项目难点

•  merchant:  商人
• 重命名列名:  user_log.rename(columns={'seller_id':'merchant_id'}, inplace=True)
• 数据类型转换:  user_log['item_id'] = user_log['item_id'].astype('int32')
• 主要使用方法: xgboost, lightbm

• 竞赛地址:  天猫复购预测之挑战Baseline_学习赛_天池大赛-阿里云天池
• 排名: 448/9361     score: 0.680989

项目简介:

阿里巴巴天池天猫复购预测的机器学习项目, 使用数据分析, 通过机器学习中的线性分类算法, 进行建模, 从而预测消费者行为, 复购情况 .

• 数据分析
• 特征工程
• 算法使用
• 算法集成

1  数据处理

1.1  模型导入

import gc    # 垃圾回收
import pandas as pd
import numpy as np
import warnings
warnings.filterwarnings('ignore')# 导入分析库
# 数据拆分
from sklearn.model_selection import train_test_split
# 同分布数据拆分
from sklearn.model_selection import StratifiedGroupKFold
import lightgbm as lgb
import xgboost as xgb

1.2  加载数据

%%time
# 加载数据
# 用户行为日志
user_log = pd.read_csv('./data_format1/user_log_format1.csv', dtype = {'time_stamp':'str'})
# 用户画像
user_info = pd.read_csv('./data_format1/user_info_format1.csv')
# 训练数据和测试数据
train_data = pd.read_csv('./data_format1/train_format1.csv')
test_data = pd.read_csv('./data_format1/test_format1.csv')

1.3  查看数据

print('---data shape---')     
for data in [user_log, user_info, train_data, test_data]:print(data.shape)

print('---data info ---')
for data in [user_log, user_info, train_data, test_data]:print(data.info())

display(user_info.head())

display(train_data.head(),test_data.head())

1.4  数据集成

train_data['origin'] = 'train'
test_data['origin'] = 'test'
# 集成
all_data = pd.concat([train_data, test_data], ignore_index=True, sort=False)
# prob测试数据中特有的一列
all_data.drop(['prob'], axis=1, inplace=True) # 删除概率这一列
display(all_data.head(),all_data.shape)

# 连接user_info表，通过user_id关联
all_data = all_data.merge(user_info, on='user_id', how='left')
display(all_data.shape,all_data.head())

# 使用 merchant_id（原列名seller_id）
user_log.rename(columns={'seller_id':'merchant_id'}, inplace=True)

del train_data,test_data,user_info
gc.collect()

1.5  数据类型转换

%%time
display(user_log.info())

%%time
display(user_log.head())

%%time
# 用户行为数据类型转换
user_log['user_id'] = user_log['user_id'].astype('int32')
user_log['merchant_id'] = user_log['merchant_id'].astype('int32')
user_log['item_id'] = user_log['item_id'].astype('int32')
user_log['cat_id'] = user_log['cat_id'].astype('int32')
user_log['brand_id'].fillna(0, inplace=True)
user_log['brand_id'] = user_log['brand_id'].astype('int32')
user_log['time_stamp'] = pd.to_datetime(user_log['time_stamp'], format='%H%M')
user_log['action_type'] = user_log['action_type'].astype('int32')
display(user_log.info(),user_log.head())

display(all_data.isnull().sum())

# 缺失值填充
all_data['age_range'].fillna(0, inplace=True)
all_data['gender'].fillna(2, inplace=True)
all_data.isnull().sum()

all_data.info()

all_data['age_range'] = all_data['age_range'].astype('int8')
all_data['gender'] = all_data['gender'].astype('int8')
all_data['label'] = all_data['label'].astype('str')
all_data['user_id'] = all_data['user_id'].astype('int32')
all_data['merchant_id'] = all_data['merchant_id'].astype('int32')
all_data.info()

1.6  用户特征工程(5min)

