
|
+
col_weights[col] = 0;
|
164
|
+
for (size_t col = 0; col < input_data.ncols_categ; col++)
|
165
|
+
if (!workspace.cols_possible[col + input_data.ncols_numeric])
|
166
|
+
col_weights[col + input_data.ncols_numeric] = 0;
|
167
|
+
workspace.col_sampler = std::discrete_distribution<size_t>(col_weights.begin(), col_weights.end());
|
168
|
+
}
|
169
|
+
|
170
|
+
bool is_col_taken(std::vector<bool> &col_is_taken, std::unordered_set<size_t> &col_is_taken_s,
|
171
|
+
InputData &input_data, size_t col_num, ColType col_type)
|
172
|
+
{
|
173
|
+
col_num += ((col_type == Categorical)? 0 : input_data.ncols_categ);
|
174
|
+
if (col_is_taken.size())
|
175
|
+
return col_is_taken[col_num];
|
176
|
+
else
|
177
|
+
return col_is_taken_s.find(col_num) != col_is_taken_s.end();
|
178
|
+
}
|
179
|
+
|
180
|
+
void set_col_as_taken(std::vector<bool> &col_is_taken, std::unordered_set<size_t> &col_is_taken_s,
|
181
|
+
InputData &input_data, size_t col_num, ColType col_type)
|
182
|
+
{
|
183
|
+
col_num += ((col_type == Categorical)? 0 : input_data.ncols_categ);
|
184
|
+
if (col_is_taken.size())
|
185
|
+
col_is_taken[col_num] = true;
|
186
|
+
else
|
187
|
+
col_is_taken_s.insert(col_num);
|
188
|
+
}
|
189
|
+
|
190
|
+
void add_separation_step(WorkerMemory &workspace, InputData &input_data, double remainder)
|
191
|
+
{
|
192
|
+
if (workspace.weights_arr.size())
|
193
|
+
increase_comb_counter(workspace.ix_arr.data(), workspace.st, workspace.end,
|
194
|
+
input_data.nrows, workspace.tmat_sep.data(), workspace.weights_arr.data(), remainder);
|
195
|
+
else if (workspace.weights_map.size())
|
196
|
+
increase_comb_counter(workspace.ix_arr.data(), workspace.st, workspace.end,
|
197
|
+
input_data.nrows, workspace.tmat_sep.data(), workspace.weights_map, remainder);
|
198
|
+
else
|
199
|
+
increase_comb_counter(workspace.ix_arr.data(), workspace.st, workspace.end,
|
200
|
+
input_data.nrows, workspace.tmat_sep.data(), remainder);
|
201
|
+
}
|
202
|
+
|
203
|
+
void add_remainder_separation_steps(WorkerMemory &workspace, InputData &input_data, long double sum_weight)
|
204
|
+
{
|
205
|
+
if (
|
206
|
+
((workspace.end - workspace.st) > 0 && !workspace.weights_arr.size() && !workspace.weights_map.size()) ||
|
207
|
+
(sum_weight > 1 && (workspace.weights_arr.size() || workspace.weights_map.size()))
|
208
|
+
)
|
209
|
+
{
|
210
|
+
double expected_dsep;
|
211
|
+
if (!workspace.weights_arr.size() && !workspace.weights_map.size())
|
212
|
+
expected_dsep = expected_separation_depth(workspace.end - workspace.st + 1);
|
213
|
+
else
|
214
|
+
expected_dsep = expected_separation_depth(sum_weight);
|
215
|
+
|
216
|
+
add_separation_step(workspace, input_data, expected_dsep + 1);
|
217
|
+
}
|
218
|
+
}
|
219
|
+
|
220
|
+
void remap_terminal_trees(IsoForest *model_outputs, ExtIsoForest *model_outputs_ext,
|
221
|
+
PredictionData &prediction_data, sparse_ix *restrict tree_num, int nthreads)
|
222
|
+
{
|
223
|
+
size_t ntrees = (model_outputs != NULL)? model_outputs->trees.size() : model_outputs_ext->hplanes.size();
|
224
|
+
size_t max_tree, curr_term;
|
225
|
+
std::vector<sparse_ix> tree_mapping;
|
226
|
+
if (model_outputs != NULL)
|
227
|
+
{
|
228
|
+
max_tree = std::accumulate(model_outputs->trees.begin(),
|
229
|
+
model_outputs->trees.end(),
|
230
|
+
(size_t)0,
|
231
|
+
[](const size_t curr_max, const std::vector<IsoTree> &tr)
|
232
|
+
{return std::max(curr_max, tr.size());});
|
233
|
+
tree_mapping.resize(max_tree);
|
234
|
+
for (size_t tree = 0; tree < ntrees; tree++)
|
235
|
+
{
|
236
|
+
std::fill(tree_mapping.begin(), tree_mapping.end(), (size_t)0);
|
237
|
+
curr_term = 0;
|
238
|
+
for (size_t node = 0; node < model_outputs->trees[tree].size(); node++)
|
239
|
+
if (model_outputs->trees[tree][node].score >= 0)
|
240
|
+
tree_mapping[node] = curr_term++;
|
241
|
+
|
242
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) shared(tree_num, tree_mapping, tree, prediction_data)
|
243
|
+
for (size_t_for row = 0; row < prediction_data.nrows; row++)
|
244
|
+
tree_num[row + tree * prediction_data.nrows] = tree_mapping[tree_num[row + tree * prediction_data.nrows]];
|
245
|
+
}
|
246
|
+
}
|
247
|
+
|
248
|
+
else
|
249
|
+
{
|
250
|
+
max_tree = std::accumulate(model_outputs_ext->hplanes.begin(),
|
251
|
+
model_outputs_ext->hplanes.end(),
|
252
|
+
(size_t)0,
|
253
|
+
[](const size_t curr_max, const std::vector<IsoHPlane> &tr)
|
254
|
+
{return std::max(curr_max, tr.size());});
|
255
|
+
tree_mapping.resize(max_tree);
|
256
|
+
for (size_t tree = 0; tree < ntrees; tree++)
|
257
|
+
{
|
258
|
+
std::fill(tree_mapping.begin(), tree_mapping.end(), (size_t)0);
|
259
|
+
curr_term = 0;
|
260
|
+
for (size_t node = 0; node < model_outputs_ext->hplanes[tree].size(); node++)
|
261
|
+
if (model_outputs_ext->hplanes[tree][node].score >= 0)
|
262
|
+
tree_mapping[node] = curr_term++;
|
263
|
+
|
264
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) shared(tree_num, tree_mapping, tree, prediction_data)
|
265
|
+
for (size_t_for row = 0; row < prediction_data.nrows; row++)
|
266
|
+
tree_num[row + tree * prediction_data.nrows] = tree_mapping[tree_num[row + tree * prediction_data.nrows]];
|
267
|
+
}
|
268
|
+
}
|
269
|
+
}
|
270
|
+
|
271
|
+
void backup_recursion_state(WorkerMemory &workspace, RecursionState &recursion_state)
|
272
|
+
{
|
273
|
+
recursion_state.st = workspace.st;
|
274
|
+
recursion_state.st_NA = workspace.st_NA;
|
275
|
+
recursion_state.end_NA = workspace.end_NA;
|
276
|
+
recursion_state.split_ix = workspace.split_ix;
|
277
|
+
recursion_state.end = workspace.end;
|
278
|
+
recursion_state.cols_possible = workspace.cols_possible;
|
279
|
+
recursion_state.col_sampler = workspace.col_sampler;
|
280
|
+
|
281
|
+
/* for the extended model, it's not necessary to copy everything */
|
282
|
+
if (!workspace.comb_val.size())
|
283
|
+
{
|
284
|
+
/* TODO: here only need to copy the left half, as the right one is untouched */
|
285
|
+
recursion_state.ix_arr = workspace.ix_arr;
|
286
|
+
recursion_state.weights_map = workspace.weights_map;
|
287
|
+
recursion_state.weights_arr = workspace.weights_arr;
|
288
|
+
}
|
289
|
+
}
|
290
|
+
|
291
|
+
|
292
|
+
void restore_recursion_state(WorkerMemory &workspace, RecursionState &recursion_state)
|
293
|
+
{
|
294
|
+
workspace.st = recursion_state.st;
|
295
|
+
workspace.st_NA = recursion_state.st_NA;
|
296
|
+
workspace.end_NA = recursion_state.end_NA;
|
297
|
+
workspace.split_ix = recursion_state.split_ix;
|
298
|
+
workspace.end = recursion_state.end;
|
299
|
+
workspace.cols_possible = std::move(recursion_state.cols_possible);
|
300
|
+
workspace.col_sampler = std::move(recursion_state.col_sampler);
|
301
|
+
|
302
|
+
if (!workspace.comb_val.size())
|
303
|
+
{
|
304
|
+
/* TODO: here only need to copy the left half, as the right one is untouched */
|
305
|
+
workspace.ix_arr = std::move(recursion_state.ix_arr);
|
306
|
+
workspace.weights_map = std::move(recursion_state.weights_map);
|
307
|
+
workspace.weights_arr = std::move(recursion_state.weights_arr);
|
308
|
+
}
|
309
|
+
}
|
@@ -0,0 +1,1205 @@
|
|
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
|
+
|
48
|
+
/* TODO: this file is a complete mess, needs a refactor from scratch along with the data structs */
|
49
|
+
|
50
|
+
/* Impute missing values in new data
|
51
|
+
*
|
52
|
+
* Parameters
|
53
|
+
* ==========
|
54
|
+
* - numeric_data[nrows * ncols_numeric] (in, out)
|
55
|
+
* Pointer to numeric data in which missing values will be imputed. Must be ordered by columns like Fortran,
|
56
|
+
* not ordered by rows like C (i.e. entries 1..n contain column 0, n+1..2n column 1, etc.),
|
57
|
+
* and the column order must be the same as in the data that was used to fit the model.
