pytofi 0.0.1__py3-none-any.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -0,0 +1,51 @@
1
+ Metadata-Version: 2.4
2
+ Name: pytofi
3
+ Version: 0.0.1
4
+ Summary: Tests on Feature Importance
5
+ Author-email: Po-Hsien Huang <psyphh@nccu.edu.tw>
6
+ Classifier: Programming Language :: Python :: 3
7
+ Classifier: License :: OSI Approved :: MIT License
8
+ Classifier: Operating System :: OS Independent
9
+ Requires-Python: >=3.9
10
+ Description-Content-Type: text/markdown
11
+ License-File: LICENSE
12
+ Requires-Dist: scikit-learn>=1.5.0
13
+ Requires-Dist: pandas>=2.0.3
14
+ Dynamic: license-file
15
+
16
+ # TOFI: Tests on Feature Importance
17
+
18
+
19
+ ## Overview
20
+
21
+ `tofi` is a python package for conducting tests on feature importance in machine learning. It is written on the basis of [`scikit-learn`](https://scikit-learn.org/stable/index.html). Hence, in principle, any regressor/classifier learned by `scikit-learn` can be inferred by `tofi`.
22
+
23
+ `tofi` provides two types of `Inferer` for making inferences: `CIT` for conditional independence test and `RIT` for risk invariance test. `CIT` can implement the holdout randomization test (HRT; Tansey et al., 2022), residual permutation test (RPT; Huang, 2024), and conditional predictive impact (CPI; Watson & Wright, 2021). On the other hand, `RIT` can conduct leave-one-covariate-out (LOCO; Lei et al., 2018) and plug-in estimation (PIE; Williamson et al., 2023). In addition, `tofi` also constructs a `Crosser` class. With the help of `Crosser`, cross-fitting and $p$-value combination can be easily implemented to improve the stability and even statistical power of the CITs and RITs.
24
+
25
+ ## Installation
26
+ `tofi` can be installed via `pip`:
27
+ ```
28
+ pip install pytofi
29
+ ```
30
+
31
+ ## Tutorial
32
+
33
+ Three Jupyter notebooks are available to learn the use of `sial`:
34
+
35
+ 1. [Conditional Independence Tests](https://github.com/aipsylab/tofi/blob/main/tutorial/tutorial_1_cit.ipynb)
36
+ 2. [Risk Invariance Tests](https://github.com/aipsylab/tofi/blob/main/tutorial/tutorial_2_rit.ipynb)
37
+ 3. [Cross-Fitting and $p$-Value Combination](https://github.com/aipsylab/tofi/blob/main/tutorial/tutorial_3_crosser.ipynb)
38
+
39
+ These notebooks assume that readers have experiences in using `scikit-learn`.
40
+
41
+ ## References
42
+
43
+ Huang, P.-H. (2024). Residual Permutation Tests for Feature Importance in Machine Learning. [Manuscript submitted for publication].
44
+
45
+ Lei, J., G’Sell, M., Rinaldo, A., Tibshirani, R. J., & Wasserman, L. (2018). Distribution-free predictive inference for regression. Journal of the American Statistical Association, 113(523), 1094–1111. doi: 10.1080/01621459.2017.1307116
46
+
47
+ Tansey, W., Veitch, V., Zhang, H., Rabadan, R., & Blei, D. M. (2022). The holdout randomization test for feature selection in black box models. Journal of Computational and Graphical Statistics, 31(1), 151–162. doi: 10.1080/10618600.2021.1923520
48
+
49
+ Watson, D., & Wright, M. (2021). Testing conditional independence in supervised learning algorithms. Machine Learning, 110, 1-23. doi: 10.1007/s10994-021-06030-6
50
+
51
+ Williamson, B. D., Gilbert, P. B., Simon, N. R., & Carone, M. (2023). A general framework for inference on algorithm-agnostic variable importance. Journal of the American Statistical Association, 118(543), 1645–1658. doi: 10.1080/01621459.2021.2003200
@@ -0,0 +1,9 @@
1
+ pytofi-0.0.1.dist-info/licenses/LICENSE,sha256=rTKytVFyxeiVubTaX35pqcq2whSZYxCD5mOelzBflME,1065
2
+ tofi/__init__.py,sha256=-5hbxVx7RtokIXa3BsPje_80GXjhpW-_DQKQb5ccLBU,118
3
+ tofi/crosser.py,sha256=eD2xLHGRnsZrKF0LR0WH2PvNHLawobvsT4zjz9usGVQ,15459
4
+ tofi/inferer.py,sha256=nUHPsTmxrmBd_FGKB88i39o_v47105Hr1s5CeBrHjOM,37128
5