%%time
##### 特征处理
##### User特征处理
groups = user_log.groupby(['user_id'])# 用户交互行为数量 u1
temp = groups.size().reset_index().rename(columns={0:'u1'})
all_data = all_data.merge(temp, on='user_id', how='left')# 细分
# 使用 agg 基于列的聚合操作，统计唯一值个数 item_id, cat_id, merchant_id, brand_id
# 用户，交互行为：点了多少商品呢？
temp = groups['item_id'].agg([('u2', 'nunique')]).reset_index()
all_data = all_data.merge(temp, on='user_id', how='left')# 用户，交互行为，具体统计：类目多少
temp = groups['cat_id'].agg([('u3', 'nunique')]).reset_index()
all_data = all_data.merge(temp, on='user_id', how='left')temp = groups['merchant_id'].agg([('u4', 'nunique')]).reset_index()
all_data = all_data.merge(temp, on='user_id', how='left')temp = groups['brand_id'].agg([('u5', 'nunique')]).reset_index()
all_data = all_data.merge(temp, on='user_id', how='left')# 购物时间间隔特征 u6 按照小时
temp = groups['time_stamp'].agg([('F_time', 'min'), ('B_time', 'max')]).reset_index()
temp['u6'] = (temp['B_time'] - temp['F_time']).dt.seconds/3600
all_data = all_data.merge(temp[['user_id', 'u6']], on='user_id', how='left')# 统计操作类型为0，1，2，3的个数
temp = groups['action_type'].value_counts().unstack().reset_index().rename(columns={0:'u7', 1:'u8', 2:'u9', 3:'u10'})
all_data = all_data.merge(temp, on='user_id', how='left')del temp,groups
gc.collect()

all_data.head()

1.7  店铺特征工程(5min)

%%time
##### 商家特征处理
groups = user_log.groupby(['merchant_id'])# 商家被交互行为数量 m1
temp = groups.size().reset_index().rename(columns={0:'m1'})
all_data = all_data.merge(temp, on='merchant_id', how='left')# 统计商家被交互的 user_id, item_id, cat_id, brand_id 唯一值
temp = groups['user_id', 'item_id', 'cat_id', 'brand_id'].nunique().reset_index().rename(columns={'user_id':'m2','item_id':'m3', 'cat_id':'m4', 'brand_id':'m5'})
all_data = all_data.merge(temp, on='merchant_id', how='left')# 统计商家被交互的 action_type 唯一值
temp = groups['action_type'].value_counts().unstack().reset_index().rename(  columns={0:'m6', 1:'m7', 2:'m8', 3:'m9'})
all_data = all_data.merge(temp, on='merchant_id', how='left')del temp
gc.collect()

display(all_data.tail())

1.8  用户和店铺联合特征工程(4min)

%%time
##### 用户+商户特征
groups = user_log.groupby(['user_id', 'merchant_id'])# 用户在不同商家交互统计
temp = groups.size().reset_index().rename(columns={0:'um1'})
all_data = all_data.merge(temp, on=['user_id', 'merchant_id'], how='left')# 统计用户在不同商家交互的 item_id, cat_id, brand_id 唯一值
temp = groups['item_id', 'cat_id', 'brand_id'].nunique().reset_index().rename(columns={'item_id':'um2','cat_id':'um3','brand_id':'um4'})
all_data = all_data.merge(temp, on=['user_id', 'merchant_id'], how='left')# 统计用户在不同商家交互的 action_type 唯一值
temp = groups['action_type'].value_counts().unstack().reset_index().rename(columns={0:'um5',1:'um6',2:'um7',3:'um8'})
all_data = all_data.merge(temp, on=['user_id', 'merchant_id'], how='left')# 统计用户在不同商家购物时间间隔特征 um9 按照小时
temp = groups['time_stamp'].agg([('F_time', 'min'), ('B_time', 'max')]).reset_index()
temp['um9'] = (temp['B_time'] - temp['F_time']).dt.seconds/3600
all_data = all_data.merge(temp[['user_id','merchant_id','um9']], on=['user_id', 'merchant_id'], how='left')del temp,groups
gc.collect()

display(all_data.head())