|
58
|
+
* Pass NULL if there are no dense numeric columns.
|
59
|
+
* Can only pass one of 'numeric_data', 'Xr' + 'Xr_ind' + 'Xr_indptr'.
|
60
|
+
* Imputations will overwrite values in this same array.
|
61
|
+
* - ncols_numeric
|
62
|
+
* Number of numeric columns in the data (whether they come in a sparse matrix or dense array).
|
63
|
+
* - categ_data[nrows * ncols_categ]
|
64
|
+
* Pointer to categorical data in which missing values will be imputed. Must be ordered by columns like Fortran,
|
65
|
+
* not ordered by rows like C (i.e. entries 1..n contain column 0, n+1..2n column 1, etc.),
|
66
|
+
* and the column order must be the same as in the data that was used to fit the model.
|
67
|
+
* Pass NULL if there are no categorical columns.
|
68
|
+
* Each category should be represented as an integer, and these integers must start at zero and
|
69
|
+
* be in consecutive order - i.e. if category '3' is present, category '2' must have also been
|
70
|
+
* present when the model was fit (note that they are not treated as being ordinal, this is just
|
71
|
+
* an encoding). Missing values should be encoded as negative numbers such as (-1). The encoding
|
72
|
+
* must be the same as was used in the data to which the model was fit.
|
73
|
+
* Imputations will overwrite values in this same array.
|
74
|
+
* - ncols_categ
|
75
|
+
* Number of categorical columns in the data.
|
76
|
+
* - ncat[ncols_categ]
|
77
|
+
* Number of categories in each categorical column. E.g. if the highest code for a column is '4',
|
78
|
+
* the number of categories for that column is '5' (zero is one category).
|
79
|
+
* Must be the same as was passed to 'fit_iforest'.
|
80
|
+
* - Xr[nnz] (in, out)
|
81
|
+
* Pointer to numeric data in sparse numeric matrix in CSR format (row-compressed).
|
82
|
+
* Pass NULL if there are no sparse numeric columns.
|
83
|
+
* Can only pass one of 'numeric_data', 'Xr' + 'Xr_ind' + 'Xr_indptr'.
|
84
|
+
* Imputations will overwrite values in this same array.
|
85
|
+
* - Xr_ind[nnz]
|
86
|
+
* Pointer to column indices to which each non-zero entry in 'Xr' corresponds.
|
87
|
+
* Pass NULL if there are no sparse numeric columns in CSR format.
|
88
|
+
* - Xr_indptr[nrows + 1]
|
89
|
+
* Pointer to row index pointers that tell at entry [row] where does row 'row'
|
90
|
+
* start and at entry [row + 1] where does row 'row' end.
|
91
|
+
* Pass NULL if there are no sparse numeric columns in CSR format.
|
92
|
+
* - nrows
|
93
|
+
* Number of rows in 'numeric_data', 'Xc', 'Xr, 'categ_data'.
|
94
|
+
* - nthreads
|
95
|
+
* Number of parallel threads to use. Note that, the more threads, the more memory will be
|
96
|
+
* allocated, even if the thread does not end up being used. Ignored when not building with
|
97
|
+
* OpenMP support.
|
98
|
+
* - model_outputs
|
99
|
+
* Pointer to fitted single-variable model object from function 'fit_iforest'. Pass NULL
|
100
|
+
* if the predictions are to be made from an extended model. Can only pass one of
|
101
|
+
* 'model_outputs' and 'model_outputs_ext'.
|
102
|
+
* - model_outputs_ext
|
103
|
+
* Pointer to fitted extended model object from function 'fit_iforest'. Pass NULL
|
104
|
+
* if the predictions are to be made from a single-variable model. Can only pass one of
|
105
|
+
* 'model_outputs' and 'model_outputs_ext'.
|
106
|
+
* - impute_nodes
|
107
|
+
* Pointer to fitted imputation node obects for the same trees as in 'model_outputs' or 'model_outputs_ext',
|
108
|
+
* as produced from function 'fit_iforest',
|
109
|
+
*/
|
110
|
+
void impute_missing_values(double numeric_data[], int categ_data[],
|
111
|
+
double Xr[], sparse_ix Xr_ind[], sparse_ix Xr_indptr[],
|
112
|
+
size_t nrows, int nthreads,
|
113
|
+
IsoForest *model_outputs, ExtIsoForest *model_outputs_ext,
|
114
|
+
Imputer &imputer)
|
115
|
+
{
|
116
|
+
PredictionData prediction_data = {numeric_data, categ_data, nrows,
|
117
|
+
NULL, NULL, NULL,
|
118
|
+
Xr, Xr_ind, Xr_indptr};
|
119
|
+
|
120
|
+
std::vector<size_t> ix_arr(nrows);
|
121
|
+
std::iota(ix_arr.begin(), ix_arr.end(), (size_t) 0);
|
122
|
+
|
123
|
+
size_t end = check_for_missing(prediction_data, imputer, ix_arr.data(), nthreads);
|
124
|
+
|
125
|
+
if (end == 0)
|
126
|
+
return;
|
127
|
+
|
128
|
+
if ((size_t)nthreads > end)
|
129
|
+
nthreads = (int)end;
|
130
|
+
#ifdef _OPENMP
|
131
|
+
std::vector<ImputedData> imp_memory(nthreads);
|
132
|
+
#else
|
133
|
+
std::vector<ImputedData> imp_memory(1);
|
134
|
+
#endif
|
135
|
+
|
136
|
+
|
137
|
+
if (model_outputs != NULL)
|
138
|
+
{
|
139
|
+
#pragma omp parallel for schedule(dynamic) num_threads(nthreads) \
|
140
|
+
shared(end, imp_memory, prediction_data, model_outputs, ix_arr, imputer)
|
141
|
+
for (size_t_for row = 0; row < end; row++)
|
142
|
+
{
|
143
|
+
initialize_impute_calc(imp_memory[omp_get_thread_num()], prediction_data, imputer, ix_arr[row]);
|
144
|
+
|
145
|
+
for (std::vector<IsoTree> &tree : model_outputs->trees)
|
146
|
+
{
|
147
|
+
traverse_itree(tree,
|
148
|
+
*model_outputs,
|
149
|
+
prediction_data,
|
150
|
+
&imputer.imputer_tree[&tree - &(model_outputs->trees[0])],
|
151
|
+
&imp_memory[omp_get_thread_num()],
|
152
|
+
(double) 1,
|
153
|
+
ix_arr[row],
|
154
|
+
NULL,
|
155
|
+
(size_t) 0);
|
156
|
+
}
|
157
|
+
|
158
|
+
apply_imputation_results(prediction_data, imp_memory[omp_get_thread_num()], imputer, (size_t) ix_arr[row]);
|
159
|
+
|
160
|
+
}
|
161
|
+
}
|
162
|
+
|
163
|
+
else
|
164
|
+
{
|
165
|
+
double temp;
|
166
|
+
#pragma omp parallel for schedule(dynamic) num_threads(nthreads) \
|
167
|
+
shared(end, imp_memory, prediction_data, model_outputs_ext, ix_arr, imputer) \
|
168
|
+
private(temp)
|
169
|
+
for (size_t_for row = 0; row < end; row++)
|
170
|
+
{
|
171
|
+
initialize_impute_calc(imp_memory[omp_get_thread_num()], prediction_data, imputer, ix_arr[row]);
|
172
|
+
|
173
|
+
for (std::vector<IsoHPlane> &hplane : model_outputs_ext->hplanes)
|
174
|
+
{
|
175
|
+
traverse_hplane(hplane,
|
176
|
+
*model_outputs_ext,
|
177
|
+
prediction_data,
|
178
|
+
temp,
|
179
|
+
&imputer.imputer_tree[&hplane - &(model_outputs_ext->hplanes[0])],
|
180
|
+
&imp_memory[omp_get_thread_num()],
|
181
|
+
NULL,
|
182
|
+
ix_arr[row]);
|
183
|
+
}
|
184
|
+
|
185
|
+
apply_imputation_results(prediction_data, imp_memory[omp_get_thread_num()], imputer, (size_t) ix_arr[row]);
|
186
|
+
|
187
|
+
}
|
188
|
+
}
|
189
|
+
|
190
|
+
}
|
191
|
+
|
192
|
+
void initialize_imputer(Imputer &imputer, InputData &input_data, size_t ntrees, int nthreads)
|
193
|
+
{
|
194
|
+
imputer.ncols_numeric = input_data.ncols_numeric;
|
195
|
+
imputer.ncols_categ = input_data.ncols_categ;
|
196
|
+
imputer.ncat.assign(input_data.ncat, input_data.ncat + input_data.ncols_categ);
|
197
|
+
if (imputer.col_means.size())
|
198
|
+
{
|
199
|
+
imputer.col_means.resize(input_data.ncols_numeric);
|
200
|
+
std::fill(imputer.col_means.begin(), imputer.col_means.end(), 0);
|
201
|
+
}
|
202
|
+
|
203
|
+
else
|
204
|
+
{
|
205
|
+
imputer.col_means.resize(input_data.ncols_numeric, 0);
|
206
|
+
}
|
207
|
+
|
208
|
+
imputer.col_modes.resize(input_data.ncols_categ);
|
209
|
+
imputer.imputer_tree = std::vector<std::vector<ImputeNode>>(ntrees);
|
210
|
+
|
211
|
+
size_t offset, cnt;
|
212
|
+
if (input_data.numeric_data != NULL)
|
213
|
+
{
|
214
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) private(cnt, offset) shared(input_data, imputer)
|
215
|
+
for (size_t_for col = 0; col < input_data.ncols_numeric; col++)
|
216
|
+
{
|
217
|
+
cnt = input_data.nrows;
|
218
|
+
offset = col * input_data.nrows;
|
219
|
+
for (size_t row = 0; row < input_data.nrows; row++)
|
220
|
+
{
|
221
|
+
imputer.col_means[col] += (!is_na_or_inf(input_data.numeric_data[row + offset]))?
