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
| from svd import SVD
import numpy as np
import traceback
class SVDPlusPlus(SVD):
''' 基于SVD++的协同过滤算法 '''
def __init__(self, logger, trainset, historyset, topicNum, lr, gamma):
''' 初始化函数
Args:
logger:日志记录对象
trainset:训练集
historyset:历史数据集
topicNum:主题个数，奇异值数量
lr:学习率
gamma:正则项权重
'''
self.historyset = historyset
self.yItem = {}
super().__init__(logger, trainset, topicNum, lr, gamma)
def initModel(self):
''' 初始化模型参数：biasUser、biasItem、pUser、qItem、yItem '''
super().initModel()
for item in self.biasItem.keys():
self.yItem[item] = np.random.rand(self.topicNum) / 10
def train(self, iteration, topN):
''' 迭代多批次训练SVD模型
min power(rated - rate, 2)/2 + gamma * regularization/2
rate = mean + biasUser + biasItem + qItem * (pUser + yItemUserSum/sqrt(itemCounter))
itemCounter = sqrt(len(self.historyset[user].keys()))
yItemUserSum = sum(map(lambda key:self.yItem[key], self.historyset[user].keys()))
regularization = power(biasUser) + power(biasItem) + power(pUser) + power(qItem) + power(yItem)
故每次迭代：lr = learning rate
delta(biasUser) = lr * ((rate - rated) + gamma * biasUser)
delta(biasItem) = lr * ((rate - rated) + gamma * biasItem)
delta(qItem) = lr * ((rate - rated) * (pUser + yItemUserSum/itemCounter) + gamma * qItem)
delta(pUser) = lr * ((rate - rated) * qItem + gamma * pUser)
delta(yItem) = lr * (rate - rated) * qItem/itemCounter + gamma * yItem)
Args:
iteration:迭代次数
'''
self.logger.info('开始训练...')
sampleSize = len(self.trainMap)
sampleSeq = np.arange(sampleSize)
self.logger.info('训练前，误差为：')
self.evaluate(self.trainset, topN)
for i in range(iteration):
np.random.shuffle(sampleSeq)
for j in sampleSeq:
user, item = self.trainMap[j]
rate = self.predict(user, item)
rated = self.trainset[user][item]
itemCounter = np.sqrt(len(self.historyset[user].keys()))
yItemUserSum = sum(map(lambda key:self.yItem[key], self.historyset[user].keys()))
rateDiff = rate - rated
self.biasUser[user] -= (self.lr * (rateDiff + self.gamma * self.biasUser[user]))
self.biasItem[item] -= (self.lr * (rateDiff + self.gamma * self.biasItem[item]))
self.qItem[item] -= (self.lr * (rateDiff * (self.pUser[user] + yItemUserSum/itemCounter) +
self.gamma * self.qItem[item]))
self.pUser[user] -= (self.lr * (rateDiff * self.qItem[item] +
self.gamma * self.pUser[user]))
for historyItem in self.historyset[user].keys():
self.yItem[historyItem] -= (self.lr * (rateDiff * self.qItem[item]/itemCounter +
self.gamma * self.yItem[historyItem]))
if i % 10 == 0:
self.logger.info('训练进度[{i}/{iteration}]，训练误差为：'.format(i=i,
iteration=iteration))
self.evaluate(self.trainset, topN)
self.lr *= 0.5
self.logger.info('训练后，误差为：')
self.evaluate(self.trainset, topN)
self.logger.info('训练结束！')
def predict(self, user, item):
''' 测试SVD模型
rate = mean + biasUser + biasItem + qItem * (pUser + yItemUserSum/itemCounter)
itemCounter = sqrt(len(self.historyset[user].keys()))
yItemUserSum = sum(map(lambda key:self.yItem[key], self.historyset[user].keys()))
Args:
user:待预测评分的用户
item:待预测评分的物品
Returns:
rate:用户user对物品item的评分
'''
itemCounter = np.sqrt(len(self.historyset[user].keys()))
yItemUserSum = sum(map(lambda key:self.yItem[key], self.historyset[user].keys()))
try:
rate = (self.mean + self.biasUser[user] + self.biasItem[item] +
np.dot(self.qItem[item], self.pUser[user] + yItemUserSum/itemCounter))
except ZeroDivisionError as e:
traceback.print_exc()
print(user)
print(itemCounter)
raise e
return rate
|
