147 lines
7.3 KiB
Python
147 lines
7.3 KiB
Python
import sys
|
||
sys.path.append("..")
|
||
import random
|
||
import numpy as np
|
||
from tqdm import tqdm
|
||
import pandas as pd
|
||
from sklearn.preprocessing import LabelEncoder
|
||
from features import DenseFeat, SparseFeat, VarLenSparseFeat
|
||
from models import DSSM
|
||
import tensorflow as tf
|
||
from tensorflow.keras.preprocessing.sequence import pad_sequences
|
||
|
||
def process_data(data_path, max_len=50, neg_sample=0, type='random'):
|
||
"""读取数据
|
||
"""
|
||
data_df = pd.read_csv(data_path, sep=',')
|
||
|
||
"""
|
||
print(data_df.head())
|
||
user_id movie_id rating timestamp title genres gender age occupation zip
|
||
0 1 1193 5 978300760 One Flew Over the Cuckoo's Nest (1975) Drama F 1 10 48067
|
||
1 1 661 3 978302109 James and the Giant Peach (1996) Animation|Children's|Musical F 1 10 48067
|
||
2 1 914 3 978301968 My Fair Lady (1964) Musical|Romance F 1 10 48067
|
||
3 1 3408 4 978300275 Erin Brockovich (2000) Drama F 1 10 48067
|
||
4 1 2355 5 978824291 Bug's Life, A (1998) Animation|Children's|Comedy F 1 10 48067
|
||
"""
|
||
use_features = ['user_id', 'movie_id', 'gender', 'age', 'occupation', 'zip']
|
||
|
||
# 特征转换, 类别编码
|
||
feature_max_index_dict = {}
|
||
for feat in use_features:
|
||
lbe = LabelEncoder()
|
||
data_df[feat] = lbe.fit_transform(data_df[feat]) + 1 # 让id从1开始，0可能会被做掩码
|
||
feature_max_index_dict[feat] = data_df[feat].max() + 1
|
||
|
||
user_profile = data_df[["user_id", "gender", "age", "occupation", "zip"]].drop_duplicates('user_id')
|
||
item_profile = data_df[["movie_id"]].drop_duplicates('movie_id')
|
||
|
||
print(user_profile.head())
|
||
|
||
# 构建数据标签
|
||
# 将数据按照时间进行排序, 默认是升序
|
||
data_df.sort_values("timestamp", inplace=True)
|
||
|
||
unique_item_ids = data_df['movie_id'].unique()
|
||
|
||
train_data_list = []
|
||
test_data_list = []
|
||
# 设置最短历史序列长度
|
||
min_seq_len = 1
|
||
# 遍历每个用户构建正负样本
|
||
for user_id, hist_df in data_df.groupby('user_id'):
|
||
pos_list = hist_df['movie_id'].to_list()
|
||
rating_list = hist_df['rating'].to_list()
|
||
|
||
if neg_sample > 0:
|
||
candidate_list = list(set(unique_item_ids) - set(pos_list))
|
||
# 每个正样本对应多个负样本
|
||
neg_list = np.random.choice(candidate_list, size=len(pos_list) * neg_sample,
|
||
replace=True) # 设置为True表示可以取相同的元素
|
||
|
||
# 历史序列长度最少为1，这里也可以设置为更长
|
||
for i in range(min_seq_len, len(pos_list)):
|
||
hist = pos_list[:i]
|
||
if i != len(pos_list) - min_seq_len:
|
||
# 构造样本
|
||
# 样本的数据格式：user_id, hist_movid_id, pos_movie_id, label, hist_len
|
||
train_data_list.append((user_id, hist[::-1], pos_list[i], 1, len(hist[::-1]), rating_list[i]))
|
||
for negi in range(neg_sample):
|
||
train_data_list.append((user_id, hist[::-1], neg_list[i*neg_sample + i], 0, len(hist[::-1]), rating_list[i]))
|
||
else:
|
||
test_data_list.append((user_id, hist[::-1], pos_list[i], 1, len(hist[::-1]), rating_list[i]))
|
||
|
||
random.shuffle(train_data_list)
|
||
random.shuffle(test_data_list)
|
||
|
||
# 将输入的特征转换成字典的形式
|
||
train_data_dict = {}
|
||
test_data_dict = {}
|
||
|
||
# 构建训练集数据
|
||
train_data_dict['user_id'] = np.array([line[0] for line in train_data_list])
|
||
# 由于这个是不定长，所以不能直接转np，需要先对其进行padding
|
||
# train_hist_movie_id = [line[1] for line in train_data_list]
|