|
222
|
+
input_data.numeric_data[row + offset] : 0;
|
223
|
+
cnt -= is_na_or_inf(input_data.numeric_data[row + offset]);
|
224
|
+
}
|
225
|
+
imputer.col_means[col] /= (long double) cnt;
|
226
|
+
}
|
227
|
+
}
|
228
|
+
|
229
|
+
else if (input_data.Xc != NULL)
|
230
|
+
{
|
231
|
+
#pragma omp parallel for schedule(dynamic) num_threads(nthreads) private(cnt) shared(input_data, imputer)
|
232
|
+
for (size_t_for col = 0; col < input_data.ncols_numeric; col++)
|
233
|
+
{
|
234
|
+
cnt = input_data.nrows;
|
235
|
+
for (size_t ix = input_data.Xc_indptr[col]; ix < input_data.Xc_indptr[col + 1]; ix++)
|
236
|
+
{
|
237
|
+
imputer.col_means[col] += (!is_na_or_inf(input_data.Xc[ix]))?
|
238
|
+
input_data.Xc[ix] : 0;
|
239
|
+
cnt -= is_na_or_inf(input_data.Xc[ix]);
|
240
|
+
}
|
241
|
+
imputer.col_means[col] /= (long double) cnt;
|
242
|
+
}
|
243
|
+
}
|
244
|
+
|
245
|
+
if (input_data.categ_data != NULL)
|
246
|
+
{
|
247
|
+
std::vector<size_t> cat_counts(input_data.max_categ);
|
248
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) firstprivate(cat_counts) private(offset) shared(input_data, imputer)
|
249
|
+
for (size_t_for col = 0; col < input_data.ncols_categ; col++)
|
250
|
+
{
|
251
|
+
std::fill(cat_counts.begin(), cat_counts.end(), 0);
|
252
|
+
offset = col * input_data.nrows;
|
253
|
+
for (size_t row = 0; row < input_data.nrows; row++)
|
254
|
+
{
|
255
|
+
if (input_data.categ_data[row + offset] >= 0)
|
256
|
+
cat_counts[input_data.categ_data[row + offset]]++;
|
257
|
+
}
|
258
|
+
imputer.col_modes[col] = (int) std::distance(cat_counts.begin(),
|
259
|
+
std::max_element(cat_counts.begin(),
|
260
|
+
cat_counts.begin() + input_data.ncat[col]));
|
261
|
+
}
|
262
|
+
}
|
263
|
+
}
|
264
|
+
|
265
|
+
|
266
|
+
/* https://en.wikipedia.org/wiki/Kahan_summation_algorithm */
|
267
|
+
void build_impute_node(ImputeNode &imputer, WorkerMemory &workspace,
|
268
|
+
InputData &input_data, ModelParams &model_params,
|
269
|
+
std::vector<ImputeNode> &imputer_tree,
|
270
|
+
size_t curr_depth, size_t min_imp_obs)
|
271
|
+
{
|
272
|
+
double wsum;
|
273
|
+
bool has_weights = workspace.weights_arr.size() || workspace.weights_map.size();
|
274
|
+
if (!has_weights)
|
275
|
+
wsum = (double)(workspace.end - workspace.st + 1);
|
276
|
+
else
|
277
|
+
wsum = calculate_sum_weights(workspace.ix_arr, workspace.st, workspace.end, curr_depth,
|
278
|
+
workspace.weights_arr, workspace.weights_map);
|
279
|
+
|
280
|
+
imputer.num_sum.resize(input_data.ncols_numeric, 0);
|
281
|
+
imputer.num_weight.resize(input_data.ncols_numeric, 0);
|
282
|
+
imputer.cat_sum.resize(input_data.ncols_categ);
|
283
|
+
imputer.cat_weight.resize(input_data.ncols_categ, 0);
|
284
|
+
imputer.num_sum.shrink_to_fit();
|
285
|
+
imputer.num_weight.shrink_to_fit();
|
286
|
+
imputer.cat_sum.shrink_to_fit();
|
287
|
+
imputer.cat_weight.shrink_to_fit();
|
288
|
+
|
289
|
+
/* Note: in theory, 'num_weight' could be initialized to 'wsum',
|
290
|
+
and the entries could get subtracted the weight afterwards, but due to rounding
|
291
|
+
error, this could produce some cases of no-present observations having positive
|
292
|
+
weight, or cases of negative weight, so it's better to add it for each row after
|
293
|
+
checking for possible NAs, even though it's less computationally efficient.
|
294
|
+
For sparse matrices it's done the other way as otherwise it would be too slow. */
|
295
|
+
|
296
|
+
for (size_t col = 0; col < input_data.ncols_categ; col++)
|
297
|
+
{
|
298
|
+
imputer.cat_sum[col].resize(input_data.ncat[col]);
|
299
|
+
imputer.cat_sum[col].shrink_to_fit();
|
300
|
+
}
|
301
|
+
|
302
|
+
double xnum;
|
303
|
+
int xcat;
|
304
|
+
double weight;
|
305
|
+
size_t ix;
|
306
|
+
|
307
|
+
if ((input_data.Xc == NULL && input_data.ncols_numeric) || input_data.ncols_categ)
|
308
|
+
{
|
309
|
+
if (!has_weights)
|
310
|
+
{
|
311
|
+
size_t cnt;
|
312
|
+
if (input_data.numeric_data != NULL)
|
313
|
+
{
|
314
|
+
for (size_t col = 0; col < input_data.ncols_numeric; col++)
|
315
|
+
{
|
316
|
+
cnt = 0;
|
317
|
+
for (size_t row = workspace.st; row <= workspace.end; row++)
|
318
|
+
{
|
319
|
+
xnum = input_data.numeric_data[workspace.ix_arr[row] + col * input_data.nrows];
|
320
|
+
if (!is_na_or_inf(xnum))
|
321
|
+
{
|
322
|
+
cnt++;
|
323
|
+
imputer.num_sum[col] += (xnum - imputer.num_sum[col]) / (long double)cnt;
|
324
|
+
}
|
325
|
+
}
|
326
|
+
imputer.num_weight[col] = (double) cnt;
|
327
|
+
}
|
328
|
+
}
|
329
|
+
|
330
|
+
if (input_data.categ_data != NULL)
|
331
|
+
{
|
332
|
+
for (size_t col = 0; col < input_data.ncols_categ; col++)
|
333
|
+
{
|
334
|
+
cnt = 0;
|
335
|
+
for (size_t row = workspace.st; row <= workspace.end; row++)
|
336
|
+
{
|
337
|
+
xcat = input_data.categ_data[workspace.ix_arr[row] + col * input_data.nrows];
|
338
|
+
if (xcat >= 0)
|
339
|
+
{
|
340
|
+
cnt++;
|
341
|
+
imputer.cat_sum[col][xcat]++; /* later gets divided */
|
342
|
+
}
|
343
|
+
}
|
344
|
+
imputer.cat_weight[col] = (double) cnt;
|
345
|
+
}
|
346
|
+
}
|
347
|
+
|
348
|
+
}
|
349
|
+
|
350
|
+
else
|
351
|
+
{
|
352
|
+
long double prod_sum, corr, val, diff;
|
353
|
+
if (input_data.numeric_data != NULL)
|
354
|
+
{
|
355
|
+
for (size_t col = 0; col < input_data.ncols_numeric; col++)
|
356
|
+
{
|
357
|
+
prod_sum = 0; corr = 0;
|
358
|
+
for (size_t row = workspace.st; row <= workspace.end; row++)
|
359
|
+
{
|
360
|
+
xnum = input_data.numeric_data[workspace.ix_arr[row] + col * input_data.nrows];
|
361
|
+
if (!is_na_or_inf(xnum))
|
362
|
+
{
|
363
|
+
if (workspace.weights_arr.size())
|
364
|
+
weight = workspace.weights_arr[workspace.ix_arr[row]];
|
365
|
+
else
|
366
|
+
weight = workspace.weights_map[workspace.ix_arr[row]];
|
367
|
+
|
368
|
+
imputer.num_weight[col] += weight; /* these are always <= 1 */
|
369
|
+
val = (xnum * weight) - corr;
|
370
|
+
diff = prod_sum + val;
|
371
|
+
corr = (diff - prod_sum) - val;
|
372
|
+
