-
Notifications
You must be signed in to change notification settings - Fork 0
/
pmf-sgd-netflix.R
151 lines (133 loc) · 5.8 KB
/
pmf-sgd-netflix.R
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
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
rm(list=ls())
setwd("C:/Users/Lenovo/Code/R/r_pmf_paper")
data=read.table("netflix.txt", header = FALSE)
rating=data[,3]
movie=data[,2]
user=data[,1]
data_num=length(rating)
movie_num= max(movie)#电影数量
user_num=max(user)#用户数量
#parametre
epsilon=50 # Learning rate 学习率
lambda = 0.1#Regularization parameter正则化参数
momentum=0.7#动量优化参数
epoch=1#初始化epoch
maxepoch=50#总训练次数
num_feat = 5 #Rank 10 decomposition 隐因子数量
err_train1=rep(0,maxepoch)
err_valid1=rep(0,maxepoch)
err_random=rep(0,maxepoch)
train_num=(data_num*9)%/%10;
train_vec=data[1:train_num,];#训练集
probe_vec=data[train_num:data_num,];#测试集
movie_num= max(movie)#电影数量1682
user_num=max(user)#用户数量943
#sum_list=max(data)
mean_rating = mean(train_vec[,3])#平均评级
pairs_tr = length((train_vec[,3]))#training data 训练集长度
pairs_pr = length((probe_vec[,3]))#validation data 验证集长度
numbatches= 9 #Number of batches把数据分为9份
num_m = movie_num # Number of movies 电影数量
num_p = user_num #Number of users 用户数量
#初始化
w1_M1 = 0.1*matrix(runif(num_m*num_feat),nrow = num_m , ncol = num_feat ,byrow =T) # Movie feature vectors 生成用户物品特征矩阵d=10
w1_P1 = 0.1*matrix(runif(num_p*num_feat),nrow = num_p , ncol = num_feat ,byrow =T) # User feature vecators
w1_M1_inc = matrix(rep(0,num_m*num_feat),nrow = num_m , ncol = num_feat ,byrow =T)#生成同shape的全零矩阵
w1_P1_inc = matrix(rep(0,num_p*num_feat),nrow = num_p , ncol = num_feat ,byrow =T)
for(epoch in epoch:maxepoch){
#采用mini batch的方法，每次训练9个样本
for(batch in 1:numbatches){
N=10000 #number training triplets per batch 每次训练三元组的数量
aa_p= train_vec[((batch-1)*N+1):(batch*N),1]#读取用户列每次读取一万个
aa_m= train_vec[((batch-1)*N+1):(batch*N),2]#读取电影列
rating = train_vec[((batch-1)*N+1):(batch*N),3]#读取评级列
rating = rating-mean_rating; #Default prediction is the mean rating.
# Compute Predictions %%%%%%%%%%%%%%%%%
pred_out= apply(w1_M1[aa_m,]*w1_P1[aa_p,],1,sum)#每一行进行求和,.*为对应元素相乘 10k*1，用隐特征矩阵相乘得到1w个预测评级
f=sum((pred_out - rating)^2 + 0.5*lambda*(apply((w1_M1[aa_m,]^2 + w1_P1[aa_p,]^2),1,sum) ))
#求出损失函数
#Compute Gradients %%%%%%%%%%%%%%%%%%%迭代
IO =2*(pred_out - rating)
kkkk=num_feat-1
for(kkk in 1:kkkk){
IO =cbind(IO,2*(pred_out - rating))
}
#IO = repmat(2*(pred_out - rating),1,num_feat)
#将损失矩阵的二倍，复制10列
Ix_m=IO*w1_P1[aa_p,] + lambda*w1_M1[aa_m,]#损失*U-lambda*V 就是更新规则
Ix_p=IO*w1_M1[aa_m,] + lambda*w1_P1[aa_p,]#损失*V-lambda*U
