复购预测 java 复购模型

1.项目背景

电商经常会做大促，比如打折或者发放优惠券，尤其是在特殊的节日，比如黑色星期五，双十一等，目的是为了吸引大量的新用户。然而，许多吸引过来的购买者只是一次性交易，这些促销对于转换为长期的顾客来说可能收效甚微。提供了双十一促销期间的商家和新用户数据，目标是对于给定的商家，预测新用户是否能成为忠诚用户，即未来6个月内会购买这个商家的商品。

2.数据预处理

user_log

user_id	用户ID
item_id	商品ID
cat_id	品类ID
merchant_id	商家ID
brand_id	品牌ID
time_tamp	时间戳（格式: mmdd）
action_type	取值为{0, 1, 2, 3}, 0 代表点击，1 代表加入购物车，2 代表购买， 3 代表添加到收藏

user_info

user_id	用户ID
age_range	用户年龄范围： 1 代表 <18; 2 代表 [18,24]; 3 代表 [25,29]; 4 代表 [30,34]; 5 代表 [35,39]; 6 代表 [40,49]; 7 和 8 代表 >= 50; 0代表未知
gender	性别: 0 代表女, 1 代表男, 2和NULL 代表未知

 

训练测试数据

user_id	用户ID
merchant_id	商家ID
label	取值为{0, 1}, 1 代表重复购买者， 0 代表非重复购买者， 测试文件中该列为空

# 加载小样本
user_log = pd.read_csv('./data_format1_small/sample_user_log.csv', dtype={'time_stamp':'str'})
user_info = pd.read_csv('./data_format1_small/sample_user_info.csv')
train_data1 = pd.read_csv('./data_format1_small/train.csv')
submission = pd.read_csv('./data_format1_small/test.csv')
train_data = pd.read_csv('./data_format2/train_format2.csv')

train_data1['origin'] = 'train'
submission['origin'] = 'test'

#concat训练和测试数据
matrix = pd.concat([train_data1, submission], ignore_index=True, sort=False)


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

# 格式化
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')

对离散特征做LabelEncoder

be_merchant_id=LabelEncoder()
lbe_merchant_id.fit(np.r_[0,user_log['merchant_id'].values])
user_log['merchant_id']=lbe_merchant_id.transform(user_log['merchant_id'])
matrix['merchant_id']=lbe_merchant_id.transform(matrix['merchant_id'])

lbe_user_id=LabelEncoder()
user_log['user_id']=lbe_user_id.fit_transform(user_log['user_id'])
user_info['user_id']=lbe_user_id.transform(user_info['user_id'])
matrix['user_id']=lbe_user_id.transform(matrix['user_id'])

lbe_item_id=LabelEncoder()
user_log['item_id']=lbe_item_id.fit_transform(user_log['item_id'])
lbe_cat_id=LabelEncoder()
user_log['cat_id']=lbe_cat_id.fit_transform(user_log['cat_id'])
lbe_brand_id=LabelEncoder()
user_log['brand_id']=lbe_brand_id.fit_transform(user_log['brand_id'])

user_log['merchant_id'].max(),user_log['user_id'].max()
matrix = matrix.merge(user_info, on='user_id', how='left')

 

格式化matrix

matrix['age_range'].fillna(0, inplace=True)
# 0:female, 1:male, 2:unknown
matrix['gender'].fillna(2, inplace=True)
matrix['age_range'] = matrix['age_range'].astype('int8')
matrix['gender'] = matrix['gender'].astype('int8')
matrix['label'] = matrix['label'].astype('str')
matrix['user_id'] = matrix['user_id'].astype('int32')
matrix['merchant_id'] = matrix['merchant_id'].astype('int32')
del user_info, train_data1
gc.collect()
print(martix)

3.特征工程

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'})
matrix = matrix.merge(temp, on='user_id', how='left')

