-
Notifications
You must be signed in to change notification settings - Fork 3
/
predict.py
172 lines (152 loc) · 8.57 KB
/
predict.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
import pandas as pd
import numpy as np
import lightgbm as lgb
import os
from sklearn import model_selection, metrics, linear_model
import datetime as dt
from scipy import special
# Since the following forum topic and script show that relationships between columns don't matter, treat all columns
# individually and then build a meta model to combine the predictions.
# https://www.kaggle.com/c/santander-customer-transaction-prediction/discussion/83882
# https://www.kaggle.com/brandenkmurray/randomly-shuffled-data-also-works?scriptVersionId=11467087
# Values which repeat in a column seem to show stronger signal, so include a feature based on that. However, as
# observed in the following kernel, some of the test data seems to be simulated:
# https://www.kaggle.com/yag320/list-of-fake-samples-and-public-private-lb-split?scriptVersionId=11948999
# Therefore exclude these from the counts.
test_mode = False
# read input
train = pd.read_csv(os.path.join('input', 'train.csv'))
test = pd.read_csv(os.path.join('input', 'test.csv'))
if test_mode:
all_data = train
else:
all_data = pd.concat([train, test], ignore_index=True, sort=True).reset_index()
train_bool = all_data['target'].notnull()
# identify the fake test data as in above kernel
train_cols = [c for c in train.columns if c not in ['ID_code', 'target', 'predicted', 'size', 'index']]
unique_vals = {}
for c in train_cols:
sizes = test.groupby(c)['ID_code'].size()
unique_vals[c] = sizes.loc[sizes == 1].index
unique_bool = test.copy()
for c in train_cols:
unique_bool[c] = test[c].isin(unique_vals[c])
num_unique = unique_bool[train_cols].sum(axis=1)
test_real = test.loc[num_unique > 0]
# add the counts
count_data = pd.concat([train, test_real], axis=0, sort=True)
for i in range(200):
c = 'var_' + str(i)
all_size = count_data.groupby(c)['ID_code'].size().to_frame(c + '_size')
count_data['rank'] = count_data[c].rank(method='first')
count_data['bin'] = (count_data['rank'] - 1) // 300
count_data['size'] = count_data.groupby(c)['ID_code'].transform('size')
count_data['avg_size_in_bin'] = count_data.groupby('bin')['size'].transform('mean')
count_data[c + '_size_scaled'] = count_data['size'] / count_data['avg_size_in_bin']
all_size[c + '_size_scaled'] = count_data.groupby(c)[c + '_size_scaled'].first()
all_data = pd.merge(all_data, all_size, 'left', left_on=c, right_index=True)
def rolling_log_reg(var_arr, target_arr, window, step_size):
est = linear_model.LogisticRegression(random_state=0)
ans = np.zeros(len(range(0, len(var_arr), step_size)))
ans[:] = np.nan
for i, center in enumerate(range(0, len(var_arr), step_size)):
min_ind = max(0, center - window // 2)
max_ind = min(len(var_arr) - 1, center + window // 2)
x_curr = var_arr[min_ind:max_ind][:, np.newaxis]
y_curr = target_arr[min_ind:max_ind]
if len(np.unique(y_curr)) > 1:
est.fit(x_curr, y_curr)
ans[i] = est.predict_proba(np.array([var_arr[center]])[:, np.newaxis])[:, 1]
else:
ans[i] = np.unique(y_curr)[0]
return np.interp(var_arr, var_arr[np.arange(0, len(var_arr), step_size)], ans)
def rolling_prediction(input_df, train_bool, c):
col_copy = input_df[[c, c + '_size', 'target']].copy()
size1 = col_copy.loc[(col_copy[c + '_size'] == 1) & train_bool].copy().sort_values(c)
size2 = col_copy.loc[(col_copy[c + '_size'] > 1) & train_bool].copy().sort_values(c)
size1[c + '_rolling'] = rolling_log_reg(size1[c].values, size1['target'].values, 20000, 500)
size2[c + '_rolling'] = rolling_log_reg(size2[c].values, size2['target'].values, 20000, 500)
input_df[c + '_rolling'] = np.nan
input_df.loc[input_df[c + '_size'] <= 1, c + '_rolling'] \
= np.interp(input_df.loc[input_df[c + '_size'] <= 1, c], size1[c].values, size1[c + '_rolling'].values)
input_df.loc[input_df[c + '_size'] > 1, c + '_rolling'] \
