-
Notifications
You must be signed in to change notification settings - Fork 7
/
Classification_Experiment_sota.py
295 lines (257 loc) · 13 KB
/
Classification_Experiment_sota.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
# pylint: disable=invalid-name, line-too-long, duplicate-code, cell-var-from-loop, consider-using-enumerate, missing-function-docstring
"""
Experiment used in the introductory paper to evaluate the stability and robustness of the explanations
"""
import time
import warnings
import pickle
import numpy as np
import pandas as pd
import xgboost as xgb
from sklearn.ensemble import RandomForestClassifier, HistGradientBoostingClassifier
from sklearn.tree import DecisionTreeClassifier
from sklearn.svm import SVC
from sklearn.model_selection import train_test_split
from venn_abers import VennAbers
from shap import Explainer # version 0.44.0
from lime.lime_tabular import LimeTabularExplainer # version 0.2.0.1
from calibrated_explanations import CalibratedExplainer
# -------------------------------------------------------
# pylint: disable=invalid-name, missing-function-docstring
def debug_print(message, debug=True):
if debug:
print(message)
# ------------------------------------------------------
test_size = 20 # number of test samples per dataset
is_debug = True
num_rep = 30
descriptors = ['uncal','va',]#,'va'
Descriptors = {'uncal':'Uncal','va': 'VA'}
models = ['xGB','RF'] # ['xGB','RF','DT','SVM',] # 'NN',
# pylint: disable=line-too-long
datasets = {1:"pc1req",2:"haberman",3:"hepati",4:"transfusion",5:"spect",6:"heartS",7:"heartH",8:"heartC",9:"je4243",10:"vote",11:"kc2",12:"wbc",
13:"kc3",14:"creditA",15:"diabetes",16:"iono",17:"liver",18:"je4042",19:"sonar", 20:"spectf",21:"german",22:"ttt",23:"colic",24:"pc4",25:"kc1"}
klara = [1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25]#
tic_all = time.time()
# -----------------------------------------------------------------------------------------------------
results = {'num_rep': num_rep, 'test_size': test_size}
for dataset in klara:
dataSet = datasets[dataset]
tic_data = time.time()
print(dataSet)
fileName = 'data/' + dataSet + ".csv"
df = pd.read_csv(fileName, delimiter=';')
Xn, y = df.drop('Y',axis=1), df['Y']
no_of_classes = len(np.unique(y))
no_of_features = Xn.shape[1]
no_of_instances = Xn.shape[0]
t1 = DecisionTreeClassifier(min_weight_fraction_leaf=0.15) # Changed from min_leaf=4
t2 = DecisionTreeClassifier(min_weight_fraction_leaf=0.15)
s1 = SVC(probability=True)
s2 = SVC(probability=True)
r1 = RandomForestClassifier(n_estimators=100)
r2 = RandomForestClassifier(n_estimators=100)
h1 = HistGradientBoostingClassifier()
h2 = HistGradientBoostingClassifier()
g1 = xgb.XGBClassifier(objective='binary:logistic',use_label_encoder=False,eval_metric='logloss')
g2 = xgb.XGBClassifier(objective='binary:logistic',use_label_encoder=False,eval_metric='logloss')
model_dict = {'xGB':(g1,g2,"xGB",Xn),'RF':(r1,r2,"RF",Xn),'SVM': (s1,s2,"SVM",Xn),'DT': (t1,t2,"DT",Xn),'HGB': (h1,h2,"HGB",Xn)}#,'NN': (a1,a2,"NN",Xn)
model_struct = [model_dict[model] for model in models]
results[dataSet] = {}
for c1, c2, alg, X in model_struct:
tic_algorithm = time.time()
debug_print(dataSet+' '+alg)
results[dataSet][alg] = {}
calibrators = {}
for desc in descriptors:
calibrators[desc] = {'ce': []}
X_train, X_test, y_train, y_test = train_test_split(X.values, y.values, test_size=test_size,random_state=42)
X_prop_train, X_cal, y_prop_train, y_cal = train_test_split(X_train, y_train, test_size=0.33,random_state=42)