# 使用agg 基于列的聚合操作，统计唯一值的个数 item_id, cat_id, merchant_id, brand_id
#temp = groups['item_id', 'cat_id', 'merchant_id', 'brand_id'].nunique().reset_index().rename(columns={'item_id':'u2', 'cat_id':'u3', 'merchant_id':'u4', 'brand_id':'u5'})

temp = groups['item_id'].agg([('u2', 'nunique')]).reset_index()
matrix = matrix.merge(temp, on='user_id', how='left')
temp = groups['cat_id'].agg([('u3', 'nunique')]).reset_index()
matrix = matrix.merge(temp, on='user_id', how='left')
temp = groups['merchant_id'].agg([('u4', 'nunique')]).reset_index()
matrix = matrix.merge(temp, on='user_id', how='left')
temp = groups['brand_id'].agg([('u5', 'nunique')]).reset_index()
matrix = matrix.merge(temp, on='user_id', how='left')

# 时间间隔特征 u6 按照小时
temp = groups['time_stamp'].agg([('F_time', 'min'), ('L_time', 'max')]).reset_index()
temp['u6'] = (temp['L_time'] - temp['F_time']).dt.seconds/3600
matrix = matrix.merge(temp[['user_id', 'u6']], on='user_id', how='left')
# 统计action_type为0，1，2，3的个数（原始操作，没有补0）
temp = groups['action_type'].value_counts().unstack().reset_index().rename(columns={0:'u7', 1:'u8', 2:'u9', 3:'u10'})
matrix = matrix.merge(temp, on='user_id', how='left')
print(martix)

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'})
matrix = matrix.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'})
matrix = matrix.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'})
matrix = matrix.merge(temp, on='merchant_id', how='left')

# 按照merchant_id 统计随机负采样的个数
temp = train_data[train_data['label']==-1].groupby(['merchant_id']).size().reset_index().rename(columns={0:'m10'})
matrix = matrix.merge(temp, on='merchant_id', how='left')

# 按照user_id, merchant_id分组
groups = user_log.groupby(['user_id', 'merchant_id'])
temp = groups.size().reset_index().rename(columns={0:'um1'}) #统计行为个数
matrix = matrix.merge(temp, on=['user_id', 'merchant_id'], how='left')
temp = groups['item_id', 'cat_id', 'brand_id'].nunique().reset_index().rename(columns={'item_id':'um2', 'cat_id':'um3', 'brand_id':'um4'}) #统计item_id, cat_id, brand_id唯一个数
matrix = matrix.merge(temp, on=['user_id', 'merchant_id'], how='left')
temp = groups['action_type'].value_counts().unstack().reset_index().rename(columns={0:'um5', 1:'um6', 2:'um7', 3:'um8'})#统计不同action_type唯一个数
matrix = matrix.merge(temp, on=['user_id', 'merchant_id'], how='left')
temp = groups['time_stamp'].agg([('first', 'min'), ('last', 'max')]).reset_index()
temp['um9'] = (temp['last'] - temp['first']).dt.seconds/3600
temp.drop(['first', 'last'], axis=1, inplace=True)
print(temp)
print('-'*100)
matrix = matrix.merge(temp, on=['user_id', 'merchant_id'], how='left') #统计时间间隔
print(martix)

matrix特征处理

#用户购买点击比
matrix['r1'] = matrix['u9']/matrix['u7'] 
#商家购买点击比
matrix['r2'] = matrix['m8']/matrix['m6'] 
#不同用户不同商家购买点击比
matrix['r3'] = matrix['um7']/matrix['um5']
matrix.fillna(0, inplace=True)
# 修改age_range字段名称为 age_0, age_1, age_2... age_8
temp = pd.get_dummies(matrix['age_range'], prefix='age')
matrix = pd.concat([matrix, temp], axis=1)
temp = pd.get_dummies(matrix['gender'], prefix='g')
matrix = pd.concat([matrix, temp], axis=1)
matrix.drop(['age_range', 'gender'], axis=1, inplace=True)

lbe_action_type={0:1,1:2,2:3,3:4}
user_log['action_type']=user_log['action_type'].map(lbe_action_type)
# 用户行为sequence
# 把user_log里同user的这些数据合并成一个list
temp=pd.DataFrame(user_log.groupby('user_id')['merchant_id','action_type'].agg(lambda x:list(x)))
# 列名称改成hist_merchant_id 和 hist_action_type 
temp.columns=['hist_merchant_id','hist_action_type']
matrix = matrix.merge(temp, on=['user_id'], how='left') #统计时间间隔

