def xgb_online(train, cv, test):
train_data = train.copy()
cv_data = cv.copy()
test_data = test.copy()
train_data = pd.concat([train_data, cv_data], axis=0)
train_data.reset_index(inplace=True, drop=True)
train_Y = train_data['is_trade']
drop_cols = ['is_trade']
train_data.drop(drop_cols, axis=1, inplace=True)
test_data.drop(drop_cols, axis=1, inplace=True)
folds = 5
kf = KFold(len(train_data), n_folds=folds, shuffle=True, random_state=7)
train_preds = np.zeros(train_data.shape[0])
cv_preds = np.zeros(train_data.shape[0])
test_preds = np.zeros((test_data.shape[0], 5))
for i, (train_index, cv_index) in enumerate(kf):
print('第{}次训练...'.format(i))
train_feat = train_data.loc[train_index]
cv_feat = train_data.loc[cv_index]
train_feat = xgb.DMatrix(train_feat.values, label=train_Y[train_index])
cv_feat = xgb.DMatrix(cv_feat.values, label=train_Y[cv_index])
test_feat = xgb.DMatrix(test_data.values)
watchlist = [(train_feat, 'train'), (cv_feat, 'val')]
clf = xgb.train(params=params, dtrain=train_feat,num_boost_round=n_round,\
evals=watchlist,early_stopping_rounds=7,verbose_eval=False)
predict_train = clf.predict(train_feat)
predict_cv = clf.predict(cv_feat)
predict_test = clf.predict(test_feat)
train_preds[train_index] += predict_train
cv_preds[cv_index] += predict_cv
test_preds[:, i] = predict_test
#特征重要度
features = train_data.columns
ceate_feature_map(features)
importance = clf.get_fscore(fmap='xgb.fmap')
importance = sorted(importance.items(), key=operator.itemgetter(1))
feat_imp = pd.DataFrame(importance, columns=['feature', 'fscore'])
feat_imp['fscore'] = feat_imp['fscore'] / feat_imp['fscore'].sum()
print(clf.best_iteration)
print(clf.best_score)
print(' 训练损失:', cal_log_loss(predict_train,
train_Y.loc[train_index]))
print(' 测试损失:', cal_log_loss(predict_cv, train_Y.loc[cv_index]))
predict_test = np.median(test_preds, axis=1)
predict_test = predict_test / (predict_test + (1 - predict_test) / rate)
print(params)
print('训练损失:', cal_log_loss(train_preds / (folds - 1), train_Y))
print('测试损失:', cal_log_loss(cv_preds, train_Y))
print('test mean:', np.mean(predict_test))
submmit_result(predict_test, 'XGB')
return feat_imp, predict_test
def lgb_online(train, cv, test):
train_data = train.copy()
cv_data = cv.copy()
test_data = test.copy()
train_data = pd.concat([train_data, cv_data], axis=0)
train_data.reset_index(inplace=True, drop=True)
train_Y = train_data['is_trade']
drop_cols = ['is_trade']
train_data.drop(drop_cols, axis=1, inplace=True)
test_data.drop(drop_cols, axis=1, inplace=True)
folds = 5
kf = KFold(len(train_data), n_folds=folds, shuffle=True, random_state=7)
train_preds = np.zeros(train_data.shape[0])
cv_preds = np.zeros(train_data.shape[0])
test_preds = np.zeros((test_data.shape[0], 5))
for i, (train_index, cv_index) in enumerate(kf):
print('第{}次训练...'.format(i))
train_feat = train_data.loc[train_index]
cv_feat = train_data.loc[cv_index]
lgb_train = lgb.Dataset(train_feat.values, train_Y.loc[train_index])
lgb_cv = lgb.Dataset(cv_feat.values, train_Y.loc[cv_index])
gbm = lgb.train(params=params,
train_set=lgb_train,
num_boost_round=6000,
valid_sets=lgb_cv,
verbose_eval=False,
early_stopping_rounds=50)
#评价特征的重要性
feat_imp = pd.Series(
gbm.feature_importance(),
index=train_data.columns).sort_values(ascending=False)
predict_train = gbm.predict(train_feat.values)
predict_cv = gbm.predict(cv_feat.values)
test_preds[:, i] = gbm.predict(test_data.values)
train_preds[train_index] += predict_train
cv_preds[cv_index] += predict_cv
feat_imp = pd.Series(
gbm.feature_importance(),
index=train_data.columns).sort_values(ascending=False)
print(gbm.best_iteration)
