-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
134 lines (106 loc) · 4.53 KB
/
train.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
import numpy as np
import pandas as pd
import seaborn as sns
import matplotlib.pyplot as plt
from scipy.stats import uniform
from sklearn.model_selection import train_test_split
from sklearn.metrics import PrecisionRecallDisplay, mutual_info_score, recall_score, f1_score, roc_auc_score, roc_curve, auc, precision_recall_curve, classification_report, make_scorer
from sklearn.feature_extraction import DictVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import KFold, StratifiedKFold, GridSearchCV, RandomizedSearchCV
from imblearn.under_sampling import RandomUnderSampler
from xgboost import XGBClassifier
from imblearn.pipeline import Pipeline
from sklearn.ensemble import RandomForestClassifier
from sklearn.preprocessing import MinMaxScaler
import joblib
# reading train data
data_descriptions = pd.read_csv("data/data_descriptions.csv")
train = pd.read_csv("data/train.csv")
# removing customerID as it is not useful for predicting churn
train.drop(columns=['CustomerID'], inplace=True)
# splitting train data into train, val sets with 80%, 20% data
df_train, df_val = train_test_split(train, test_size=0.2, random_state=1)
# separating target variable
y_train = df_train.Churn.values
y_val = df_val.Churn.values
# deleting target variable from feature set
del df_train['Churn']
del df_val['Churn']
# one hot encoding the categorical variables
dv = DictVectorizer(sparse=False)
train_dict = df_train.to_dict(orient='records')
X_train = dv.fit_transform(train_dict)
val_dict = df_val.to_dict(orient='records')
X_val = dv.transform(val_dict)
prop_of_churn_cases = y_train.sum()/len(y_train)
IR = prop_of_churn_cases/(1-prop_of_churn_cases)
def get_model(n_splits=5):
cv = StratifiedKFold(**{
'n_splits': n_splits,
'shuffle': True,
'random_state': 1
})
# recall_scorer = make_scorer(
# recall_score,
# greater_is_better=True,
# needs_proba=True
# )
scoring = 'f1_macro'
scaler = MinMaxScaler()
sampler = RandomUnderSampler(**{
'random_state': 1
})
classifier = XGBClassifier(**{
'random_state': 1,
'n_jobs': -1,
'verbosity': 2,
'validate_parameters': True,
'disable_default_eval_metric': True,
'eval_metric' : 'aucpr',
'scale_pos_weight' : 1/IR # Since scale_pos_weight quantifies the cost of a false negative, we set its value to the inverse of the imbalance ratio.
})
pipeline = Pipeline([
('scaler', scaler),
('sampler', sampler),
('classifier', classifier)
])
param_grid = {
'sampler__sampling_strategy': ['all', 'auto'], #all = resamples both classes, auto = resamples 'not minority' class so no of samples is equalized in both classes.
'classifier__n_estimators': [25, 50, 100, 125, 150, 200], # no of trees
'classifier__max_depth': [3, 6, 9, 12, 15, 18, 20, 25, 30, 35],
'classifier__eta' : [0.001, 0.01, 0.1, 1],
'classifier__min_child_weight' : [1, 3, 5, 7],
'classifier__max_delta_step' : [0, 1, 3, 5], #needed when imbalance exists
'classifier__subsample' : [0.5, 0.75, 1], # 0.75 for preventing overfitting
'classifier__alpha' : [0, 1, 3, 5, 10, 15], # L1 regularization. Can act as in-built feature selection as we have seen many unimportant features
'classifier__colsample_bytree' : [0.3, 0.4, 0.5 , 0.7, 0.9],
'classifier__gamma' : [ 0.0, 0.1, 0.2 , 0.3, 0.4 ]
}
model = RandomizedSearchCV(**{
'estimator': pipeline,
'cv': cv,
'param_distributions': param_grid,
'verbose': 4,
'scoring': scoring,
'error_score': np.NaN,
'n_jobs': -1,
'random_state' : 2
})
return model
model = get_model(n_splits=3)
model.fit(X_train, y_train)
# predictions on validation set
y_pred = model.predict_proba(X_val)[:,1]
precision, recall, thresholds = precision_recall_curve(y_val, y_pred, drop_intermediate=True)
score_df = pd.DataFrame([precision, recall, thresholds]).T
score_df.columns = ['precision','recall','thresholds']
score_df['f1_score'] = (2*score_df['precision']*score_df['recall'])/(score_df['precision']+score_df['recall'])
score_df.sort_values(by='f1_score', ascending=False)
# threshold for which f1_score is highest
decision_threshold = round(score_df.sort_values(by='f1_score', ascending=False)['thresholds'].reset_index(drop=True)[0],2)
# save the churn classification model as a pickle file
model_file_name = "churn_xgb_model.sav"
dv_file_name = "dv.sav"
joblib.dump(model, model_file_name)
joblib.dump(dv, dv_file_name)