dw1_M1 = matrix(rep(0,num_m*num_feat),nrow = num_m , ncol = num_feat ,byrow =T)
dw1_P1 = matrix(rep(0,num_p*num_feat),nrow = num_p , ncol = num_feat ,byrow =T)#生成全零用户特征矩阵
for(ii in 1:N){#迭代一万次 每一行一行来 得到更新矩阵
dw1_M1[aa_m[ii],]= dw1_M1[aa_m[ii],] + Ix_m[ii,]
dw1_P1[aa_p[ii],]= dw1_P1[aa_p[ii],] + Ix_p[ii,]
}
# Update movie and user features %%%%%%%%%%%
w1_M1_inc = epsilon*dw1_M1/N
w1_M1 = w1_M1 - w1_M1_inc#原矩阵-负导数*学习率
w1_P1_inc = epsilon*dw1_P1/N
w1_P1 = w1_P1 - w1_P1_inc
}
#此时所有(9轮)batch结束
#现在已经得到了此轮epoch后的一组U和V
# Compute Predictions after Paramete Updates %%%%%%%%%%%%%%%%%
pred_out= apply(w1_M1[aa_m,]*w1_P1[aa_p,],1,sum)
f_s=sum((pred_out - rating)^2 + 0.5*lambda*(apply((w1_M1[aa_m,]^2 + w1_P1[aa_p,]^2),1,sum) ))
err_train1[epoch] = sqrt(f_s/N)
#Compute predictions on the validation set %%%%%%%%%%%%%%%%%%%%%%
NN=pairs_pr#验证集长度
aa_p = probe_vec[,1]#读取验证集的user、movie、rating
aa_m = probe_vec[,2]
rating = probe_vec[,3]
pred_out =apply(w1_M1[aa_m,]*w1_P1[aa_p,],1,sum) + mean_rating#预测结果加上mean才行
pred_out[pred_out>5]=5
pred_out[pred_out<1]=1 #使得预测结果超过评分区间的值依旧掉在区间内
err_valid1[epoch]= sqrt(sum((pred_out- rating)^2)/NN)
RMSE=sqrt(sum((pred_out- rating)^2)/NN)
MAE=sum(abs(pred_out- rating))/NN
print(paste('epoch',epoch,'Train RMSE',signif(err_train1[epoch], 4),'Test RMSE',signif(err_valid1[epoch], 4)))
}
print(paste('RMSE',signif(RMSE, 4),'MAE',signif(MAE, 4)))
#计算Pre@5 Re@5
# TOP-N 推荐
j=10
txt=numeric(user_num)
for(i in 1:user_num){
user_i_rating_real=probe_vec[(probe_vec$V1==i),]
user_i_rating_real=user_i_rating_real[order(user_i_rating_real$V3,decreasing = TRUE),]
user_i_rating=w1_P1[i,]%*%t(w1_M1[user_i_rating_real$V2,])+mean_rating
if(length(user_i_rating_real$V2)>j){
txt[i]=sum(rank(-1*user_i_rating)[1:j]<=j)/j
}
if(length(user_i_rating_real$V2)<=j){
ti=sum(user_i_rating_real$V3>=4)
if(ti!=0){
txt[i]=sum(user_i_rating[1:ti]>=4)/ti
}
}
}
Pre=mean(txt)
txt=numeric(user_num)
for(i in 1:user_num){
user_i_rating_real=probe_vec[(probe_vec$V1==i),]
user_i_rating_real=user_i_rating_real[order(user_i_rating_real$V3,decreasing = TRUE),]
user_i_rating=w1_P1[i,]%*%t(w1_M1[user_i_rating_real$V2,])+mean_rating
if(length(user_i_rating_real$V2)>j){
user_i_rating[user_i_rating<4]=0
user_i_rating[user_i_rating>4]=1
ti=sum(user_i_rating)
if(ti!=0){
bigerthan4=sum(user_i_rating_real$V3>=4)
tinri=sum(user_i_rating[1:bigerthan4])
txt[i]=tinri/ti
}
}
if(length(user_i_rating_real$V2)<=j){
ti=sum(user_i_rating_real$V3>=4)
if(ti!=0){
txt[i]=sum(user_i_rating[1:ti]>=4)/ti
}
}
}
Re=mean(txt)
print(paste('Re',signif(Re, 4),'Pre',signif(Pre, 4)))