c2.fit(X_prop_train,y_prop_train)
calibrators['uncal']['model'] = c2
if 'va' in descriptors:
calibrators['va']['model'] = VennAbers()
calibrators['va']['model'].fit(c2.predict_proba(X_cal), y_cal)
calibrators['data'] = {'X_prop_train':X_prop_train,'y_prop_train':y_prop_train,'X_cal':X_cal,'y_cal':y_cal,'X_test':X_test,'y_test':y_test,}
np.random.seed(1337)
categorical_features = [i for i in range(no_of_features) if len(np.unique(X.iloc[:,i])) < 10]
ce = CalibratedExplainer(calibrators['uncal']['model'], X_cal, y_cal, \
feature_names=df.columns, categorical_features=categorical_features)
stability = {'ce':[], 'cce':[], 'lime':[], 'lime_va':[], 'shap':[], 'shap_va':[]}
stab_timer = {'ce':[], 'cce':[], 'lime':[], 'lime_va':[], 'shap':[], 'shap_va':[]}
robustness = {'ce':[], 'cce':[], 'lime':[], 'lime_va':[], 'shap':[], 'shap_va':[], 'proba':[], 'proba_va':[]}
rob_timer = {'ce':[], 'cce':[], 'lime':[], 'lime_va':[], 'shap':[], 'shap_va':[]}
i = 0
while i < num_rep:
try:
# print(f'{i}:',end='\t')
ce.set_seed(i)
tic = time.time()
explanations = ce.explore_alternatives(X_test)
ct = time.time()-tic
stab_timer['ce'].append(ct)
# print(f'{ct:.1f}',end='\t')
stability['ce'].append([f.feature_weights for f in explanations])
ce.set_seed(i)
tic = time.time()
explanations = ce.explain_factual(X_test)
ct = time.time()-tic
stab_timer['cce'].append(ct)
# print(f'{ct:.1f}',end='\t')
stability['cce'].append([f.feature_weights for f in explanations])
lime = LimeTabularExplainer(X_cal, feature_names=df.columns, random_state=i)
model = calibrators['uncal']['model']
tic = time.time()
lime_exps_cl = []
for instance in X_test:
exp = lime.explain_instance(instance, model.predict_proba)
lime_exps_cl.append(exp)
ct = time.time()-tic
stab_timer['lime'].append(ct)
# print(f'{ct:.1f}',end='\t')
tmps = []
for j in range(len(lime_exps_cl)):
tmp = np.zeros(no_of_features)
for _,f in enumerate(lime_exps_cl[j].local_exp[1]):
tmp[f[0]] = f[1]
tmps.append(tmp)
stability['lime'].append(tmps)
model = calibrators['va']['model']
tic = time.time()
lime_exps_cl = []
for instance in X_test:
exp = lime.explain_instance(instance, lambda x: model.predict_proba(x)[0])
lime_exps_cl.append(exp)
ct = time.time()-tic
stab_timer['lime_va'].append(ct)
# print(f'{ct:.1f}',end='\t')
tmps = []
for j in range(len(lime_exps_cl)):
tmp = np.zeros(no_of_features)
for _,f in enumerate(lime_exps_cl[j].local_exp[1]):
tmp[f[0]] = f[1]
tmps.append(tmp)
stability['lime_va'].append(tmps)
shap = Explainer(lambda x: calibrators['uncal']['model'].predict_proba(x)[:,1], X_cal, \
feature_names=df.columns)
shap.random_state = i
tic = time.time()
explanations = shap(X_test)
ct = time.time()-tic
stab_timer['shap'].append(ct)
# print(f'{ct:.1f}',end='\t')
stability['shap'].append(explanations.values) # pylint: disable=no-member
shap_va = Explainer(lambda x: calibrators['va']['model'].predict_proba(x)[0][:,1], X_cal, \
feature_names=df.columns)
shap.random_state = i
tic = time.time()
explanations = shap_va(X_test)
ct = time.time()-tic
stab_timer['shap_va'].append(ct)
# print(f'{ct:.1f}')
stability['shap_va'].append(explanations.values) # pylint: disable=no-member
i += 1
except Exception as e: # pylint: disable=broad-exception-caught
warnings.warn(f'Error: {e}')
print('_',end='')
print('')
results[dataSet][alg]['stability'] = stability
results[dataSet][alg]['stab_timer'] = stab_timer
i = 0
while i < num_rep:
np.random.seed(i)
if alg == 'xGB':