# 截取，不缺到定长M个
M=500
for feature in ['hist_merchant_id','hist_action_type']:
    matrix[feature]=matrix[feature].map(lambda x:np.array(x+[0]*(M-len(x)))[:M])


# 分割训练数据和测试数据
train_data = matrix[matrix['origin'] == 'train'].drop(['origin'], axis=1)
test_data = matrix[matrix['origin'] == 'test'].drop(['label', 'origin'], axis=1)
train_X, train_y = train_data.drop(['label'], axis=1), train_data['label']

4.DIN模型

from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestClassifier
from sklearn.linear_model import LinearRegression
from sklearn.metrics import log_loss
from deepctr.feature_column import SparseFeat,VarLenSparseFeat,DenseFeat,get_feature_names
from deepctr.models import DIN, DIEN, DSIN
from sklearn.metrics import classification_report

train_X['action_type']=3
feature_columns = []
for column in train_X.columns:
  if column != 'hist_merchant_id' and column != 'hist_action_type':
    num = train_X[column].nunique()
    if num > 10000:
        dim = 10
    else:
        if num > 1000:
            dim = 8
        else:
            dim = 4

    if column  == 'user_id':
        feature_columns += [SparseFeat(column, 19111+1, embedding_dim=dim)]
    elif column  == 'merchant_id':
        feature_columns += [SparseFeat(column, 4994+1, embedding_dim=dim)]
    elif column  == 'action_type':
        feature_columns += [SparseFeat(column, 4+1, embedding_dim=dim)]
    else:
        feature_columns += [DenseFeat(column, 1)]


# maxlen为历史信息的长度，vocabulary_size为onehot的长度
feature_columns += [VarLenSparseFeat('hist_merchant_id', maxlen=M, vocabulary_size=19111+1, embedding_dim=8, embedding_name='merchant_id'),
                   VarLenSparseFeat('hist_action_type', maxlen=M, vocabulary_size=4+1, embedding_dim=4, embedding_name='action_type')]
hist_features=['merchant_id','action_type']
print(feature_columns)

# 使用DIN模型
model=DIN(feature_columns, hist_features)
# 使用Adam优化器，二分类的交叉熵
model.compile('adam', 'binary_crossentropy', metrics=['binary_crossentropy'])

# 组装train_model_input，得到feature names，将train_X转换为字典格式
feature_names=list(train_X.columns)
train_model_input = {name:train_X[name].values for name in feature_names}
# histroy输入必须是二维数组
from tqdm import tqdm
for fea in ['hist_merchant_id','hist_action_type']:
    l = []
    for i in tqdm(train_model_input[fea]):
        l.append(i)
    train_model_input[fea]=np.array(l)
history = model.fit(train_model_input, train_y, verbose=True, epochs=10, validation_split=0.2,batch_size=512)

# 转换test__model_input
test_data['action_type']=3
test_model_input = {name:test_data[name].values for name in feature_names}
from tqdm import tqdm
for fea in ['hist_merchant_id','hist_action_type']:
    l = []
    for i in tqdm(test_model_input[fea]):
        l.append(i)
    test_model_input[fea]=np.array(l)

# 得到预测结果
prob = model.predict(test_model_input)
submission['prob'] = prob
submission.drop(['origin'], axis=1, inplace=True)
submission.to_csv('prediction.csv', index=False)