1.9  购买点击比

all_data['r1'] = all_data['u9']/all_data['u7']    # 用户购买点击比
all_data['r2'] = all_data['m8']/all_data['m6']    # 商家购买点击比
all_data['r3'] = all_data['um7']/all_data['um5']  # 不同用户不同商家购买点击比
display(all_data.head())

1.10  空数据填充

display(all_data.isnull().sum())

all_data.fillna(0, inplace=True)
all_data.isnull().sum()

1.11  年龄性别类别型转换

all_data['age_range']

%%time
# 修改age_range字段名称为 age_0, age_1, age_2... age_8
# 独立编码
temp = pd.get_dummies(all_data['age_range'], prefix='age')
display(temp.head(10))
all_data = pd.concat([all_data, temp], axis=1)

# 性别转换
temp = pd.get_dummies(all_data['gender'], prefix='g')
all_data = pd.concat([all_data, temp], axis=1) # 列进行合并# 删除原数据
all_data.drop(['age_range', 'gender'], axis=1, inplace=True)del temp
gc.collect()

all_data.head()

1.12  数据存储

%%time
# train_data、test-data
train_data = all_data[all_data['origin'] == 'train'].drop(['origin'], axis=1)
test_data = all_data[all_data['origin'] == 'test'].drop(['label', 'origin'], axis=1)train_data.to_csv('train_data.csv')
test_data.to_csv('test_data.csv')

2  算法建模预测

# 训练数据和目标值
train_X, train_y = train_data.drop(['label'], axis=1), train_data['label']# 数据拆分保留20%作为测试数据
X_train, X_valid, y_train, y_valid = train_test_split(train_X, train_y, test_size=.2)

2.1  LGB 模型

def lgb_train(X_train, y_train, X_valid, y_valid, verbose=True):model_lgb = lgb.LGBMClassifier(max_depth=10, # 8 # 树最大的深度n_estimators=5000, # 集成算法，树数量min_child_weight=100, colsample_bytree=0.7, # 特征筛选subsample=0.9,  # 样本采样比例learning_rate=0.1) # 学习率model_lgb.fit(X_train, y_train,eval_metric='auc',eval_set=[(X_train, y_train), (X_valid, y_valid)],verbose=verbose, # 是否打印输出训练过程early_stopping_rounds=10) # 早停，等10轮决策，评价指标不在变化，停止print(model_lgb.best_score_['valid_1']['auc'])return model_lgb

X_train

model_lgb = lgb_train(X_train.values, y_train, X_valid.values, y_valid, verbose=True)

%%time
prob = model_lgb.predict_proba(test_data.values) # 预测
submission = pd.read_csv('./data_format1/test_format1.csv')# 复购的概率
submission['prob'] = pd.Series(prob[:,1]) # 预测数据赋值给提交数据
display(submission.head())
submission.to_csv('submission_lgb.csv', index=False)del submission
gc.collect()

2.2  XGB 模型

def xgb_train(X_train, y_train, X_valid, y_valid, verbose=True):model_xgb = xgb.XGBClassifier(max_depth=10, # raw8n_estimators=5000,min_child_weight=300, colsample_bytree=0.7, subsample=0.9, learing_rate=0.1)model_xgb.fit(X_train, y_train,eval_metric='auc',eval_set=[(X_train, y_train), (X_valid, y_valid)],verbose=verbose,early_stopping_rounds=10)  # 早停法，如果auc在10epoch没有进步就stopprint(model_xgb.best_score)return model_xgb

模型训练

model_xgb = xgb_train(X_train, y_train, X_valid, y_valid, verbose=False)

模型预测

%%time
prob = model_xgb.predict_proba(test_data)
submission = pd.read_csv('./data_format1/test_format1.csv')
submission['prob'] = pd.Series(prob[:,1])
submission.to_csv('submission_xgb.csv', index=False)
display(submission.head())
del submission
gc.collect()