= np.interp(input_df.loc[input_df[c + '_size'] > 1, c].values, size2[c].values, size2[c + '_rolling'].values)
return input_df
# define lgb model
params = {'objective': 'binary', 'learning_rate': 0.01, 'lambda_l1': 0, 'lambda_l2': 0, 'num_leaves': 4,
'bagging_fraction': 0.8, 'bagging_freq': 1, 'max_depth': 2, 'seed': 0, 'verbosity': -1}
num_rounds = 2000
# controls the contribution of rolling average model to final prediction
rolling_factor = 0.3
# log file
log_file = 'Log_{}.txt'.format(dt.datetime.now().strftime('%y%m%d_%H%M'))
# cross validate and find the optimal number of iterations for each column
best_iter = {c: [] for c in [c + '_size' for c in train_cols] + [c + '_size_scaled' for c in train_cols]}
auc_train = []
auc_test = []
cv = model_selection.KFold(5, shuffle=True, random_state=0)
train = all_data.loc[all_data['target'].notnull()].copy()
for train_ind, test_ind in list(cv.split(train)):
lgb_results = train[train_cols].copy()
rolling_results = train[train_cols].copy()
train_bool = train.index.isin(train_ind)
test_bool = train.index.isin(test_ind)
# individual column models
for c in train_cols:
# lgb: average 2 models with different features
lgb_results[c] = 0
for size_col in [c + '_size', c + '_size_scaled']:
lgb_train = lgb.Dataset(train.loc[train_bool, [c, size_col]],
label=train.loc[train_bool, 'target'].values)
lgb_valid = lgb.Dataset(train.loc[test_bool, [c, size_col]],
label=train.loc[test_bool, 'target'].values)
est = lgb.train(params=params, train_set=lgb_train, valid_sets=[lgb_train, lgb_valid],
valid_names=['train', 'valid'], num_boost_round=num_rounds, early_stopping_rounds=100)
lgb_results[c] += 0.5 * special.logit(est.predict(train[[c, size_col]],
num_iteration=max(est.best_iteration + 100, num_rounds)))
best_iter[size_col].append(est.best_iteration)
# rolling logistic model
train = rolling_prediction(train, train_bool, c)
rolling_results[c] = special.logit(train[c + '_rolling'])
# combine predictions and train meta model
comb_pred = (1 - rolling_factor) * lgb_results + rolling_factor * rolling_results
train['predicted'] = special.expit(comb_pred.mean(axis=1))
# score
auc_train.append(metrics.roc_auc_score(train.loc[train_bool, 'target'].values,
train.loc[train_bool, 'predicted'].values))
auc_test.append(metrics.roc_auc_score(train.loc[test_bool, 'target'].values,
train.loc[test_bool, 'predicted'].values))
print('train auc: ', auc_train)
print('test auc: ', auc_test)
with open(log_file, 'a') as f:
f.write('\ntrain auc: {}'.format(auc_train))
f.write('\ntest auc: {}'.format(auc_test))
print('mean train auc: ', np.mean(auc_train))
print('mean test auc: ', np.mean(auc_test))
with open(log_file, 'a') as f:
f.write('\nmean train auc: {}'.format(np.mean(auc_train)))
f.write('\nmean test auc: {}'.format(np.mean(auc_test)))
# train on all data and write output
if not(test_mode):
train_bool = all_data['target'].notnull()
lgb_results = all_data[train_cols].copy()
rolling_results = all_data[train_cols].copy()
for c in train_cols:
# lgb: average 2 models with different features
lgb_results[c] = 0
for size_col in [c + '_size', c + '_size_scaled']:
lgb_train = lgb.Dataset(all_data.loc[train_bool, [c, size_col]],
label=all_data.loc[train_bool, 'target'].values)
est = lgb.train(params=params, train_set=lgb_train, valid_sets=[lgb_train],
valid_names=['train'],
num_boost_round=min(num_rounds, max(1, int(np.mean(best_iter[size_col])))))
lgb_results[c] += 0.5 * special.logit(est.predict(all_data[[c, size_col]]))
# rolling mean
all_data = rolling_prediction(all_data, train_bool, c)
rolling_results[c] = special.logit(all_data[c + '_rolling'])
# combine predictions and train meta model
comb_pred = (1 - rolling_factor) * lgb_results + rolling_factor * rolling_results
all_data['predicted'] = special.expit(comb_pred.mean(axis=1))
all_data['target'] = all_data['predicted']
all_data.loc[~train_bool, ['ID_code', 'target']].to_csv(
'Submission_{}.csv'.format(dt.datetime.now().strftime('%y%m%d_%H%M')), index=False)