c2 = xgb.XGBClassifier(objective='binary:logistic',use_label_encoder=False,eval_metric='logloss', random_state=i)
else:
c2 = RandomForestClassifier(n_estimators=100, random_state=i)
X_prop_train, X_cal, y_prop_train, y_cal = train_test_split(X_train, y_train, test_size=0.33,random_state=i)
c2.fit(X_prop_train,y_prop_train)
calibrators['uncal']['model'] = c2
if 'va' in descriptors:
calibrators['va']['model'] = VennAbers()
calibrators['va']['model'].fit(c2.predict_proba(X_cal), y_cal)
ce = CalibratedExplainer(calibrators['uncal']['model'], X_cal, y_cal, \
feature_names=df.columns, categorical_features=categorical_features)
robustness['proba'].append(calibrators['uncal']['model'].predict_proba(X_test)[:,1])
robustness['proba_va'].append(calibrators['va']['model'].predict_proba(X_test)[0][:,1])
try:
# print(f'{i}:',end='\t')
# ce.set_random_state(i)
tic = time.time()
explanations = ce.explain_factual(X_test)
ct = time.time()-tic
rob_timer['ce'].append(ct)
# print(f'{ct:.1f}',end='\t')
robustness['ce'].append([f.feature_weights for f in explanations])
# ce.set_random_state(i)
tic = time.time()
explanations = ce.explore_alternatives(X_test)
ct = time.time()-tic
rob_timer['cce'].append(ct)
# print(f'{ct:.1f}',end='\t')
robustness['cce'].append([f.feature_weights for f in explanations])
lime = LimeTabularExplainer(X_cal, feature_names=df.columns)
model = calibrators['uncal']['model']
tic = time.time()
lime_exps_cl = []
for instance in X_test:
exp = lime.explain_instance(instance, model.predict_proba)
lime_exps_cl.append(exp)
ct = time.time()-tic
rob_timer['lime'].append(ct)
# print(f'{ct:.1f}',end='\t')
tmps = []
for j in range(len(lime_exps_cl)):
tmp = np.zeros(no_of_features)
for _,f in enumerate(lime_exps_cl[j].local_exp[1]):
tmp[f[0]] = f[1]
tmps.append(tmp)
robustness['lime'].append(tmps)
model = calibrators['va']['model']
tic = time.time()
lime_exps_cl = []
for instance in X_test:
exp = lime.explain_instance(instance, lambda x: model.predict_proba(x)[0])
lime_exps_cl.append(exp)
ct = time.time()-tic
rob_timer['lime_va'].append(ct)
# print(f'{ct:.1f}',end='\t')
tmps = []
for j in range(len(lime_exps_cl)):
tmp = np.zeros(no_of_features)
for _,f in enumerate(lime_exps_cl[j].local_exp[1]):
tmp[f[0]] = f[1]
tmps.append(tmp)
robustness['lime_va'].append(tmps)
shap = Explainer(lambda x: calibrators['uncal']['model'].predict_proba(x)[:,1], X_cal, \
feature_names=df.columns)
# shap.random_state = i
tic = time.time()
explanations = shap(X_test)
ct = time.time()-tic
rob_timer['shap'].append(ct)
# print(f'{ct:.1f}',end='\t')
robustness['shap'].append(explanations.values) # pylint: disable=no-member
shap_va = Explainer(lambda x: calibrators['va']['model'].predict_proba(x)[0][:,1], X_cal, \
feature_names=df.columns)
# shap.random_state = i
tic = time.time()
explanations = shap_va(X_test)
ct = time.time()-tic
rob_timer['shap_va'].append(ct)
# print(f'{ct:.1f}',end='\t')
robustness['shap_va'].append(explanations.values) # pylint: disable=no-member
i += 1
except Exception as e: # pylint: disable=broad-exception-caught
warnings.warn(f'Error: {e}')
print('_',end='')
print('')
results[dataSet][alg]['robustness'] = robustness
results[dataSet][alg]['rob_timer'] = rob_timer
toc_data = time.time()
debug_print(dataSet + ': ' +str(toc_data-tic_data),is_debug )
with open('evaluation/results_sota.pkl', 'wb') as f:
pickle.dump(results, f)
# pickle.dump(results, open('evaluation/results_stab_rob.pkl', 'wb'))
toc_all = time.time()
debug_print(str(toc_data-tic_data),is_debug )