-
Notifications
You must be signed in to change notification settings - Fork 1
/
optimize_params.py
251 lines (211 loc) · 10.8 KB
/
optimize_params.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
import EncoderFactory
from DatasetManager import DatasetManager
import BucketFactory
import pandas as pd
import numpy as np
from sklearn.metrics import roc_auc_score
from sklearn.pipeline import FeatureUnion, Pipeline
from sklearn.preprocessing import StandardScaler
import time
import os
import sys
from sys import argv
import pickle
from collections import defaultdict
from sklearn.ensemble import RandomForestClassifier
import xgboost as xgb
from sklearn.linear_model import LogisticRegression
from sklearn.svm import SVC
from hyperopt import Trials, STATUS_OK, tpe, fmin, hp
import hyperopt
from hyperopt.pyll.base import scope
from hyperopt.pyll.stochastic import sample
def create_and_evaluate_model(args):
global trial_nr
trial_nr += 1
start = time.time()
score = 0
for cv_iter in range(n_splits):
dt_test_prefixes = dt_prefixes[cv_iter]
dt_train_prefixes = pd.DataFrame()
for cv_train_iter in range(n_splits):
if cv_train_iter != cv_iter:
dt_train_prefixes = pd.concat([dt_train_prefixes, dt_prefixes[cv_train_iter]], axis=0, sort=False)
# Bucketing prefixes based on control flow
bucketer_args = {'encoding_method':bucket_encoding,
'case_id_col':dataset_manager.case_id_col,
'cat_cols':[dataset_manager.activity_col],
'num_cols':[],
'random_state':random_state}
if bucket_method == "cluster":
bucketer_args["n_clusters"] = args["n_clusters"]
bucketer = BucketFactory.get_bucketer(bucket_method, **bucketer_args)
bucket_assignments_train = bucketer.fit_predict(dt_train_prefixes)
bucket_assignments_test = bucketer.predict(dt_test_prefixes)
preds_all = []
test_y_all = []
if "prefix" in method_name:
scores = defaultdict(int)
for bucket in set(bucket_assignments_test):
relevant_train_cases_bucket = dataset_manager.get_indexes(dt_train_prefixes)[bucket_assignments_train == bucket]
relevant_test_cases_bucket = dataset_manager.get_indexes(dt_test_prefixes)[bucket_assignments_test == bucket]
dt_test_bucket = dataset_manager.get_relevant_data_by_indexes(dt_test_prefixes, relevant_test_cases_bucket)
test_y = dataset_manager.get_label_numeric(dt_test_bucket)
if len(relevant_train_cases_bucket) == 0:
preds = [class_ratios[cv_iter]] * len(relevant_test_cases_bucket)
else:
dt_train_bucket = dataset_manager.get_relevant_data_by_indexes(dt_train_prefixes, relevant_train_cases_bucket) # one row per event
train_y = dataset_manager.get_label_numeric(dt_train_bucket)
if len(set(train_y)) < 2:
preds = [train_y[0]] * len(relevant_test_cases_bucket)
else:
feature_combiner = FeatureUnion([(method, EncoderFactory.get_encoder(method, **cls_encoder_args)) for method in methods])
if cls_method == "rf":
cls = RandomForestClassifier(n_estimators=500,
max_features=args['max_features'],
random_state=random_state)
elif cls_method == "xgboost":
cls = xgb.XGBClassifier(objective='binary:logistic',
n_estimators=500,
learning_rate= args['learning_rate'],
subsample=args['subsample'],
max_depth=int(args['max_depth']),
colsample_bytree=args['colsample_bytree'],
min_child_weight=int(args['min_child_weight']),
seed=random_state)
elif cls_method == "logit":
cls = LogisticRegression(C=2**args['C'],
random_state=random_state)
elif cls_method == "svm":
cls = SVC(C=2**args['C'],
gamma=2**args['gamma'],
random_state=random_state)
if cls_method == "svm" or cls_method == "logit":
pipeline = Pipeline([('encoder', feature_combiner), ('scaler', StandardScaler()), ('cls', cls)])
else:
pipeline = Pipeline([('encoder', feature_combiner), ('cls', cls)])
pipeline.fit(dt_train_bucket, train_y)
if cls_method == "svm":
preds = pipeline.decision_function(dt_test_bucket)
else:
preds_pos_label_idx = np.where(cls.classes_ == 1)[0][0]
preds = pipeline.predict_proba(dt_test_bucket)[:,preds_pos_label_idx]
if "prefix" in method_name:
auc = 0.5
if len(set(test_y)) == 2:
auc = roc_auc_score(test_y, preds)
scores[bucket] += auc
preds_all.extend(preds)
test_y_all.extend(test_y)
score += roc_auc_score(test_y_all, preds_all)
if "prefix" in method_name:
for k, v in args.items():
for bucket, bucket_score in scores.items():
fout_all.write("%s;%s;%s;%s;%s;%s;%s;%s\n" % (trial_nr, dataset_name, cls_method, method_name, bucket, k, v, bucket_score / n_splits))
fout_all.write("%s;%s;%s;%s;%s;%s;%s;%s\n" % (trial_nr, dataset_name, cls_method, method_name, 0, "processing_time", time.time() - start, 0))
