@@ -312,6 +312,171 @@ def create_binary_newsgroups_data():
312312 return newsgroups_train , newsgroups_test , class_names
313313
314314
315+ def create_random_forest_vectorizer ():
316+ from sklearn .base import TransformerMixin
317+ from sklearn .ensemble import RandomForestClassifier
318+ from sklearn .feature_extraction .text import CountVectorizer
319+ from sklearn .pipeline import Pipeline
320+
321+ vectorizer = CountVectorizer (lowercase = False , min_df = 0.0 , binary = True )
322+
323+ class DenseTransformer (TransformerMixin ):
324+ def fit (self , X , y = None , ** fit_params ):
325+ return self
326+
327+ def transform (self , X , y = None , ** fit_params ):
328+ return X .toarray ()
329+
330+ rf = RandomForestClassifier (n_estimators = 500 , random_state = 777 )
331+ return Pipeline ([("vectorizer" , vectorizer ), ("to_dense" , DenseTransformer ()), ("rf" , rf )])
332+
333+
334+ def test_sklearn_random_forest_newsgroups ():
335+ import shap
336+ # from sklearn.ensemble import RandomForestClassifier
337+
338+ # note: this test used to fail in native TreeExplainer code due to memory corruption
339+ newsgroups_train , newsgroups_test , _ = create_binary_newsgroups_data ()
340+ pipeline = create_random_forest_vectorizer ()
341+ pipeline .fit (newsgroups_train .data , newsgroups_train .target )
342+ rf = pipeline .named_steps ["rf" ]
343+ vectorizer = pipeline .named_steps ["vectorizer" ]
344+ densifier = pipeline .named_steps ["to_dense" ]
345+
346+ dense_bg = densifier .transform (vectorizer .transform (newsgroups_test .data [0 :20 ]))
347+
348+ test_row = newsgroups_test .data [83 :84 ]
349+ explainer = shap .TreeExplainer (rf , dense_bg , feature_perturbation = "interventional" )
350+ vec_row = vectorizer .transform (test_row )
351+ dense_row = densifier .transform (vec_row )
352+ explainer .shap_values (dense_row )
353+
354+
355+ def test_sklearn_decision_tree_multiclass ():
356+ import numpy as np
357+ from sklearn .tree import DecisionTreeClassifier
358+
359+ import shap
360+
361+ X , y = shap .datasets .iris ()
362+ y [y == 2 ] = 1
363+ model = DecisionTreeClassifier (max_depth = None , min_samples_split = 2 , random_state = 0 )
364+ model .fit (X , y )
365+
366+ explainer = shap .TreeExplainer (model )
367+ shap_values = explainer .shap_values (X )
368+ assert np .abs (shap_values [0 ][0 , 0 ] - 0.05 ) < 1e-1
369+ assert np .abs (shap_values [1 ][0 , 0 ] + 0.05 ) < 1e-1
370+
371+
372+ def _common_lightgbm_regressor_test (create_model ):
373+ import numpy as np
374+
375+ import shap
376+
377+ # train lightgbm model on california housing price regression dataset
378+ X , y = shap .datasets .california ()
379+ model = create_model ()
380+ model .fit (X , y )
381+
382+ # explain the model's predictions using SHAP values
383+ ex = shap .TreeExplainer (model )
384+ shap_values = ex .shap_values (X )
385+
386+ predicted = model .predict (X , raw_score = True )
387+
388+ assert np .abs (shap_values .sum (1 ) + ex .expected_value - predicted ).max () < 1e-4 , (
389+ "SHAP values don't sum to model output!"
390+ )
391+
392+
393+ def test_lightgbm ():
394+ lightgbm = pytest .importorskip ("lightgbm" )
395+
396+ def create_model ():
397+ return lightgbm .sklearn .LGBMRegressor (categorical_feature = [8 ])
398+
399+ _common_lightgbm_regressor_test (create_model )
400+
401+
402+ def test_lightgbm_mse_regressor ():
403+ lightgbm = pytest .importorskip ("lightgbm" )
404+
405+ # train the lightgbm model on a dataset with MSE objective
406+ def create_model ():
407+ return lightgbm .sklearn .LGBMRegressor (categorical_feature = [8 ], objective = "mean_squared_error" )
408+
409+ _common_lightgbm_regressor_test (create_model )
410+
411+
412+ def _common_lightgbm_nonsklearn_api (dataset , params , validation ):
413+ import lightgbm
414+ from sklearn .model_selection import train_test_split
415+
416+ import shap
417+
418+ # train the lightgbm model using non-sklearn API with binary classification dataset
419+ X_train , X_test , y_train , y_test = train_test_split (* dataset , test_size = 0.2 , random_state = 0 )
420+ lgb_train = lightgbm .Dataset (X_train , y_train )
421+ lgb_test = lightgbm .Dataset (X_test , y_test , reference = lgb_train )
422+
423+ booster = lightgbm .train (params , lgb_train , valid_sets = [lgb_train , lgb_test ])
424+ # explain the model's predictions using SHAP values
425+ ex = shap .TreeExplainer (booster )
426+ shap_values = ex .shap_values (X_test )
427+
428+ predicted = booster .predict (X_test , raw_score = True )
429+
430+ validation (shap_values , ex .expected_value , predicted )
431+
432+
433+ def test_lightgbm_nonsklearn_api_binary ():
434+ import numpy as np
435+
436+ import shap
437+
438+ # train the lightgbm model using non-sklearn API with binary classification dataset
439+ params = {
440+ "objective" : "binary" ,
441+ "num_threads" : 4 ,
442+ "n_estimators" : 8000 ,
443+ "early_stopping_round" : 50 ,
444+ "metric" : ["binary_error" ],
445+ "random_state" : 7 ,
446+ "verbose" : 1 ,
447+ }
448+
449+ def validation (shap_values , expected_value , predicted ):
450+ assert np .abs (shap_values .sum (1 ) + expected_value - predicted ).max () < 1e-4 , (
451+ "SHAP values don't sum to model output!"
452+ )
453+
454+ _common_lightgbm_nonsklearn_api (dataset = shap .datasets .iris (), params = params , validation = validation )
455+
456+
457+ def test_lightgbm_nonsklearn_api_regressor ():
458+ import numpy as np
459+
460+ import shap
461+
462+ # train the lightgbm model using non-sklearn API with regression dataset
463+ params = {
464+ "num_threads" : 4 ,
465+ "n_estimators" : 8000 ,
466+ "early_stopping_round" : 50 ,
467+ "metric" : ["rmse" ],
468+ "random_state" : 7 ,
469+ "verbose" : 1 ,
470+ }
471+
472+ def validation (shap_values , expected_value , predicted ):
473+ assert np .abs (shap_values .sum (1 ) + expected_value - predicted ).max () < 1e-4 , (
474+ "SHAP values don't sum to model output!"
475+ )
476+
477+ _common_lightgbm_nonsklearn_api (dataset = shap .datasets .adult (), params = params , validation = validation )
478+
479+
315480def test_gpboost ():
316481 gpboost = pytest .importorskip ("gpboost" )
317482 # train gpboost model
0 commit comments