prod_sum = diff;
|
373
|
+
}
|
374
|
+
}
|
375
|
+
imputer.num_sum[col] = prod_sum / imputer.num_weight[col];
|
376
|
+
}
|
377
|
+
}
|
378
|
+
|
379
|
+
|
380
|
+
if (input_data.ncols_categ)
|
381
|
+
{
|
382
|
+
for (size_t row = workspace.st; row <= workspace.end; row++)
|
383
|
+
{
|
384
|
+
ix = workspace.ix_arr[row];
|
385
|
+
if (workspace.weights_arr.size())
|
386
|
+
weight = workspace.weights_arr[ix];
|
387
|
+
else
|
388
|
+
weight = workspace.weights_map[ix];
|
389
|
+
|
390
|
+
for (size_t col = 0; col < input_data.ncols_categ; col++)
|
391
|
+
{
|
392
|
+
xcat = input_data.categ_data[ix + col * input_data.nrows];
|
393
|
+
if (xcat >= 0)
|
394
|
+
{
|
395
|
+
imputer.cat_sum[col][xcat] += weight; /* later gets divided */
|
396
|
+
imputer.cat_weight[col] += weight;
|
397
|
+
}
|
398
|
+
}
|
399
|
+
}
|
400
|
+
}
|
401
|
+
}
|
402
|
+
}
|
403
|
+
|
404
|
+
if (input_data.Xc != NULL) /* sparse numeric */
|
405
|
+
{
|
406
|
+
size_t *ix_arr = workspace.ix_arr.data();
|
407
|
+
size_t st_col, end_col, ind_end_col, curr_pos;
|
408
|
+
std::fill(imputer.num_weight.begin(), imputer.num_weight.end(), wsum);
|
409
|
+
|
410
|
+
for (size_t col = 0; col < input_data.ncols_numeric; col++)
|
411
|
+
{
|
412
|
+
st_col = input_data.Xc_indptr[col];
|
413
|
+
end_col = input_data.Xc_indptr[col + 1] - 1;
|
414
|
+
ind_end_col = input_data.Xc_ind[end_col];
|
415
|
+
curr_pos = st_col;
|
416
|
+
for (size_t *row = std::lower_bound(ix_arr + workspace.st, ix_arr + workspace.end + 1, input_data.Xc_ind[st_col]);
|
417
|
+
row != ix_arr + workspace.end + 1 && curr_pos != end_col + 1 && ind_end_col >= *row;
|
418
|
+
)
|
419
|
+
{
|
420
|
+
if (input_data.Xc_ind[curr_pos] == *row)
|
421
|
+
{
|
422
|
+
xnum = input_data.Xc[curr_pos];
|
423
|
+
if (workspace.weights_arr.size())
|
424
|
+
weight = workspace.weights_arr[*row];
|
425
|
+
else if (workspace.weights_map.size())
|
426
|
+
weight = workspace.weights_map[*row];
|
427
|
+
else
|
428
|
+
weight = 1;
|
429
|
+
|
430
|
+
if (!is_na_or_inf(xnum))
|
431
|
+
{
|
432
|
+
imputer.num_sum[col] += weight * xnum;
|
433
|
+
}
|
434
|
+
|
435
|
+
else
|
436
|
+
{
|
437
|
+
imputer.num_weight[col] -= weight;
|
438
|
+
}
|
439
|
+
|
440
|
+
if (row == ix_arr + workspace.end || curr_pos == end_col) break;
|
441
|
+
curr_pos = std::lower_bound(input_data.Xc_ind + curr_pos, input_data.Xc_ind + end_col + 1, *(++row)) - input_data.Xc_ind;
|
442
|
+
}
|
443
|
+
|
444
|
+
else
|
445
|
+
{
|
446
|
+
if (input_data.Xc_ind[curr_pos] > *row)
|
447
|
+
row = std::lower_bound(row + 1, ix_arr + workspace.end + 1, input_data.Xc_ind[curr_pos]);
|
448
|
+
else
|
449
|
+
curr_pos = std::lower_bound(input_data.Xc_ind + curr_pos + 1, input_data.Xc_ind + end_col + 1, *row) - input_data.Xc_ind;
|
450
|
+
}
|
451
|
+
}
|
452
|
+
|
453
|
+
imputer.num_sum[col] /= imputer.num_weight[col];
|
454
|
+
}
|
455
|
+
}
|
456
|
+
|
457
|
+
/* if any value is not possible to impute, look it up from the parent tree, but assign a lesser weight
|
458
|
+
Note: in theory, the parent node should always have some imputation value for every variable, but due to
|
459
|
+
numeric rounding errors, it might have a weight of zero, so in those cases it's looked up higher up the
|
460
|
+
tree instead. */
|
461
|
+
size_t look_aboves, curr_tree;
|
462
|
+
double min_imp_obs_dbl = (double) min_imp_obs;
|
463
|
+
if (imputer.num_sum.size())
|
464
|
+
{
|
465
|
+
for (size_t col = 0; col < input_data.ncols_numeric; col++)
|
466
|
+
{
|
467
|
+
if (imputer.num_weight[col] < min_imp_obs_dbl)
|
468
|
+
{
|
469
|
+
look_aboves = 1;
|
470
|
+
curr_tree = imputer.parent;
|
471
|
+
while (true)
|
472
|
+
{
|
473
|
+
if (!is_na_or_inf(imputer_tree[curr_tree].num_sum[col]))
|
474
|
+
{
|
475
|
+
imputer.num_sum[col] = imputer_tree[curr_tree].num_sum[col] / imputer_tree[curr_tree].num_weight[col];
|
476
|
+
imputer.num_weight[col] = wsum / (double)(2 * look_aboves);
|
477
|
+
break;
|
478
|
+
}
|
479
|
+
|
480
|
+
else if (curr_tree > 0)
|
481
|
+
{
|
482
|
+
curr_tree = imputer_tree[curr_tree].parent;
|
483
|
+
look_aboves++;
|
484
|
+
}
|
485
|
+
|
486
|
+
else /* will only happen if every single value is missing */
|
487
|
+
{
|
488
|
+
imputer.num_sum[col] = NAN;
|
489
|
+
imputer.num_weight[col] = 0;
|
490
|
+
break;
|
491
|
+
}
|
492
|
+
}
|
493
|
+
}
|
494
|
+
}
|
495
|
+
}
|
496
|
+
|
497
|
+
if (imputer.cat_sum.size())
|
498
|
+
{
|
499
|
+
for (size_t col = 0; col < input_data.ncols_categ; col++)
|
500
|
+
{
|
501
|
+
if (imputer.cat_weight[col] >= min_imp_obs_dbl)
|
502
|
+
{
|
503
|
+
for (double &cat : imputer.cat_sum[col])
|
504
|
+
cat /= imputer.cat_weight[col];
|
505
|
+
}
|
506
|
+
|
507
|
+
else
|
508
|
+
{
|
509
|
+
look_aboves = 1;
|
510
|
+
curr_tree = imputer.parent;
|
511
|
+
while (true)
|
512
|
+
{
|
513
|
+
if (imputer_tree[curr_tree].cat_weight[col] > 0)
|
514
|
+
{
|
515
|
+
for (int cat = 0; cat < input_data.ncat[col]; cat++)
|
516
|
+
{
|
517
|
+
imputer.cat_sum[col][cat] += imputer_tree[curr_tree].cat_sum[col][cat] / imputer.cat_weight[col];
|
518
|
+
imputer.cat_weight[col] = wsum / (double)(2 * look_aboves);
|
519
|
+
}
|
520
|
+
break;
|
521
|
+
}
|
522
|
+
|
523
|
+
else if (curr_tree > 0)
|
524
|
+
{
|
525
|
+
curr_tree = imputer_tree[curr_tree].parent;
|
526
|
+
look_aboves++;
|
527
|
+
}
|
528
|
+
|
529
|
+
else /* will only happen if every single value is missing */
|
530
|
+
{
|
531
|
+
break;
|
532
|
+
}
|
533
|
+
}
|
534
|
+
imputer.cat_weight[col] = std::accumulate(imputer.cat_sum[col].begin(),
|
535
|
+
imputer.cat_sum[col].end(),
|
536
|
+
(double) 0);
|
537
|
+
}
|
538
|
+
}
|
539
|
+
}
|
540
|
+
|
541
|
+
/* re-adjust the weights according to parameters
|
542
|
+
(note that by this point, the weights are a sum) */
|
543
|
+
switch(model_params.weigh_imp_rows)
|
544
|
+
{
|
545
|
+
case Inverse:
|
546
|
+
{
|
547
|
+
double wsum_div = wsum * sqrt(wsum);
|
548
|
+
for (double &w : imputer.num_weight)
|
549
|
+