+ tofi/utils.py,sha256=A4mg6vu-VWUqrfEYsjSfZruM7tRwNYB2mzs1ZQV3Rvw,1453
6
+ pytofi-0.0.1.dist-info/METADATA,sha256=4XdZOpZXFrYo5BbbdSFKdzPQcYv-jTPp3UXz3_lHYEA,3111
7
+ pytofi-0.0.1.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
8
+ pytofi-0.0.1.dist-info/top_level.txt,sha256=8s92NUTG5IVmFjb8EXdb0TCiLrPPflcgpwq7dr58_pI,5
9
+ pytofi-0.0.1.dist-info/RECORD,,
@@ -0,0 +1,5 @@
1
+ Wheel-Version: 1.0
2
+ Generator: setuptools (80.9.0)
3
+ Root-Is-Purelib: true
4
+ Tag: py3-none-any
5
+
@@ -0,0 +1,21 @@
1
+ MIT License
2
+
3
+ Copyright (c) 2025 aipsylab
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
@@ -0,0 +1 @@
1
+ tofi
tofi/__init__.py ADDED
@@ -0,0 +1,6 @@
1
+ from .crosser import Crosser
2
+ from .inferer import CIT, RIT
3
+
4
+ __all__ = ["Crosser",
5
+ "CIT",
6
+ "RIT"]
tofi/crosser.py ADDED
@@ -0,0 +1,490 @@
1
+ import math
2
+ import scipy
3
+ import warnings
4
+ import numpy as np
5
+ import pandas as pd
6
+ from sklearn.base import clone
7
+ from sklearn import model_selection
8
+ from sklearn.base import is_classifier
9
+ from sklearn.utils.validation import check_is_fitted
10
+ from sklearn.preprocessing import LabelBinarizer
11
+ from sklearn.metrics import get_scorer, get_scorer_names
12
+ from .utils import _sample
13
+
14
+ class Crosser():
15
+ """
16
+ An object for repeated cross-fitting
17
+
18
+ Parameters
19
+ ----------
20
+ estimator : estimator object
21
+ The base estimator for repeated cross-fitting.
22
+
23
+ cv : cross-validation generator
24
+ The cross-validation generator for repeated cross-fitting.
25
+ It must be a class member of `KFold`, `RepeatedKFold`,
26
+ `ShuffleSplit`, `StratifiedShuffleSplit`.
27
+
28
+ scoring : str or callable, default=None
29
+ Strategy to evaluate the performance of the cross-validated
30
+ model when using `summarize`. To see availble scoring methods
31
+ via `str`, use `sklearn.get_scorer_names()`.
32
+
33
+ - For regression tasks, `None` means "r2".
34
+ - For classification tasks, `None` means "accuracy".
35
+
36
+ """
37
+
38
+ def __init__(
39
+ self,
40
+ estimator,
41
+ cv,
42
+ scoring = None
43
+ ):
44
+ self.estimator = estimator
45
+ self.cv = cv
46
+ self.scoring = scoring
47
+
48
+ self._check()
49
+ self._setup()
50
+
51
+ def fit(
52
+ self,
53
+ X,
54
+ y
55
+ ):
56
+ """
57
+ Fit by repeated cross-fitting
58
+
59
+ Parameters
60
+ ----------
61
+
62
+ X : array-like of shape (n_samples, n_features)
63
+ Training array, where `n_samples` is the sample size and
64
+ `n_features` is the number of features.
65
+
66
+ y : array-like of shape (n_samples,)
67
+ Target relative to X for classification or regression.
68
+
69
+ Returns
70
+ -------
71
+ self : object
72
+ Instance of fitted crosser.
73
+ """
74
+
75
+ cv = self.cv
76
+ estimators = []
77
+ features = self._features(X, y, on = "train")
78
+ targets = self._targets(X, y, on = "train")
79
+ for feature, target in zip(features, targets):
80
+ estimator = clone(self.estimator)
81
+ _ = estimator.fit(feature, target)
82
+ estimators.append(estimator)
83
+ self.estimators_ = estimators
84
+
85
+ if is_classifier(self):
86
+ label_binarizer = LabelBinarizer()
87
+ _ = label_binarizer.fit(y)
88
+ self.label_binarizer_ = label_binarizer
89
+ self.classes_ = sorted(set(y))
90
+ self._is_fitted = True
91
+
92
+ val_scores, train_scores, test_scores = self._scores(X, y)
93
+ self.val_scores_ = val_scores
94
+ self.train_scores_ = train_scores
95
+ self.test_scores_ = test_scores
96
+
97
+ return self
98
+
99
+
100
+ def summarize(
101
+ self,
102
+ cross_fit = None,
103
+ combine = None,
104
+ reverse = False,
105
+ verbose = True
106
+ ):
107
+ """
108
+ Summarize repeated cross-fitting results
109
+
110
+ Parameters
111
+ ----------
112
+
113
+ cross_fit : bool, default=None
114
+ If `True`, score values will be averaged across folds.