print(gbm.best_score)
print(' 训练损失:', cal_log_loss(predict_train,
train_Y.loc[train_index]))
print(' 测试损失:', cal_log_loss(predict_cv, train_Y.loc[cv_index]))
predict_test = np.median(test_preds, axis=1)
predict_test = predict_test / (predict_test + (1 - predict_test) / rate)
print(params)
print('训练损失:', cal_log_loss(train_preds / (folds - 1), train_Y))
print('测试损失:', cal_log_loss(cv_preds, train_Y))
print('test mean:', np.mean(predict_test))
submmit_result(predict_test, 'LGB')
return feat_imp, predict_test
def LR_online(train_data, cv_data, test_data):
print('on line')
train_data = pd.concat([train_data, cv_data], axis=0)
train_data.reset_index(inplace=True, drop=True)
train_Y = train_data['is_trade']
fold = 5
kf = KFold(len(train_data), n_folds=fold, shuffle=True, random_state=520)
train_preds = np.zeros(train_data.shape[0])
cv_preds = np.zeros(train_data.shape[0])
test_preds = np.zeros((test_data.shape[0], fold))
data_in_process = [(train_data.loc[train_index], train_data.loc[cv_index],
test_data)
for i, (train_index, cv_index) in enumerate(kf)]
index_all = [(train_index, cv_index)
for i, (train_index, cv_index) in enumerate(kf)]
with multiprocessing.Pool(fold) as p:
k_val_list = p.map(_LR_train, data_in_process)
for i, index, val in zip(range(fold), index_all, k_val_list):
print('no %d train' % (i))
train_index, cv_index = index
predict_train, predict_cv, predict_test = val
train_preds[train_index] += predict_train
cv_preds[cv_index] += predict_cv
test_preds[:, i] = predict_test
print(' 训练损失:', cal_log_loss(predict_train, train_Y[train_index]))
print(' 测试损失:', cal_log_loss(predict_cv, train_Y[cv_index]))
'''
for i, (train_index, cv_index) in enumerate(kf):
print('no %d train:'%(i))
train_feat = train_data.loc[train_index]
cv_feat = train_data.loc[cv_index]
predict_train,predict_cv,predict_test = _LR_train((train_feat,cv_feat,test_data))
train_preds[train_index] += predict_train
cv_preds[cv_index] += predict_cv
test_preds[:,i] = predict_test
print(' 训练损失:',cal_log_loss(predict_train, train_Y[train_index]))
print(' 测试损失:',cal_log_loss(predict_cv, train_Y[cv_index]))
'''
predict_test = np.median(test_preds, axis=1)
print('mean:', np.mean(predict_test))
print('训练损失:', cal_log_loss(train_preds / (fold - 1), train_Y))
print('测试损失:', cal_log_loss(cv_preds, train_Y))
submmit_result(predict_test, 'LR')
def LR_online(train_data, cv_data, test_data):
train_data = pd.concat([train_data, cv_data],axis=0)
train_data = build_train_dataset(train_data, rate)
train_data.reset_index(inplace=True,drop=True)
train_Y = train_data['is_trade']
drop_cols = ['is_trade']
train_data.drop(drop_cols,axis=1,inplace=True)
test_data.drop(drop_cols,axis=1,inplace=True)
fold = 5
kf = KFold(len(train_data), n_folds = fold, shuffle=True, random_state=520)
train_preds = np.zeros(train_data.shape[0])
cv_preds = np.zeros(train_data.shape[0])
test_preds = np.zeros((test_data.shape[0], fold))
for i, (train_index, cv_index) in enumerate(kf):
train_feat = train_data.loc[train_index]
cv_feat = train_data.loc[cv_index]
clf = LogisticRegression(C=1.2, fit_intercept=True, max_iter=3000,class_weight={0:0.5, 1:0.5})
clf.fit(X=train_feat.values, y=train_Y[train_index])
predict_train = clf.predict_proba(train_feat.values)[:,1]
predict_cv = clf.predict_proba(cv_feat.values)[:,1]
predict_test = clf.predict_proba(test_data.values)[:,1]
train_preds[train_index] += predict_train
cv_preds[cv_index] += predict_cv
test_preds[:,i] = predict_test
print(' 训练损失:',cal_log_loss(predict_train, train_Y[train_index]))
print(' 测试损失:',cal_log_loss(predict_cv, train_Y[cv_index]))
predict_test = np.median(test_preds,axis=1)