w /= wsum_div;
|
550
|
+
|
551
|
+
for (double &w : imputer.cat_weight)
|
552
|
+
w /= wsum_div;
|
553
|
+
break;
|
554
|
+
}
|
555
|
+
|
556
|
+
case Flat:
|
557
|
+
{
|
558
|
+
for (double &w : imputer.num_weight)
|
559
|
+
w /= wsum;
|
560
|
+
for (double &w : imputer.cat_weight)
|
561
|
+
w /= wsum;
|
562
|
+
break;
|
563
|
+
}
|
564
|
+
|
565
|
+
/* TODO: maybe divide by nrows for prop */
|
566
|
+
}
|
567
|
+
|
568
|
+
double curr_depth_dbl = (double) (curr_depth + 1);
|
569
|
+
switch(model_params.depth_imp)
|
570
|
+
{
|
571
|
+
case Lower:
|
572
|
+
{
|
573
|
+
for (double &w : imputer.num_weight)
|
574
|
+
w /= curr_depth_dbl;
|
575
|
+
for (double &w : imputer.cat_weight)
|
576
|
+
w /= curr_depth_dbl;
|
577
|
+
break;
|
578
|
+
}
|
579
|
+
|
580
|
+
case Higher:
|
581
|
+
{
|
582
|
+
for (double &w : imputer.num_weight)
|
583
|
+
w *= curr_depth_dbl;
|
584
|
+
for (double &w : imputer.cat_weight)
|
585
|
+
w *= curr_depth_dbl;
|
586
|
+
break;
|
587
|
+
}
|
588
|
+
}
|
589
|
+
|
590
|
+
/* now re-adjust sums */
|
591
|
+
if (model_params.weigh_imp_rows != Prop || model_params.depth_imp != Same)
|
592
|
+
{
|
593
|
+
for (size_t col = 0; col < input_data.ncols_numeric; col++)
|
594
|
+
imputer.num_sum[col] *= imputer.num_weight[col];
|
595
|
+
|
596
|
+
for (size_t col = 0; col < input_data.ncols_categ; col++)
|
597
|
+
for (int cat = 0; cat < input_data.ncat[col]; cat++)
|
598
|
+
imputer.cat_sum[col][cat] *= imputer.cat_weight[col];
|
599
|
+
}
|
600
|
+
}
|
601
|
+
|
602
|
+
|
603
|
+
void shrink_impute_node(ImputeNode &imputer)
|
604
|
+
{
|
605
|
+
imputer.num_sum.clear();
|
606
|
+
imputer.num_weight.clear();
|
607
|
+
imputer.cat_sum.clear();
|
608
|
+
imputer.cat_weight.clear();
|
609
|
+
|
610
|
+
imputer.num_sum.shrink_to_fit();
|
611
|
+
imputer.num_weight.shrink_to_fit();
|
612
|
+
imputer.cat_sum.shrink_to_fit();
|
613
|
+
imputer.cat_weight.shrink_to_fit();
|
614
|
+
}
|
615
|
+
|
616
|
+
void drop_nonterminal_imp_node(std::vector<ImputeNode> &imputer_tree,
|
617
|
+
std::vector<IsoTree> *trees,
|
618
|
+
std::vector<IsoHPlane> *hplanes)
|
619
|
+
{
|
620
|
+
if (trees != NULL)
|
621
|
+
{
|
622
|
+
for (size_t tr = 0; tr < trees->size(); tr++)
|
623
|
+
{
|
624
|
+
if ((*trees)[tr].score <= 0)
|
625
|
+
{
|
626
|
+
shrink_impute_node(imputer_tree[tr]);
|
627
|
+
}
|
628
|
+
|
629
|
+
else
|
630
|
+
{
|
631
|
+
/* cat_weight is not needed for anything else */
|
632
|
+
imputer_tree[tr].cat_weight.clear();
|
633
|
+
imputer_tree[tr].cat_weight.shrink_to_fit();
|
634
|
+
}
|
635
|
+
}
|
636
|
+
}
|
637
|
+
|
638
|
+
else
|
639
|
+
{
|
640
|
+
for (size_t tr = 0; tr < hplanes->size(); tr++)
|
641
|
+
{
|
642
|
+
if ((*hplanes)[tr].score <= 0)
|
643
|
+
{
|
644
|
+
shrink_impute_node(imputer_tree[tr]);
|
645
|
+
}
|
646
|
+
|
647
|
+
else
|
648
|
+
{
|
649
|
+
/* cat_weight is not needed for anything else */
|
650
|
+
imputer_tree[tr].cat_weight.clear();
|
651
|
+
imputer_tree[tr].cat_weight.shrink_to_fit();
|
652
|
+
}
|
653
|
+
}
|
654
|
+
}
|
655
|
+
|
656
|
+
imputer_tree.shrink_to_fit();
|
657
|
+
}
|
658
|
+
|
659
|
+
void combine_imp_single(ImputedData &imp_addfrom, ImputedData &imp_addto)
|
660
|
+
{
|
661
|
+
size_t col;
|
662
|
+
for (size_t ix = 0; ix < imp_addfrom.n_missing_num; ix++)
|
663
|
+
{
|
664
|
+
imp_addto.num_sum[ix] += imp_addfrom.num_sum[ix];
|
665
|
+
imp_addto.num_weight[ix] += imp_addfrom.num_weight[ix];
|
666
|
+
}
|
667
|
+
|
668
|
+
for (size_t ix = 0; ix < imp_addfrom.n_missing_cat; ix++)
|
669
|
+
{
|
670
|
+
col = imp_addfrom.missing_cat[ix];
|
671
|
+
for (size_t cat = 0; cat < imp_addto.cat_sum[col].size(); cat++)
|
672
|
+
{
|
673
|
+
imp_addto.cat_sum[col][cat] += imp_addfrom.cat_sum[col][cat];
|
674
|
+
}
|
675
|
+
}
|
676
|
+
|
677
|
+
for (size_t ix = 0; ix < imp_addfrom.n_missing_sp; ix++)
|
678
|
+
{
|
679
|
+
imp_addto.sp_num_sum[ix] += imp_addfrom.sp_num_sum[ix];
|
680
|
+
imp_addto.sp_num_weight[ix] += imp_addfrom.sp_num_weight[ix];
|
681
|
+
}
|
682
|
+
}
|
683
|
+
|
684
|
+
void combine_tree_imputations(WorkerMemory &workspace,
|
685
|
+
std::vector<ImputedData> &impute_vec,
|
686
|
+
std::unordered_map<size_t, ImputedData> &impute_map,
|
687
|
+
std::vector<char> &has_missing,
|
688
|
+
int nthreads)
|
689
|
+
{
|
690
|
+
if (workspace.impute_vec.size())
|
691
|
+
{
|
692
|
+
#pragma omp parallel for schedule(dynamic) num_threads(nthreads) shared(has_missing, workspace, impute_vec)
|
693
|
+
for (size_t_for row = 0; row < has_missing.size(); row++)
|
694
|
+
if (has_missing[row])
|
695
|
+
combine_imp_single(workspace.impute_vec[row], impute_vec[row]);
|
696
|
+
}
|
697
|
+
|
698
|
+
else if (workspace.impute_map.size())
|
699
|
+
{
|
700
|
+
#pragma omp parallel for schedule(dynamic) num_threads(nthreads) shared(has_missing, workspace, impute_map)
|
701
|
+
for (size_t_for row = 0; row < has_missing.size(); row++)
|
702
|
+
if (has_missing[row])
|
703
|
+
combine_imp_single(workspace.impute_map[row], impute_map[row]);
|
704
|
+
}
|
705
|
+
}
|
706
|
+
|
707
|
+
|
708
|
+
void add_from_impute_node(ImputeNode &imputer, ImputedData &imputed_data, double w)
|
709
|
+
{
|
710
|
+
size_t col;
|
711
|
+
for (size_t ix = 0; ix < imputed_data.n_missing_num; ix++)
|
712
|
+
{
|
713
|
+
col = imputed_data.missing_num[ix];
|
714
|
+
imputed_data.num_sum[ix] += (!is_na_or_inf(imputer.num_sum[col]))? (w * imputer.num_sum[col]) : 0;
|
715
|
+
imputed_data.num_weight[ix] += w * imputer.num_weight[ix];
|
716
|
+
}
|
717
|
+
|
718
|
+
for (size_t ix = 0; ix < imputed_data.n_missing_sp; ix++)
|
719
|
+
{
|
720
|
+
col = imputed_data.missing_sp[ix];
|
721
|
+
imputed_data.sp_num_sum[ix] += (!is_na_or_inf(imputer.num_sum[col]))? (w * imputer.num_sum[col]) : 0;
|
722
|
+
imputed_data.sp_num_weight[ix] += w * imputer.num_weight[ix];
|
723
|
+
}
|
724
|
+
|
725
|
+
for (size_t ix = 0; ix < imputed_data.n_missing_cat; ix++)
|
726
|
+
{
|
727
|
+
col = imputed_data.missing_cat[ix];
|
728
|
+