else:
for k, v in args.items():
fout_all.write("%s;%s;%s;%s;%s;%s;%s\n" % (trial_nr, dataset_name, cls_method, method_name, k, v, score / n_splits))
fout_all.write("%s;%s;%s;%s;%s;%s;%s\n" % (trial_nr, dataset_name, cls_method, method_name, "processing_time", time.time() - start, 0))
fout_all.flush()
return {'loss': -score / n_splits, 'status': STATUS_OK, 'model': cls}
# dataset_ref = argv[1]
# params_dir = argv[2]
# n_iter = int(argv[3])
# bucket_method = argv[4]
# cls_encoding = argv[5]
# cls_method = argv[6]
dataset_ref = "bpic2012"
params_dir = "params"
n_iter = 3
bucket_method = "single"
cls_encoding = "agg"
cls_method = "rf"
if bucket_method == "state":
bucket_encoding = "last"
else:
bucket_encoding = "agg"
method_name = "%s_%s"%(bucket_method, cls_encoding)
dataset_ref_to_datasets = {
"bpic2011": ["bpic2011_f%s"%formula for formula in range(4,5)],
"bpic2015": ["bpic2015_%s_f2"%(municipality) for municipality in range(5,6)],
"insurance": ["insurance_activity", "insurance_followup"],
"bpic2012" : ["bpic2012_accepted"],
"sepsis_cases": ["sepsis_cases_1", "sepsis_cases_2", "sepsis_cases_4"]
}
encoding_dict = {
"laststate": ["static", "last"],
"agg": ["static", "agg"],
"index": ["static", "index"],
"combined": ["static", "last", "agg"]
}
datasets = [dataset_ref] if dataset_ref not in dataset_ref_to_datasets else dataset_ref_to_datasets[dataset_ref]
methods = encoding_dict[cls_encoding]
print(datasets)
train_ratio = 0.8
n_splits = 3
random_state = 22
# create results directory
if not os.path.exists(os.path.join(params_dir)):
os.makedirs(os.path.join(params_dir))
for dataset_name in datasets:
# read the data
dataset_manager = DatasetManager(dataset_name)
data = dataset_manager.read_dataset()
cls_encoder_args = {'case_id_col': dataset_manager.case_id_col,
'static_cat_cols': dataset_manager.static_cat_cols,
'static_num_cols': dataset_manager.static_num_cols,
'dynamic_cat_cols': dataset_manager.dynamic_cat_cols,
'dynamic_num_cols': dataset_manager.dynamic_num_cols,
'fillna': True}
# determine min and max (truncated) prefix lengths
min_prefix_length = 1
if "traffic_fines" in dataset_name:
max_prefix_length = 10
elif "bpic2017" in dataset_name:
max_prefix_length = min(20, dataset_manager.get_pos_case_length_quantile(data, 0.90))
else:
max_prefix_length = min(40, dataset_manager.get_pos_case_length_quantile(data, 0.90))
# split into training and test
print("splitting data")
train, _ = dataset_manager.split_data_strict(data, train_ratio, split="temporal")
# prepare chunks for CV
dt_prefixes = []
class_ratios = []
for train_chunk, test_chunk in dataset_manager.get_stratified_split_generator(train, n_splits=n_splits):
class_ratios.append(dataset_manager.get_class_ratio(train_chunk))
# generate data where each prefix is a separate instance
dt_prefixes.append(dataset_manager.generate_prefix_data(test_chunk, min_prefix_length, max_prefix_length))
del train
# set up search space
if cls_method == "rf":
space = {'max_features': hp.uniform('max_features', 0, 1)}
elif cls_method == "xgboost":
space = {'learning_rate': hp.uniform("learning_rate", 0, 1),
'subsample': hp.uniform("subsample", 0.5, 1),
'max_depth': scope.int(hp.quniform('max_depth', 4, 30, 1)),
'colsample_bytree': hp.uniform("colsample_bytree", 0.5, 1),
'min_child_weight': scope.int(hp.quniform('min_child_weight', 1, 6, 1))}
elif cls_method == "logit":
space = {'C': hp.uniform('C', -15, 15)}
elif cls_method == "svm":
space = {'C': hp.uniform('C', -15, 15),
'gamma': hp.uniform('gamma', -15, 15)}
if bucket_method == "cluster":
space['n_clusters'] = scope.int(hp.quniform('n_clusters', 2, 50, 1))
# optimize parameters
trial_nr = 1
trials = Trials()
fout_all = open(os.path.join(params_dir, "param_optim_all_trials_%s_%s_%s.csv" % (cls_method, dataset_name, method_name)), "w")
if "prefix" in method_name:
fout_all.write("%s;%s;%s;%s;%s;%s;%s;%s\n" % ("iter", "dataset", "cls", "method", "nr_events", "param", "value", "score"))
else:
fout_all.write("%s;%s;%s;%s;%s;%s;%s\n" % ("iter", "dataset", "cls", "method", "param", "value", "score"))
best = fmin(create_and_evaluate_model, space, algo=tpe.suggest, max_evals=n_iter, trials=trials, verbose=True)
fout_all.close()
# write the best parameters
best_params = hyperopt.space_eval(space, best)
outfile = os.path.join(params_dir, "optimal_params_%s_%s_%s.pickle" % (cls_method, dataset_name, method_name))
# write to file
with open(outfile, "wb") as fout:
pickle.dump(best_params, fout)