115
+ It is applicable if `n_folds` > 1. If applicable, `None`
116
+ means `True`; Otherwise, `None` means `False`.
117
+
118
+ combine : bool, default=None
119
+ If `True`, score values will be averaged across folds
120
+ and repeats. It is applicable if `n_splits` > 1. If
121
+ `n_repeats` > 1, `None` means `True`; Otherwise, `None`
122
+ means `False`.
123
+
124
+ reverse : bool, default=True
125
+ If `True`, negative score values will be reported.
126
+ Note that by default a larger score value means better
127
+ in `scikit-learn`.
128
+
129
+ verbose : bool, default=True
130
+ Controls the verbosity.
131
+
132
+ Returns
133
+ -------
134
+ summary : DataFrame
135
+ A summary for validation, train, and test scores (`val_score`,
136
+ `train_score`, and `test_score`). The validation scores
137
+ are only presented if `estimator.best_score_` is available.
138
+
139
+ """
140
+ check_is_fitted(self)
141
+ estimator = self.estimator
142
+ cv = self.cv
143
+ scoring = self.scoring
144
+ n_splits = self._n_splits
145
+ n_repeats = self._n_repeats
146
+ n_folds = self._n_folds
147
+
148
+ val_scores = self.val_scores_
149
+ train_scores = self.train_scores_
150
+ test_scores = self.test_scores_
151
+
152
+ if combine is None:
153
+ if n_repeats > 1:
154
+ combine = True
155
+ else:
156
+ combine = False
157
+
158
+ if cross_fit is None:
159
+ if n_folds > 1:
160
+ cross_fit = True
161
+ else:
162
+ cross_fit = False
163
+
164
+ if reverse is True:
165
+ val_scores, train_scores, test_scores = -val_scores, -train_scores, -test_scores
166
+
167
+ if cross_fit:
168
+ val_scores = val_scores.reshape((n_repeats, n_folds)).mean(axis = 1)
169
+ train_scores = train_scores.reshape((n_repeats, n_folds)).mean(axis = 1)
170
+ test_scores = test_scores.reshape((n_repeats, n_folds)).mean(axis = 1)
171
+
172
+ if not combine:
173
+ if cross_fit:
174
+ index = pd.Index(
175
+ range(n_repeats),
176
+ name = "repeat")
177
+ else:
178
+ splits = list(range(self._n_splits))
179
+ repeats = [split // n_folds for split in splits]
180
+ folds = [split % n_folds for split in splits]
181
+ index = pd.MultiIndex.from_tuples(
182
+ list(zip(splits, repeats, folds)),
183
+ names = ["split", "repeat", "fold"])
184
+ summary = pd.DataFrame(
185
+ {"val_score": val_scores,
186
+ "train_score": train_scores,
187
+ "test_score": test_scores},
188
+ index = index
189
+ )
190
+ else:
191
+ index = pd.Index(
192
+ ["mean", "std"],
193
+ name = "method")
194
+ summary = pd.DataFrame(
195
+ {"val_score": [np.mean(val_scores), np.std(val_scores)],
196
+ "train_score": [np.mean(train_scores), np.std(train_scores)],
197
+ "test_score": [np.mean(test_scores), np.std(test_scores)]},
198
+ index = index
199
+ )
200
+
201
+ if verbose:
202
+ print("Crosser Summary", end = " ")
203
+ print("(cross_fit=", cross_fit, ", " ,
204
+ "combine=", combine, ")", sep = "")
205
+ print(" + Estimator: ",
206
+ estimator.__class__.__name__, sep = "")
207
+ print(" + Cross-Validator: ",
208
+ cv.__class__.__name__,
209
+ " (", "n_folds=", n_folds,
210
+ ", n_repeats=", n_repeats, ")", sep = "")
211
+ if callable(scoring):
212
+ print(" + Train/Test Scorer: ",
213
+ "Callable",
214
+ " (", "reverse=", reverse, ")", sep = "")
215
+ else:
216
+ print(" + Train/Test Scorer: ",
217
+ scoring.replace("_", " ").title(),
218
+ " (", "reverse=", reverse, ")", sep = "")
219
+ return summary
220
+
221
+
222
+ def predict(
223