for (size_t cat = 0; cat < imputer.cat_sum[col].size(); cat++)
|
729
|
+
imputed_data.cat_sum[col][cat] += w * imputer.cat_sum[col][cat];
|
730
|
+
}
|
731
|
+
}
|
732
|
+
|
733
|
+
|
734
|
+
void add_from_impute_node(ImputeNode &imputer, WorkerMemory &workspace, InputData &input_data)
|
735
|
+
{
|
736
|
+
if (workspace.impute_vec.size())
|
737
|
+
{
|
738
|
+
if (!workspace.weights_arr.size() && !workspace.weights_map.size())
|
739
|
+
{
|
740
|
+
for (size_t row = workspace.st; row <= workspace.end; row++)
|
741
|
+
if (input_data.has_missing[workspace.ix_arr[row]])
|
742
|
+
add_from_impute_node(imputer,
|
743
|
+
workspace.impute_vec[workspace.ix_arr[row]],
|
744
|
+
(double)1);
|
745
|
+
}
|
746
|
+
|
747
|
+
else if (workspace.weights_arr.size())
|
748
|
+
{
|
749
|
+
for (size_t row = workspace.st; row <= workspace.end; row++)
|
750
|
+
if (input_data.has_missing[workspace.ix_arr[row]])
|
751
|
+
add_from_impute_node(imputer,
|
752
|
+
workspace.impute_vec[workspace.ix_arr[row]],
|
753
|
+
workspace.weights_arr[workspace.ix_arr[row]]);
|
754
|
+
}
|
755
|
+
|
756
|
+
else
|
757
|
+
{
|
758
|
+
for (size_t row = workspace.st; row <= workspace.end; row++)
|
759
|
+
if (input_data.has_missing[workspace.ix_arr[row]])
|
760
|
+
add_from_impute_node(imputer,
|
761
|
+
workspace.impute_vec[workspace.ix_arr[row]],
|
762
|
+
workspace.weights_map[workspace.ix_arr[row]]);
|
763
|
+
}
|
764
|
+
}
|
765
|
+
|
766
|
+
else if (workspace.impute_map.size())
|
767
|
+
{
|
768
|
+
if (!workspace.weights_arr.size() && !workspace.weights_map.size())
|
769
|
+
{
|
770
|
+
for (size_t row = workspace.st; row <= workspace.end; row++)
|
771
|
+
if (input_data.has_missing[workspace.ix_arr[row]])
|
772
|
+
add_from_impute_node(imputer,
|
773
|
+
workspace.impute_map[workspace.ix_arr[row]],
|
774
|
+
(double)1);
|
775
|
+
}
|
776
|
+
|
777
|
+
else if (workspace.weights_arr.size())
|
778
|
+
{
|
779
|
+
for (size_t row = workspace.st; row <= workspace.end; row++)
|
780
|
+
if (input_data.has_missing[workspace.ix_arr[row]])
|
781
|
+
add_from_impute_node(imputer,
|
782
|
+
workspace.impute_map[workspace.ix_arr[row]],
|
783
|
+
workspace.weights_arr[workspace.ix_arr[row]]);
|
784
|
+
}
|
785
|
+
|
786
|
+
else
|
787
|
+
{
|
788
|
+
for (size_t row = workspace.st; row <= workspace.end; row++)
|
789
|
+
if (input_data.has_missing[workspace.ix_arr[row]])
|
790
|
+
add_from_impute_node(imputer,
|
791
|
+
workspace.impute_map[workspace.ix_arr[row]],
|
792
|
+
workspace.weights_map[workspace.ix_arr[row]]);
|
793
|
+
}
|
794
|
+
}
|
795
|
+
}
|
796
|
+
|
797
|
+
template <class imp_arr>
|
798
|
+
void apply_imputation_results(imp_arr &impute_vec,
|
799
|
+
Imputer &imputer,
|
800
|
+
InputData &input_data,
|
801
|
+
int nthreads)
|
802
|
+
{
|
803
|
+
size_t col;
|
804
|
+
|
805
|
+
if (input_data.Xc != NULL)
|
806
|
+
{
|
807
|
+
std::vector<size_t> row_pos(input_data.nrows, 0);
|
808
|
+
size_t row;
|
809
|
+
|
810
|
+
for (size_t col = 0; col < input_data.ncols_numeric; col++)
|
811
|
+
{
|
812
|
+
for (sparse_ix ix = input_data.Xc_indptr[col]; ix < input_data.Xc_indptr[col + 1]; ix++)
|
813
|
+
{
|
814
|
+
if (is_na_or_inf(input_data.Xc[ix]))
|
815
|
+
{
|
816
|
+
row = input_data.Xc_ind[ix];
|
817
|
+
if (impute_vec[row].sp_num_weight[row_pos[row]] > 0 && !is_na_or_inf(impute_vec[row].sp_num_sum[row_pos[row]]))
|
818
|
+
input_data.Xc[ix]
|
819
|
+
=
|
820
|
+
impute_vec[row].sp_num_sum[row_pos[row]]
|
821
|
+
/
|
822
|
+
impute_vec[row].sp_num_weight[row_pos[row]];
|
823
|
+
else
|
824
|
+
input_data.Xc[ix]
|
825
|
+
=
|
826
|
+
imputer.col_means[col];
|
827
|
+
|
828
|
+
row_pos[row]++;
|
829
|
+
}
|
830
|
+
}
|
831
|
+
}
|
832
|
+
}
|
833
|
+
|
834
|
+
#pragma omp parallel for schedule(dynamic) num_threads(nthreads) shared(input_data, impute_vec, imputer) private(col)
|
835
|
+
for (size_t_for row = 0; row < input_data.nrows; row++)
|
836
|
+
{
|
837
|
+
if (input_data.has_missing[row])
|
838
|
+
{
|
839
|
+
for (size_t ix = 0; ix < impute_vec[row].n_missing_num; ix++)
|
840
|
+
{
|
841
|
+
col = impute_vec[row].missing_num[ix];
|
842
|
+
if (impute_vec[row].num_weight[ix] > 0 && !is_na_or_inf(impute_vec[row].num_sum[ix]))
|
843
|
+
input_data.numeric_data[row + col * input_data.nrows]
|
844
|
+
=
|
845
|
+
impute_vec[row].num_sum[ix] / impute_vec[row].num_weight[ix];
|
846
|
+
else
|
847
|
+
input_data.numeric_data[row + col * input_data.nrows]
|
848
|
+
=
|
849
|
+
imputer.col_means[col];
|
850
|
+
}
|
851
|
+
|
852
|
+
for (size_t ix = 0; ix < impute_vec[row].n_missing_cat; ix++)
|
853
|
+
{
|
854
|
+
col = impute_vec[row].missing_cat[ix];
|
855
|
+
input_data.categ_data[row + col * input_data.nrows]
|
856
|
+
=
|
857
|
+
std::distance(impute_vec[row].cat_sum[col].begin(),
|
858
|
+
std::max_element(impute_vec[row].cat_sum[col].begin(),
|
859
|
+
impute_vec[row].cat_sum[col].end()));
|
860
|
+
|
861
|
+
if (input_data.categ_data[row + col * input_data.nrows] == 0 && impute_vec[row].cat_sum[col][0] <= 0)
|
862
|
+
input_data.categ_data[row + col * input_data.nrows]
|
863
|
+
=
|
864
|
+
imputer.col_modes[col];
|
865
|
+
}
|
866
|
+
}
|
867
|
+
}
|
868
|
+
}
|
869
|
+
|
870
|
+
void apply_imputation_results(std::vector<ImputedData> &impute_vec,
|
871
|
+
std::unordered_map<size_t, ImputedData> &impute_map,
|
872
|
+
Imputer &imputer,
|
873
|
+
InputData &input_data,
|
874
|
+
int nthreads)
|
875
|
+
{
|
876
|
+
if (impute_vec.size())
|
877
|
+
apply_imputation_results(impute_vec, imputer, input_data, nthreads);
|
878
|
+
else if (impute_map.size())
|
879
|
+
apply_imputation_results(impute_map, imputer, input_data, nthreads);
|
880
|
+
}
|
881
|
+
|
882
|
+
|
883
|
+
void apply_imputation_results(PredictionData &prediction_data,
|
884
|
+
ImputedData &imp,
|
885
|
+
Imputer &imputer,
|
886
|
+
size_t row)
|
887
|
+
{
|
888
|
+
size_t col;
|
889
|
+
size_t pos = 0;
|
890
|
+
for (size_t ix = 0; ix < imp.n_missing_num; ix++)
|
891
|
+
{
|
892
|
+