predict_test = predict_test/(predict_test+(1-predict_test)/rate)
print('训练损失:',cal_log_loss(train_preds/4, train_Y))
print('测试损失:',cal_log_loss(cv_preds, train_Y))
submmit_result(predict_test, 'LR')
def online(train_data, cv_data, test_data):
#剔除历史数据，保留老用户的历史数据
cv_data.index += len(train_data)
train_data = pd.concat([train_data, cv_data], axis=0)
history_cols = ['user_id_cvr_smooth', 'user_id_buy_count']
old_user_data_train = train_data[history_cols]
old_user_data_test = test_data[history_cols]
#对数据集进行训练
train_Y = train_data['is_trade']
drop_cols = ['is_trade']
train_data.drop(drop_cols, axis=1, inplace=True)
test_data.drop(drop_cols, axis=1, inplace=True)
kf = KFold(len(train_data), n_folds=5, shuffle=True, random_state=520)
train_preds = np.zeros(train_data.shape[0])
cv_preds = np.zeros(train_data.shape[0])
test_preds = np.zeros((test_data.shape[0], 5))
for i, (train_index, cv_index) in enumerate(kf):
train_feat = train_data.loc[train_index]
cv_feat = train_data.loc[cv_index]
print('第{}次训练...'.format(i))
lgb_train = lgb.Dataset(train_feat.values, train_Y.loc[train_index])
lgb_cv = lgb.Dataset(cv_feat.values, train_Y.loc[cv_index])
gbm = lgb.train(params=params,
train_set=lgb_train,
num_boost_round=6000,
valid_sets=lgb_cv,
verbose_eval=False,
early_stopping_rounds=200)
#评价特征的重要性
feat_imp = pd.Series(
gbm.feature_importance(),
index=train_data.columns).sort_values(ascending=False)
predict_train = gbm.predict(train_feat.values)
predict_cv = gbm.predict(cv_feat.values)
test_preds[:, i] = gbm.predict(test_data.values)
train_preds[train_index] += predict_train
cv_preds[cv_index] += predict_cv
feat_imp = pd.Series(
gbm.feature_importance(),
index=train_data.columns).sort_values(ascending=False)
print(gbm.best_iteration)
print(gbm.best_score)
print(' 训练损失:', cal_log_loss(predict_train,
train_Y.loc[train_index]))
print(' 测试损失:', cal_log_loss(predict_cv, train_Y.loc[cv_index]))
test_preds = np.median(test_preds, axis=1)
print(params)
print('训练损失:', cal_log_loss(train_preds / 4, train_Y))
print('测试损失:', cal_log_loss(cv_preds, train_Y))
#划分出新老用户的分数并计算损失情况
train_old_user_index = old_user_data_train.loc[
old_user_data_train.user_id_cvr_smooth != -1, :].index
test_old_user_index = old_user_data_test.loc[
old_user_data_test.user_id_cvr_smooth != -1, :].index
# train_new_user_index = old_user_data_train.loc[old_user_data_train.user_id_cvr_smooth==-1,:].index
# test_new_user_index = old_user_data_test.loc[old_user_data_test.user_id_cvr_smooth==-1,:].index
train_old_score = cv_preds[train_old_user_index]
test_old_score = test_preds[test_old_user_index]
# train_new_score = cv_preds[train_new_user_index]
# test_new_score = test_preds[test_new_user_index]
new_train_data = old_user_data_train.loc[
old_user_data_train.user_id_cvr_smooth != -1, :]
new_test_data = old_user_data_test.loc[
old_user_data_test.user_id_cvr_smooth != -1, :]
new_train_data['y'] = train_old_score
new_test_data['y'] = test_old_score
new_train_Y = train_Y[train_old_user_index]
#对老用户单独训练
clf = LogisticRegression(C=12,
fit_intercept=True,
max_iter=3000,
class_weight={
0: 0.5,
1: 0.5
})
clf.fit(X=new_train_data.values, y=new_train_Y)
train_LR_score = clf.predict_proba(new_train_data.values)[:, 1]
test_LR_score = clf.predict_proba(new_test_data.values)[:, 1]
cv_preds[train_old_user_index] = train_LR_score
test_preds[test_old_user_index] = test_LR_score
#记录老用户的损失情况
print('LR train:', cal_log_loss(train_LR_score, new_train_Y))
#拼接结果看看总体的损失情况
print('All train:', cal_log_loss(cv_preds, train_Y))
submmit_result(test_preds, 'old_and_new')
return test_preds, feat_imp
