-
Notifications
You must be signed in to change notification settings - Fork 16
/
Copy patheli.py
313 lines (273 loc) · 13.5 KB
/
eli.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
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
#%matplotlib inline
import matplotlib
import matplotlib.pyplot as plt
from sklearn import (datasets, metrics, model_selection,
linear_model, ensemble)
import xgboost as xgb
import pandas as pd
import numpy as np
import eli5
from collections import defaultdict
import json
import sys
class XGBExplainer(object):
def __init__(self, verbose=2):
self.verbose = verbose
def __repr__(self):
return "XGBExplainer(verbose=%s)"%(self.verbose)
def explain_model(self,
model,
X,
y,
feature_names=[],
target_names=[],
visualize=True,
importance_type="gain",
temporal=False,
max_samples=0):
def _infer_problem_from_model(model):
if "Classifier" in model.__class__.__name__:
print("\nClassification Problem")
return True
else:
print("\nRegression Problem")
return False
def _classification_report(y_test, preds):
if self.verbose > 0:
y_mean = np.mean(y_test)
print("\n%s:\t%f (benchmark: 0.5)"%("ROC AUC".ljust(20),
metrics.roc_auc_score(y_test, preds)))
print("%s:\t%f (benchmark: %f)"%("Accuracy".ljust(20),
metrics.accuracy_score(y_test,
preds > 0.5),
y_mean if y_mean > 0.5 else 1-y_mean))
print("%s:\t%f (benchmark: %f)\n"%("Log Loss".ljust(20),
metrics.log_loss(y_test, preds),
metrics.log_loss(y_test, [y_mean for p in preds])))
if len(target_names) > 0:
print(metrics.classification_report(y_test, preds > 0.5, target_names=target_names))
else:
print(metrics.classification_report(y_test, preds > 0.5))
print("Confusion Matrix:\n%s"%(metrics.confusion_matrix(y_test, preds > 0.5)))
def _regression_report(y_test, preds):
if self.verbose > 0:
y_mean = np.mean(y_test)
print("\nStandard Deviation of predictions: %f"%(np.std(preds)))
print("\n%s:\t%f (benchmark: %f)"%("Mean Absolute Error".ljust(20),
metrics.mean_absolute_error(y_test,
preds),
metrics.mean_absolute_error(y_test, [y_mean for p in preds])))
print("%s:\t%f (benchmark: %f)"%("Mean Squared Error".ljust(20),
metrics.mean_squared_error(y_test,
preds),
metrics.mean_squared_error(y_test, [y_mean for p in preds])))
print("%s:\t%f (benchmark: %f)"%("R2-Score".ljust(20),
metrics.r2_score(y_test,
preds),
0.))
print("%s:\t%f (benchmark: %f)"%("Explained Variance".ljust(20),
metrics.explained_variance_score(y_test,
preds),
0.))
def _predict(model, X_test, classification):
if classification:
return model.predict_proba(X_test)[:,1]
else:
return model.predict(X_test)
# Verbosity
if self.verbose > 0:
print("Explaining model: %s"%(model))
# Infer classification or regression problem
classification = _infer_problem_from_model(model)
# Split into train and test data
if self.verbose > 0:
print("\nSplitting data into 67%% train and 33%% train. Is temporal? %s"%(temporal))
if temporal:
cut_off = int(X.shape[0] * 0.67)
X_train, X_test, y_train, y_test = X[:cut_off],\
X[cut_off:],\
y[:cut_off],\
y[cut_off:]
else:
X_train, X_test, y_train, y_test = model_selection.train_test_split(X,
y,
test_size=0.33,
random_state=1)
# Fitting the model
if self.verbose > 0:
print("\nFitting on data shaped %s. Mean target: %f"%(X_train.shape,
np.mean(y_train)))
model.fit(X_train, y_train)
# Check if feature_names was passed or needs default names
# Else check if length of feature names matches length of fitted features
fscores = model.booster().get_fscore()
if len(feature_names) == 0:
feature_names = fscores.keys()
else:
if len(feature_names) != len(fscores):
sys.exit("`feature_names` length (%d) does not equal length of"%(len(feature_names)) \
+ " number of features used during fitting (%d)"%(len(fscores)))
# Verbosity
if self.verbose > 1:
print("\nUsing %d features: %s"%(len(feature_names), feature_names))
# Calculating feature importances
d = eli5.formatters.as_dict.format_as_dict(
eli5.explain_weights_xgboost(model,
feature_names=feature_names,
importance_type=importance_type))
# Format weights and feature names as a sorted list
l = sorted([(k["weight"], k["feature"]) for k in \
d["feature_importances"]["importances"]], reverse=True)
l_names = [k[1] for k in l]
for feature_name in feature_names:
if feature_name not in l_names:
l.append((0., feature_name))
l = sorted(l, reverse=True)
# Verbosity
if self.verbose > 0:
print("\nImportance: %s\t%s")%(importance_type.ljust(2), "Feature Name".ljust(20))
print("%s\t%s")%(str("-"*20).rjust(20), str("-"*20).ljust(20))
for weight, feature in l:
print("%s\t%s")%(str("%f"%(weight)).rjust(20), str(feature).ljust(20))
# Testing
if self.verbose > 0:
print("\nTesting on data shaped %s. Mean target: %f"%(X_test.shape,
np.mean(y_test)))
preds = _predict(model, X_test, classification)
if classification:
if self.verbose > 0:
_classification_report(y_test, preds)
else:
preds = model.predict(X_test)
if self.verbose > 0:
_regression_report(y_test, preds)
# Violin Density Plots of contributions
if self.verbose > 0:
print("\nCreating explanations for test set.")