col = imp.missing_num[ix];
|
893
|
+
if (imp.num_weight[ix] > 0 && !is_na_or_inf(imp.num_sum[ix]))
|
894
|
+
prediction_data.numeric_data[row + col * prediction_data.nrows]
|
895
|
+
=
|
896
|
+
imp.num_sum[ix] / imp.num_weight[ix];
|
897
|
+
else
|
898
|
+
prediction_data.numeric_data[row + col * prediction_data.nrows]
|
899
|
+
=
|
900
|
+
imputer.col_means[col];
|
901
|
+
}
|
902
|
+
|
903
|
+
if (prediction_data.Xr != NULL)
|
904
|
+
for (size_t ix = prediction_data.Xr_indptr[row]; ix < prediction_data.Xr_indptr[row + 1]; ix++)
|
905
|
+
{
|
906
|
+
if (is_na_or_inf(prediction_data.Xr[ix]))
|
907
|
+
{
|
908
|
+
if (imp.sp_num_weight[pos] > 0 && !is_na_or_inf(imp.sp_num_sum[pos]))
|
909
|
+
prediction_data.Xr[ix]
|
910
|
+
=
|
911
|
+
imp.sp_num_sum[pos] / imp.sp_num_weight[pos];
|
912
|
+
else
|
913
|
+
prediction_data.Xr[ix]
|
914
|
+
=
|
915
|
+
imputer.col_means[imp.missing_sp[pos]];
|
916
|
+
pos++;
|
917
|
+
}
|
918
|
+
}
|
919
|
+
|
920
|
+
for (size_t ix = 0; ix < imp.n_missing_cat; ix++)
|
921
|
+
{
|
922
|
+
col = imp.missing_cat[ix];
|
923
|
+
prediction_data.categ_data[row + col * prediction_data.nrows]
|
924
|
+
=
|
925
|
+
std::distance(imp.cat_sum[col].begin(),
|
926
|
+
std::max_element(imp.cat_sum[col].begin(), imp.cat_sum[col].end()));
|
927
|
+
|
928
|
+
if (prediction_data.categ_data[row + col * prediction_data.nrows] == 0 && imp.cat_sum[col][0] <= 0)
|
929
|
+
prediction_data.categ_data[row + col * prediction_data.nrows]
|
930
|
+
=
|
931
|
+
imputer.col_modes[col];
|
932
|
+
}
|
933
|
+
}
|
934
|
+
|
935
|
+
|
936
|
+
void initialize_impute_calc(ImputedData &imp, InputData &input_data, size_t row)
|
937
|
+
{
|
938
|
+
imp.n_missing_num = 0;
|
939
|
+
imp.n_missing_cat = 0;
|
940
|
+
imp.n_missing_sp = 0;
|
941
|
+
|
942
|
+
if (input_data.numeric_data != NULL)
|
943
|
+
{
|
944
|
+
imp.missing_num.resize(input_data.ncols_numeric);
|
945
|
+
for (size_t col = 0; col < input_data.ncols_numeric; col++)
|
946
|
+
if (is_na_or_inf(input_data.numeric_data[row + col * input_data.nrows]))
|
947
|
+
imp.missing_num[imp.n_missing_num++] = col;
|
948
|
+
imp.missing_num.resize(imp.n_missing_num);
|
949
|
+
imp.num_sum.assign(imp.n_missing_num, 0);
|
950
|
+
imp.num_weight.assign(imp.n_missing_num, 0);
|
951
|
+
}
|
952
|
+
|
953
|
+
else if (input_data.Xc != NULL)
|
954
|
+
{
|
955
|
+
imp.missing_sp.resize(input_data.ncols_numeric);
|
956
|
+
sparse_ix *res;
|
957
|
+
for (size_t col = 0; col < input_data.ncols_numeric; col++)
|
958
|
+
{
|
959
|
+
res = std::lower_bound(input_data.Xc_ind + input_data.Xc_indptr[col],
|
960
|
+
input_data.Xc_ind + input_data.Xc_indptr[col + 1],
|
961
|
+
(sparse_ix) row);
|
962
|
+
if (
|
963
|
+
res != input_data.Xc_ind + input_data.Xc_indptr[col + 1] &&
|
964
|
+
*res == row &&
|
965
|
+
is_na_or_inf(input_data.Xc[res - input_data.Xc_ind])
|
966
|
+
)
|
967
|
+
{
|
968
|
+
imp.missing_sp[imp.n_missing_sp++] = col;
|
969
|
+
}
|
970
|
+
}
|
971
|
+
imp.sp_num_sum.assign(imp.n_missing_sp, 0);
|
972
|
+
imp.sp_num_weight.assign(imp.n_missing_sp, 0);
|
973
|
+
}
|
974
|
+
|
975
|
+
if (input_data.categ_data != NULL)
|
976
|
+
{
|
977
|
+
imp.missing_cat.resize(input_data.ncols_categ);
|
978
|
+
for (size_t col = 0; col < input_data.ncols_categ; col++)
|
979
|
+
if (input_data.categ_data[row + col * input_data.nrows] < 0)
|
980
|
+
imp.missing_cat[imp.n_missing_cat++] = col;
|
981
|
+
imp.missing_cat.resize(imp.n_missing_cat);
|
982
|
+
imp.cat_weight.assign(imp.n_missing_cat, 0);
|
983
|
+
imp.cat_sum.resize(input_data.ncols_categ);
|
984
|
+
for (size_t cat = 0; cat < imp.n_missing_cat; cat++)
|
985
|
+
imp.cat_sum[imp.missing_cat[cat]].assign(input_data.ncat[imp.missing_cat[cat]], 0);
|
986
|
+
}
|
987
|
+
}
|
988
|
+
|
989
|
+
void initialize_impute_calc(ImputedData &imp, PredictionData &prediction_data, Imputer &imputer, size_t row)
|
990
|
+
{
|
991
|
+
imp.n_missing_num = 0;
|
992
|
+
imp.n_missing_cat = 0;
|
993
|
+
imp.n_missing_sp = 0;
|
994
|
+
|
995
|
+
if (prediction_data.numeric_data != NULL)
|
996
|
+
{
|
997
|
+
if (!imp.missing_num.size())
|
998
|
+
imp.missing_num.resize(imputer.ncols_numeric);
|
999
|
+
for (size_t col = 0; col < imputer.ncols_numeric; col++)
|
1000
|
+
if (is_na_or_inf(prediction_data.numeric_data[row + col * prediction_data.nrows]))
|
1001
|
+
imp.missing_num[imp.n_missing_num++] = col;
|
1002
|
+
|
1003
|
+
if (!imp.num_sum.size())
|
1004
|
+
{
|
1005
|
+
imp.num_sum.resize(imputer.ncols_numeric, 0);
|
1006
|
+
imp.num_weight.resize(imputer.ncols_numeric, 0);
|
1007
|
+
}
|
1008
|
+
|
1009
|
+
else
|
1010
|
+
{
|
1011
|
+
std::fill(imp.num_sum.begin(), imp.num_sum.begin() + imp.n_missing_num, 0);
|
1012
|
+
std::fill(imp.num_weight.begin(), imp.num_weight.begin() + imp.n_missing_num, 0);
|
1013
|
+
}
|
1014
|
+
}
|
1015
|
+
|
1016
|
+
else if (prediction_data.Xr != NULL)
|
1017
|
+
{
|
1018
|
+
if (!imp.missing_sp.size())
|
1019
|
+
imp.missing_sp.resize(imputer.ncols_numeric);
|
1020
|
+
for (size_t ix = prediction_data.Xr_indptr[row]; ix < prediction_data.Xr_indptr[row + 1]; ix++)
|
1021
|
+
if (is_na_or_inf(prediction_data.Xr[ix]))
|
1022
|
+
imp.missing_sp[imp.n_missing_sp++] = prediction_data.Xr_ind[ix];
|
1023
|
+
|
1024
|
+
if (!imp.sp_num_sum.size())
|
1025
|
+
{
|
1026
|
+
imp.sp_num_sum.resize(imputer.ncols_numeric, 0);
|
1027
|
+
imp.sp_num_weight.resize(imputer.ncols_numeric, 0);
|
1028
|
+
}
|
1029
|
+
|
1030
|
+
else
|
1031
|
+
{
|
1032
|
+
std::fill(imp.sp_num_sum.begin(), imp.sp_num_sum.begin() + imp.n_missing_sp, 0);
|
1033
|
+
std::fill(imp.sp_num_weight.begin(), imp.sp_num_weight.begin() + imp.n_missing_sp, 0);
|
1034
|
+
}
|
1035
|
+
}
|
1036
|
+
|
1037
|
+
if (prediction_data.categ_data != NULL)
|
1038
|
+
{
|
1039
|
+
if (!imp.missing_cat.size())
|
1040
|
+
imp.missing_cat.resize(imputer.ncols_categ);
|
1041
|
+
for (size_t col = 0; col < imputer.ncols_categ; col++)
|
1042
|
+
{
|
1043
|
+
if (prediction_data.categ_data[row + col * prediction_data.nrows] < 0)