3  交叉验证多轮建模

# 构造训练集和测试集
def get_train_test_datas(train_df,label_df):skv = StratifiedKFold(n_splits=10, shuffle=True)trainX = []trainY = []testX = []testY = []# 索引：训练数据索引train_index,目标值的索引test_indexfor train_index, test_index in skv.split(X=train_df, y=label_df):  # 10轮for循环train_x, train_y, test_x, test_y = train_df.iloc[train_index, :], label_df.iloc[train_index], \train_df.iloc[test_index, :], label_df.iloc[test_index]trainX.append(train_x)trainY.append(train_y)testX.append(test_x)testY.append(test_y)return trainX, testX, trainY, testY

3.1  LGB 模型（1min）

%%time
train_X, train_y = train_data.drop(['label'], axis=1), train_data['label']# 拆分为10份训练数据和验证数据
X_train, X_valid, y_train, y_valid = get_train_test_datas(train_X, train_y)print('----训练数据，长度',len(X_train))
print('----验证数据，长度',len(X_valid))pred_lgbms = [] # 列表，接受目标值，10轮，平均值for i in range(10):print('\n=========LGB training use Data {}/10===========\n'.format(i+1))model_lgb = lgb.LGBMClassifier(max_depth=10, # 8n_estimators=1000,min_child_weight=100,colsample_bytree=0.7,subsample=0.9,learning_rate=0.05)model_lgb.fit(X_train[i].values, y_train[i],eval_metric='auc',eval_set=[(X_train[i].values, y_train[i]), (X_valid[i].values, y_valid[i])],verbose=False,early_stopping_rounds=10)print(model_lgb.best_score_['valid_1']['auc'])pred = model_lgb.predict_proba(test_data.values)pred = pd.DataFrame(pred[:,1]) # 将预测概率（复购）去处理，转换成DataFramepred_lgbms.append(pred)# 求10轮平均值生成预测结果，保存
# 每一轮的结果，作为一列，进行了添加
pred_lgbms = pd.concat(pred_lgbms, axis=1) # 级联，列进行级联# 加载提交数据
submission = pd.read_csv('./data_format1/test_format1.csv')
submission['prob'] = pred_lgbms.mean(axis=1) # 10轮训练的平均值
submission.to_csv('submission_KFold_lgb.csv', index=False)

3.2  XGB 模型（4min）

# 构造训练集和测试集
def get_train_test_datas(train_df,label_df):skv = StratifiedKFold(n_splits=20, shuffle=True)trainX = []trainY = []testX = []testY = []# 索引：训练数据索引train_index,目标值的索引test_indexfor train_index, test_index in skv.split(X=train_df, y=label_df):# 10轮for循环train_x, train_y, test_x, test_y = train_df.iloc[train_index, :], label_df.iloc[train_index], \train_df.iloc[test_index, :], label_df.iloc[test_index]trainX.append(train_x)trainY.append(train_y)testX.append(test_x)testY.append(test_y)return trainX, testX, trainY, testY

%%time
train_X, train_y = train_data.drop(['label'], axis=1), train_data['label']# 拆分为20份训练数据和验证数据
X_train, X_valid, y_train, y_valid = get_train_test_datas(train_X, train_y)print('------数据长度',len(X_train),len(y_train))pred_xgbs = []
for i in range(20):print('\n============XGB training use Data {}/20========\n'.format(i+1))model_xgb = xgb.XGBClassifier(max_depth=10, # raw8n_estimators=5000,min_child_weight=200, colsample_bytree=0.7, subsample=0.9,learning_rate = 0.1)model_xgb.fit(X_train[i], y_train[i],eval_metric='auc',eval_set=[(X_train[i], y_train[i]), (X_valid[i], y_valid[i])],verbose=False,early_stopping_rounds=10 # 早停法，如果auc在10epoch没有进步就stop)    print(model_xgb.best_score)pred = model_xgb.predict_proba(test_data)pred = pd.DataFrame(pred[:,1])pred_xgbs.append(pred)# 求20轮平均值生成预测结果，保存
pred_xgbs = pd.concat(pred_xgbs, axis=1)
submission = pd.read_csv('./data_format1/test_format1.csv')
submission['prob'] = pred_xgbs.mean(axis=1)
submission.to_csv('submission_KFold_xgb.csv', index=False)