if max_samples == 0:
max_samples = X_test.shape[0]
if max_samples > X_test.shape[0]:
if self.verbose > 0:
print("\n`max_samples` larger than test set size." \
+ " Resetting `max_samples` to %d\n"%(X_test.shape[0]))
max_samples = X_test.shape[0]
importances = defaultdict(list)
values = defaultdict(list)
for i in range(max_samples):
if self.verbose > 0:
if i % 100 == 0:
print("%d/%d"%(i+1, max_samples))
d = eli5.formatters.as_dict.format_as_dict(
eli5.explain_prediction(model, X_test[i], feature_names=feature_names))
for k in d["targets"][0]["feature_weights"]["neg"] \
+ d["targets"][0]["feature_weights"]["pos"]:
importances[k["feature"]].append(k["weight"])
values[k["feature"]].append(k["value"])
del importances["<BIAS>"]
pos = list(reversed(range(len(importances))))
data = [importances[k] for k in feature_names]
ax = plt.figure(figsize=(16,10))
plt.axvline(0, linestyle='-', color='k', alpha=0.1)
plt.violinplot(data, pos, points=max_samples, vert=False, widths=0.7,
showmeans=True, showextrema=True, showmedians=False)
plt.yticks(pos, [k for v, k in l], rotation='horizontal')
plt.grid(color='b', linestyle='--', linewidth=1, alpha=0.05)
plt.title("Feature Contributions\n", fontsize=30)
if visualize:
plt.show()
if visualize:
# Feature contribution by value plots
for feature in feature_names:
x, ys = [], []
for val, imp in sorted(zip(values[feature], importances[feature])):
x.append(val)
ys.append(imp)
plt.figure()
plt.scatter(x, ys, alpha=0.5)
plt.title(feature)
plt.grid(color='b', linestyle='--', linewidth=1, alpha=0.05)
trend = ensemble.RandomForestRegressor(random_state=1,
min_samples_leaf=5)
trend.fit([[xs] for xs in x], ys)
p = trend.predict([[xs] for xs in x])
trend = linear_model.Ridge(random_state=1)
trend.fit([[xs] for xs in x], ys)
p2 = trend.predict([[xs] for xs in x])
p = (p2+p+p)/3.
plt.plot(x, p, color=(0,0,0), linewidth=5, alpha=0.33)
plt.ylabel("Feature Contribution")
plt.xlabel("Feature Value")
plt.show()
self.explainer = {
"pos": pos,
"data": data,
"mean_X": np.mean(X, axis=0),
"std_X": np.std(X, axis=0),
"mean_y": np.mean(y),
"model": model,
"feature_names": np.array(feature_names),
"target_names": target_names,
"classification": classification,
"f_weights": l,
"max_samples": max_samples
}
def explain_sample(self, j, y="?"):
try:
self.explainer
except:
sys.exit("You need to explain the model first with `.explain_model()` before" \
+ " you can explain examples.")
print("\nExplaining sample shaped %s"%(j.shape))
d = eli5.formatters.as_dict.format_as_dict(
eli5.explain_prediction(self.explainer["model"],
j,
feature_names=[k for v, k in self.explainer["f_weights"]]))
# Fetch prediction depending on classification or regression problem
if self.explainer["classification"]:
p = d["targets"][0]["proba"]
else:
p = d["targets"][0]["score"]
# Fetch scatter points
s = {}
for k in d["targets"][0]["feature_weights"]["neg"] \
+ d["targets"][0]["feature_weights"]["pos"]:
s[k["feature"]] = k["weight"]
ds = []
for i, k in enumerate([k for v, k in self.explainer["f_weights"]]):
if k in s:
ds.append(s[k])
else:
ds.append(np.nan)
# Plot violin scatter plot
ax = plt.figure(figsize=(16,10))
plt.violinplot(self.explainer["data"],
self.explainer["pos"],
points=self.explainer["max_samples"],
vert=False,
widths=0.7,
showmeans=True,
showextrema=True,
showmedians=False)
plt.yticks(self.explainer["pos"], [k for v, k in self.explainer["f_weights"]],
rotation='horizontal')
plt.grid(color='b', linestyle='--', linewidth=1, alpha=0.05)
plt.scatter(ds, self.explainer["pos"], color='r')
plt.title("Target: %s\nPredicted:%f\nMean:%f"%(y, p, self.explainer["mean_y"]))
plt.ylabel("Feature Importance")
plt.xlabel("Feature Contribution")
plt.show()
# Print 1+ STD statistics by order of feature importance
std_m = np.sqrt((j - self.explainer["mean_X"])**2) / self.explainer["std_X"]
s = {}
for i, f in enumerate(self.explainer["feature_names"]):
s[f] = (j[i], std_m[i], self.explainer["mean_X"][i])
print("\n%s %s %s %s"%("FEATURE".ljust(20),
"VALUE".ljust(10),
"STD".ljust(10),
"MEAN".ljust(10)))
for w, f in self.explainer["f_weights"]:
if s[f][1] > 1.:
print("%s %s %s %s"%(f.ljust(20),
str(s[f][0]).ljust(10),
str("%f"%(s[f][1])).ljust(10),
str("%f"%(s[f][2])).ljust(10)))