||
# train_data_dict['hist_movie_ids'] = pad_sequences(train_hist_movie_id,
|
||
# maxlen=max_len, padding='post', truncating='post', value=0)
|
||
train_data_dict['movie_id'] = np.array([line[2] for line in train_data_list])
|
||
train_data_dict['label'] = np.array([line[3] for line in train_data_list])
|
||
# train_data_dict['hist_len'] = np.array([line[4] for line in train_data_list])
|
||
# train_data_dict['rating'] = np.array([line[5] for line in train_data_list])
|
||
for key in ["gender", "age", "occupation", "zip"]:
|
||
# 将样本中所有user_id对应的其他的特征都索引到
|
||
tmp_list = []
|
||
for i in range(len(train_data_dict['user_id'])):
|
||
tmp_list.append(user_profile[user_profile['user_id'] ==
|
||
train_data_dict['user_id'][i]][key])
|
||
train_data_dict[key] = np.array(tmp_list)
|
||
|
||
# 构建测试集数据
|
||
test_data_dict['user_id'] = np.array([line[0] for line in test_data_list])
|
||
# 由于这个是不定长，所以不能直接转np，需要先对其进行padding
|
||
# train_hist_movie_id = [line[1] for line in test_data_list]
|
||
# test_data_dict['hist_movie_ids'] = pad_sequences(train_hist_movie_id,
|
||
# maxlen=max_len, padding='post', truncating='post', value=0)
|
||
test_data_dict['movie_id'] = np.array([line[2] for line in test_data_list])
|
||
test_data_dict['label'] = np.array([line[3] for line in test_data_list])
|
||
# test_data_dict['hist_len'] = np.array([line[4] for line in test_data_list])
|
||
# test_data_dict['rating'] = np.array([line[5] for line in test_data_list])
|
||
|
||
for key in ["gender", "age", "occupation", "zip"]:
|
||
# 将样本中所有user_id对应的其他的特征都索引到
|
||
tmp_list = []
|
||
for i in range(len(test_data_dict['user_id'])):
|
||
tmp_list.append(user_profile[user_profile['user_id'] ==
|
||
test_data_dict['user_id'][i]][key])
|
||
test_data_dict[key] = np.array(tmp_list)
|
||
|
||
|
||
return feature_max_index_dict, train_data_dict, test_data_dict
|
||
|
||
|
||
if __name__ == "__main__":
|
||
data_path = '/home/ryluo/recsys/data/movielens_sample.txt'
|
||
|
||
feature_max_index_dict, train_data_dict, test_data_dict = process_data(data_path)
|
||
train_label = train_data_dict['label']
|
||
print(train_data_dict.keys())
|
||
train_data_dict.pop("label")
|
||
embedding_dim = 4
|
||
user_feature_columns = [SparseFeat('user_id', feature_max_index_dict['user_id'], embedding_dim),
|
||
SparseFeat("gender", feature_max_index_dict['gender'], embedding_dim),
|
||
SparseFeat("age", feature_max_index_dict['age'], embedding_dim),
|
||
SparseFeat("occupation", feature_max_index_dict['occupation'], embedding_dim),
|
||
SparseFeat("zip", feature_max_index_dict['zip'], embedding_dim),
|
||
# VarLenSparseFeat(SparseFeat('hist_movie_id', feature_max_idx['movie_id'], embedding_dim,
|
||
# embedding_name="movie_id"), SEQ_LEN, 'mean', 'hist_len'),
|
||
]
|
||
|
||
item_feature_columns = [SparseFeat('movie_id', feature_max_index_dict['movie_id'], embedding_dim)]
|
||
print(user_feature_columns + item_feature_columns)
|
||
|
||
model = DSSM(user_feature_columns, item_feature_columns)
|
||
|
||
model.compile(optimizer="adam", loss='binary_crossentropy', metrics=[tf.keras.metrics.AUC])
|
||
|
||
model.fit(train_data_dict, train_label, # train_label,
|
||
batch_size=256, epochs=1, verbose=1, validation_split=0.1, ) |