+ self,
224
+ X,
225
+ *,
226
+ split = None
227
+ ):
228
+ check_is_fitted(self)
229
+ if split is None:
230
+ preds = np.array(
231
+ [estimator.predict(X)
232
+ for estimator in self.estimators_])
233
+ if is_classifier(self):
234
+ pred = np.apply_along_axis(
235
+ lambda x: np.argmax(np.bincount(x)),
236
+ axis = 0,
237
+ arr = preds)
238
+ else:
239
+ pred = preds.mean(axis = 0)
240
+ else:
241
+ pred = self.estimators_[split].predict(X)
242
+ return pred
243
+
244
+
245
+ def predict_proba(
246
+ self,
247
+ X,
248
+ *,
249
+ split = None
250
+ ):
251
+ check_is_fitted(self)
252
+ if split is None:
253
+ preds = np.array(
254
+ [estimator.predict_proba(X)
255
+ for estimator in self.estimators_])
256
+ pred = preds.mean(axis = 0)
257
+ else:
258
+ pred = self.estimators_[split].predict_proba(X)
259
+ return pred
260
+
261
+
262
+ def predict_log_proba(
263
+ self,
264
+ X,
265
+ *,
266
+ split = None
267
+ ):
268
+ check_is_fitted(self)
269
+ if split is None:
270
+ preds = np.array(
271
+ [estimator.predict_log_proba(X)
272
+ for estimator in self.estimators_])
273
+ pred = preds.mean(axis = 0)
274
+ else:
275
+ pred = self.estimators_[split].predict_log_proba(X)
276
+ return pred
277
+
278
+
279
+ def decision_function(
280
+ self,
281
+ X,
282
+ *,
283
+ split = None
284
+ ):
285
+ check_is_fitted(self)
286
+ if split is None:
287
+ preds = np.array(
288
+ [estimator.decision_function(X)
289
+ for estimator in self.estimators_])
290
+ pred = preds.mean(axis = 0)
291
+ else:
292
+ pred = self.estimators_[split].decision_function(X)
293
+ return pred
294
+
295
+ def _scores(
296
+ self,
297
+ X,
298
+ y
299
+ ):
300
+ estimators = self.estimators_
301
+ scorer = self._scorer
302
+ n_repeats = self._n_repeats
303
+ n_folds = self._n_folds
304
+
305
+ features = zip(self._features(X, y, on = "train"),
306
+ self._features(X, y, on = "test"))
307
+ targets = zip(self._targets(X, y, on = "train"),
308
+ self._targets(X, y, on = "test"))
309
+
310
+ train_scores = []
311
+ val_scores = []
312
+ test_scores = []
313
+ for estimator, feature, target in zip(estimators, features, targets):
314
+ feature_train, feature_test = feature
315
+ target_train, target_test = target
316
+ train_score = scorer(estimator, feature_train, target_train)
317
+ test_score = scorer(estimator, feature_test, target_test)
318
+ if hasattr(estimator, "best_score_"):
319
+ val_score = estimator.best_score_
320
+ else:
321
+ val_score = np.nan
322
+ val_scores.append(val_score)
323
+ train_scores.append(train_score)
324
+ test_scores.append(test_score)
325
+ val_scores = np.array(val_scores)
326
+ train_scores = np.array(train_scores)
327
+ test_scores = np.array(test_scores)
328
+ return val_scores, train_scores, test_scores
329
+
330
+
331
+ def _features(
332
+ self,
333
+ X,
334
+ y,
335
+ on = "test"
336
+ ):
337
+ cv = self.cv
338
+ features = []
339
+ for train_index, test_index in cv.split(X, y):
340
+ if on == "test":
341
+ index = test_index
342
+ else:
343
+ index = train_index
344
+ if isinstance(X, pd.DataFrame):
345
+ feature = X.iloc[index, :]
346
+ else:
347
+ feature = X[index, :]
348
+ features.append(feature)
349
+ return features
350
+
351
+ def _targets(
352
+ self,
353
+ X,
354
+ y,
355
+ binarize = None,
356
+ on = "test"
357
+ ):
358
+ cv = self.cv
359
+ targets = []
360
+ for train_index, test_index in cv.split(X, y):
361
+ if on == "test":
362