|
1044
|
+
imp.missing_cat[imp.n_missing_cat++] = col;
|
1045
|
+
}
|
1046
|
+
|
1047
|
+
if (!imp.cat_weight.size())
|
1048
|
+
{
|
1049
|
+
imp.cat_weight.resize(imputer.ncols_categ, 0);
|
1050
|
+
imp.cat_sum.resize(imputer.ncols_categ);
|
1051
|
+
for (size_t col = 0; col < imputer.ncols_categ; col++)
|
1052
|
+
imp.cat_sum[col].resize(imputer.ncat[col], 0);
|
1053
|
+
}
|
1054
|
+
|
1055
|
+
else
|
1056
|
+
{
|
1057
|
+
std::fill(imp.cat_weight.begin(), imp.cat_weight.begin() + imp.n_missing_cat, 0);
|
1058
|
+
for (size_t col = 0; col < imp.n_missing_cat; col++)
|
1059
|
+
std::fill(imp.cat_sum[imp.missing_cat[col]].begin(),
|
1060
|
+
imp.cat_sum[imp.missing_cat[col]].end(),
|
1061
|
+
0);
|
1062
|
+
}
|
1063
|
+
}
|
1064
|
+
}
|
1065
|
+
|
1066
|
+
ImputedData::ImputedData(InputData &input_data, size_t row)
|
1067
|
+
{
|
1068
|
+
initialize_impute_calc(*this, input_data, row);
|
1069
|
+
}
|
1070
|
+
|
1071
|
+
void allocate_imp_vec(std::vector<ImputedData> &impute_vec, InputData &input_data, int nthreads)
|
1072
|
+
{
|
1073
|
+
impute_vec.resize(input_data.nrows);
|
1074
|
+
#pragma omp parallel for schedule(dynamic) num_threads(nthreads) shared(impute_vec, input_data)
|
1075
|
+
for (size_t_for row = 0; row < input_data.nrows; row++)
|
1076
|
+
if (input_data.has_missing[row])
|
1077
|
+
initialize_impute_calc(impute_vec[row], input_data, row);
|
1078
|
+
}
|
1079
|
+
|
1080
|
+
|
1081
|
+
void allocate_imp_map(std::unordered_map<size_t, ImputedData> &impute_map, InputData &input_data)
|
1082
|
+
{
|
1083
|
+
for (size_t row = 0; row < input_data.nrows; row++)
|
1084
|
+
if (input_data.has_missing[row])
|
1085
|
+
impute_map[row] = ImputedData(input_data, row);
|
1086
|
+
}
|
1087
|
+
|
1088
|
+
void allocate_imp(InputData &input_data,
|
1089
|
+
std::vector<ImputedData> &impute_vec,
|
1090
|
+
std::unordered_map<size_t, ImputedData> &impute_map,
|
1091
|
+
int nthreads)
|
1092
|
+
{
|
1093
|
+
if (input_data.n_missing == 0)
|
1094
|
+
return;
|
1095
|
+
else if (input_data.n_missing <= input_data.nrows / (nthreads * 10))
|
1096
|
+
allocate_imp_map(impute_map, input_data);
|
1097
|
+
else
|
1098
|
+
allocate_imp_vec(impute_vec, input_data, nthreads);
|
1099
|
+
}
|
1100
|
+
|
1101
|
+
void check_for_missing(InputData &input_data,
|
1102
|
+
std::vector<ImputedData> &impute_vec,
|
1103
|
+
std::unordered_map<size_t, ImputedData> &impute_map,
|
1104
|
+
int nthreads)
|
1105
|
+
{
|
1106
|
+
input_data.has_missing.assign(input_data.nrows, false);
|
1107
|
+
|
1108
|
+
if (input_data.Xc != NULL)
|
1109
|
+
{
|
1110
|
+
for (size_t col = 0; col < input_data.ncols_numeric; col++)
|
1111
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) shared(col, input_data)
|
1112
|
+
for (size_t_for ix = input_data.Xc_indptr[col]; ix < input_data.Xc_indptr[col + 1]; ix++)
|
1113
|
+
if (is_na_or_inf(input_data.Xc[ix]))
|
1114
|
+
input_data.has_missing[input_data.Xc_ind[ix]] = true;
|
1115
|
+
#pragma omp barrier
|
1116
|
+
}
|
1117
|
+
|
1118
|
+
if (input_data.numeric_data != NULL || input_data.categ_data != NULL)
|
1119
|
+
{
|
1120
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) shared(input_data)
|
1121
|
+
for (size_t_for row = 0; row < input_data.nrows; row++)
|
1122
|
+
{
|
1123
|
+
for (size_t col = 0; col < input_data.ncols_numeric; col++)
|
1124
|
+
{
|
1125
|
+
if (is_na_or_inf(input_data.numeric_data[row + col * input_data.nrows]))
|
1126
|
+
{
|
1127
|
+
input_data.has_missing[row] = true;
|
1128
|
+
break;
|
1129
|
+
}
|
1130
|
+
}
|
1131
|
+
|
1132
|
+
if (!input_data.has_missing[row])
|
1133
|
+
for (size_t col = 0; col < input_data.ncols_categ; col++)
|
1134
|
+
{
|
1135
|
+
if (input_data.categ_data[row + col * input_data.nrows] < 0)
|
1136
|
+
{
|
1137
|
+
input_data.has_missing[row] = true;
|
1138
|
+
break;
|
1139
|
+
}
|
1140
|
+
}
|
1141
|
+
}
|
1142
|
+
}
|
1143
|
+
|
1144
|
+
input_data.n_missing = std::accumulate(input_data.has_missing.begin(), input_data.has_missing.end(), (size_t)0);
|
1145
|
+
allocate_imp(input_data, impute_vec, impute_map, nthreads);
|
1146
|
+
}
|
1147
|
+
|
1148
|
+
size_t check_for_missing(PredictionData &prediction_data,
|
1149
|
+
Imputer &imputer,
|
1150
|
+
size_t ix_arr[],
|
1151
|
+
int nthreads)
|
1152
|
+
{
|
1153
|
+
std::vector<char> has_missing(prediction_data.nrows, false);
|
1154
|
+
|
1155
|
+
#pragma omp parallel for schedule(static) num_threads(nthreads) shared(has_missing, prediction_data, imputer)
|
1156
|
+
for (size_t_for row = 0; row < prediction_data.nrows; row++)
|
1157
|
+
{
|
1158
|
+
if (prediction_data.numeric_data != NULL)
|
1159
|
+
for (size_t col = 0; col < imputer.ncols_numeric; col++)
|
1160
|
+
{
|
1161
|
+
if (is_na_or_inf(prediction_data.numeric_data[row + col * prediction_data.nrows]))
|
1162
|
+
{
|
1163
|
+
has_missing[row] = true;
|
1164
|
+
break;
|
1165
|
+
}
|
1166
|
+
}
|
1167
|
+
else if (prediction_data.Xr != NULL)
|
1168
|
+
for (size_t ix = prediction_data.Xr_indptr[row]; ix < prediction_data.Xr_indptr[row + 1]; ix++)
|
1169
|
+
{
|
1170
|
+
if (is_na_or_inf(prediction_data.Xr[ix]))
|
1171
|
+
{
|
1172
|
+
has_missing[row] = true;
|
1173
|
+
break;
|
1174
|
+
}
|
1175
|
+
}
|
1176
|
+
|
1177
|
+
if (!has_missing[row])
|
1178
|
+
for (size_t col = 0; col < imputer.ncols_categ; col++)
|
1179
|
+
{
|
1180
|
+
if (prediction_data.categ_data[row + col * prediction_data.nrows] < 0)
|
1181
|
+
{
|
1182
|
+
has_missing[row] = true;
|
1183
|
+
break;
|
1184
|
+
}
|
1185
|
+
}
|
1186
|
+
}
|
1187
|
+
|
1188
|
+
size_t st = 0;
|
1189
|
+
size_t temp;
|
1190
|
+
for (size_t row = 0; row < prediction_data.nrows; row++)
|
1191
|
+
{
|
1192
|
+
if (has_missing[row])
|
1193
|
+
{
|
1194
|
+
temp = ix_arr[st];
|
1195
|
+
ix_arr[st] = ix_arr[row];
|
1196
|
+
ix_arr[row] = temp;
|
1197
|
+
st++;
|
1198
|
+
}
|
1199
|
+
}
|
1200
|
+
|
1201
|
+
if (st == 0)
|
1202
|
+
return 0;
|
1203
|
+
|
1204
|
+
return st;
|
1205
|
+
}
|