clf = LinearRegression(fit_intercept=True, normalize=True, n_jobs=-1)
clf.fit(X=predict_day7, y=train_Y)
predict_train_2 = clf.predict(predict_day7)
predict_cv_2 = clf.predict(predict_cv)
print('train:', cal_log_loss(predict_train_2, train_Y))
print('test:', cal_log_loss(predict_cv_2, cv_Y))
print('train mean:', np.mean(predict_train_2))
print('cv mean:', np.mean(predict_cv_2))
return predict_cv_2
if __name__ == '__main__':
t0 = time.time()
day7_data = _load_splited_df(path=cache_pkl_path + 'Tree_day/train_7')
cv_data = _load_splited_df(path=cache_pkl_path + 'Tree_day/cv')
test_data = _load_splited_df(path=cache_pkl_path + 'Tree_day/test')
print('off line')
_ = XGB_model_second(day7_data, cv_data)
print('on line')
# day7_data = pd.concat([day7_data, cv_data],axis=0)
predict_cv_2 = XGB_model_second(day7_data, test_data)
submmit_result(predict_cv_2, 'XGB_LR')
gbc.fit(train_data.values, train_Y)
predict_train = gbc.predict_proba(train_data.values)[:,1]
predict_cv = gbc.predict_proba(cv_data.values)[:,1]
predict_test = gbc.predict_proba(test_data.values)[:,1]
# print(gbc.get_params)
print('训练损失:',cal_log_loss(predict_train, train_Y))
print('测试损失:',cal_log_loss(predict_cv, cv_Y))
t1 = time.time()
print('训练用时:',t1-t0)
new_train, new_cv, new_test = gen_gbdt_feature(gbc, train_data, cv_data, test_data)
print('train shap:',new_train.shape)
print('cv shape', new_cv.shape)
print('test shape', new_test.shape)
# LR预测
clf = LogisticRegression(C=0.8, fit_intercept=True, max_iter=3000,class_weight={0:0.5, 1:0.5})
clf.fit(X=new_train, y=np.squeeze(train_Y))
predict_train = clf.predict_proba(new_train)[:,1]
predict_cv = clf.predict_proba(new_cv)[:,1]
predict_test = clf.predict_proba(new_test)[:,1]
print('训练损失:',cal_log_loss(predict_train, train_Y))
print('测试损失:',cal_log_loss(predict_cv, cv_Y))
t1 = time.time()
print('训练用时:',t1-t0)
submmit_result(predict_test,'GBDT_LR')
# -*- coding: utf-8 -*-
"""
Created on Mon Apr 16 22:13:45 2018
@author: weiqing
"""
import pandas as pd
import numpy as np
from utils import load_pickle, raw_data_path, feature_data_path, cache_pkl_path, result_path, model_path, submmit_result
if __name__ == '__main__':
XGB = pd.read_csv('../result/XGB_20180421_211412.txt', sep=' ')
LGB = pd.read_csv('../result/LGB_20180421_172434.txt', sep=' ')
FFM = pd.read_csv('../result/FFM_20180421_215653.txt', sep=' ')
result = np.zeros((len(XGB), 3))
result[:, 0] = XGB['predicted_score'].values
result[:, 1] = LGB['predicted_score'].values
result[:, 2] = FFM['predicted_score'].values
median = np.median(result, axis=1)
submmit_result(median, 'median')
def change_to_result():
preds = pd.read_csv('../result/ffm_online_result.csv', header=None)
submmit_result(np.squeeze(preds.values), 'FFM')
if __name__ == '__main__':
t0 = time.time()
day7_data = _load_splited_df(path=cache_pkl_path + 'LR_day/train_7')
cv_data = _load_splited_df(path=cache_pkl_path + 'LR_day/cv')
test_data = _load_splited_df(path=cache_pkl_path + 'LR_day/test')
predict_day7, predict_cv, predict_test = LR_model_first(
day7_data, cv_data, test_data)
#model 2
train_Y = day7_data['is_trade'].values
cv_Y = cv_data['is_trade'].values
clf = LinearRegression(fit_intercept=True, normalize=True, n_jobs=-1)
clf.fit(X=predict_day7, y=train_Y)
predict_train_2 = clf.predict(predict_day7)
predict_cv_2 = clf.predict(predict_cv)
predict_test_2 = clf.predict(predict_test)
print('train:', cal_log_loss(predict_train_2, train_Y))
print('test:', cal_log_loss(predict_cv_2, cv_Y))
print('train mean:', np.mean(predict_train_2))
print('cv mean:', np.mean(predict_cv_2))
print('test mean:', np.mean(predict_test_2))
submmit_result(predict_test_2, 'LR_LR')