+ index = test_index
363
+ else:
364
+ index = train_index
365
+ if isinstance(y, pd.Series):
366
+ target = y.iloc[index]
367
+ else:
368
+ target = y[index]
369
+ if is_classifier(self.estimator):
370
+ if binarize is True:
371
+ target = self.label_binarizer_.transform(target)
372
+ if target.shape[1] == 1:
373
+ target = np.append(1 - target, target, axis=1)
374
+ targets.append(target)
375
+ return targets
376
+
377
+
378
+ def _rvs(
379
+ self,
380
+ X,
381
+ y,
382
+ n_copies = None,
383
+ random_state = None
384
+ ):
385
+ estimators = self.estimators_
386
+ features = self._features(X, y)
387
+ targets = self._targets(X, y)
388
+ rvs = []
389
+ for estimator, feature, target in zip(estimators, features, targets):
390
+ if hasattr(estimator, "sample"):
391
+ rv = estimator.sample(
392
+ feature,
393
+ target,
394
+ n_copies = n_copies,
395
+ random_state = random_state)
396
+ else:
397
+ rv = _sample(
398
+ estimator,
399
+ feature,
400
+ target,
401
+ n_copies = n_copies,
402
+ random_state = random_state)
403
+ rvs.append(rv)
404
+ return rvs
405
+
406
+ def _check(
407
+ self
408
+ ):
409
+ cv = self.cv
410
+ scoring = self.scoring
411
+
412
+ allowed_cvs = {"KFold", "RepeatedKFold", "ShuffleSplit"}
413
+ if not isinstance(
414
+ cv,
415
+ tuple(getattr(model_selection, allowed_cv)
416
+ for allowed_cv in allowed_cvs)
417
+ ):
418
+ raise ValueError(
419
+ "Support `cv` types are {}.".format(allowed_cvs))
420
+ else:
421
+ if isinstance(
422
+ cv, getattr(model_selection, "KFold")):
423
+ if cv.shuffle:
424
+ if cv.random_state is None:
425
+ raise ValueError(
426
+ "When `cv` is `KFold` and `cv.shuffle` is `True`, "
427
+ "`cv.random_state` cannot be `None`.")
428
+ else:
429
+ if cv.random_state is None:
430
+ raise ValueError(
431
+ "When `cv` is `RepeatedKFold` or `ShuffleSplit`, "
432
+ "`cv.random_state` cannot be `None`.")
433
+
434
+ if scoring is not None:
435
+ if not (isinstance(scoring, str) or callable(scoring)):
436
+ raise ValueError(
437
+ "Support `scoring` types are `str`, `callable`, or `None`")
438
+ if isinstance(scoring, str):
439
+ scorer_names = get_scorer_names()
440
+ if not scoring in get_scorer_names():
441
+ raise ValueError(
442
+ "Support `scoring` names are {}.".format(scorer_names))
443
+
444
+
445
+ def _setup(
446
+ self
447
+ ):
448
+ estimator = self.estimator
449
+ cv = self.cv
450
+ scoring = self.scoring
451
+
452
+ kf_cvs = {"KFold", "RepeatedKFold"}
453
+ ss_cvs = {"ShuffleSplit"}
454
+ n_splits = cv.get_n_splits()
455
+ if isinstance(
456
+ cv,
457
+ tuple(getattr(model_selection, kf_cv)
458
+ for kf_cv in kf_cvs)):
459
+ if hasattr(cv, "n_repeats"):
460
+ n_repeats = cv.n_repeats
461
+ else:
462
+ n_repeats = 1
463
+ else:
464
+ n_repeats = cv.get_n_splits()
465
+ n_folds = n_splits // n_repeats
466
+
467
+ if scoring is None:
468
+ if is_classifier(estimator):
469
+ scoring = "accuracy"
470
+ else:
471
+ scoring = "r2"
472
+ if isinstance(scoring, str):
473
+ scorer = get_scorer(scoring)
474
+ elif callable(scoring):
475
+ scorer = scoring
476
+
477
+ self.scoring = scoring
478
+ self._is_fitted = False
479
+ self._estimator_type = estimator._estimator_type
480
+ self._n_splits = n_splits
481
+ self._n_repeats = n_repeats
482
+ self._n_folds = n_folds
483
+ self._scorer = scorer
484
+
485
+ def __sklearn_is_fitted__(self):
486
+ return hasattr(self, "_is_fitted") and self._is_fitted
487
+
488
+
489
+
490
+