Add files via upload

edtion1.0

implicit-feedback.Rmd

@@ -0,0 +1,269 @@
+
+
+```{r}
+#参数设置
+rm(list=ls())
+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=10#总训练次数
+err_train=rep(0,maxepoch)
+err_valid=rep(0,maxepoch)
+err_random=rep(0,maxepoch)
+num_feat = 10 #Rank 10 decomposition 隐因子数量
+```
+
+```{r}
+#导入数据
+data=read.csv("training.txt", header = FALSE,sep=' ')
+data[,3]=5
+rating=data[,3]
+movie=data[,2]
+user=data[,1]
+data_num=length(rating)
+movie_num= max(movie)#电影数量1682
+user_num=max(user)#用户数量943
+negative_sample=sample(1:user_num*movie_num,data_num,replace=TRUE)
+#负采样 同样数量的负样本
+record_user=rep(0,0)
+record_movie=rep(0,0)
+record_rating=rep(1,data_num)
+for (i in negative_sample){
+  movie_i=i%/%user_num
+  user_i=i%%user_num+1
+  record_user=append(record_user,user_i)
+  record_movie=append(record_movie,movie_i)
+}
+negative_data=data.frame(V1=record_user, V2=record_movie, V3=record_rating)
+```
+
+
+```{r}
+#融合正样本和负样本
+data=rbind(data,negative_data)
+#data=rbind(data,data)
+#
+rating=data[,3]
+movie=data[,2]
+user=data[,1]
+data_num=length(rating)
+movie_num= max(movie)#电影数量1682
+user_num=max(user)#用户数量943
+#打乱数据集
+rr=sample(1:data_num,data_num)
+data= data[rr,]
+```
+
+
+```{r}
+data_num/(movie_num*user_num)#稀疏度
+```
+
+```{r}
+#划分训练集和测试集
+train_num=80000#(data_num*9)%/%10;
+train_vec=data[1:train_num,];#训练集
+probe_vec=data[train_num:84000,];#测试集
+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= 8 #Number of batches把数据分为8份
+num_m = movie_num  # Number of movies 电影数量
+num_p = user_num  #Number of users 用户数量
+
+```
+
+```{r}
+#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)
+
+
+```
+
+```{r}
+for(epoch in epoch:maxepoch){
+  
+  #采用mini batch的方法，每次训练10000个样本
+  for(batch in 1:numbatches){
+    
+    #print(c('epoch',epoch,'batch',batch))
+    
+    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)
+    #求出损失函数
+    #其实之所以这样计算是为了降低计算量，如果每一次都uv相乘，计算量太大
+    #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3列
+    Ix_m=IO*w1_P1[aa_p,] #损失*U-lambda*V 就是更新规则
+    Ix_p=IO*w1_M1[aa_m,] #损失*V-lambda*U
+    #还是不太懂他为啥能batch后这么计算 而且矩阵形式本来就更复杂
+    dw1_M1 = matrix(rep(0,num_m*num_feat),nrow = num_m , ncol = num_feat ,byrow =T)
+    #生成全零movie特征矩阵
+    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 %%%%%%%%%%%
+    #真正开始更新 新的矩阵=过去矩阵+学习率*导数
+    #全零特征矩阵=上次的得到的矩阵*动量+学习率*导数/1w
+    w1_M1_inc = momentum*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 = momentum*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 )  
+  err_train[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_valid[epoch]= sqrt(sum((pred_out- rating)^2)/NN)
+  print(paste('epoch',epoch,'Train RMSE',signif(err_train[epoch], 4),'Test RMSE',signif(err_valid[epoch], 4)))
+  ################################输出到屏幕
+  
+ # err_random[epoch]=sqrt(sum((runif(10001,1,5) - rating)^2)/NN)
+  
+}
+```
+
+
+
+```{r}
+#画出迭代的损失函数图
+err_random=sqrt(sum((runif(10001,1,5) - rating)^2)/NN)
+
+plot(err_train,type="l",col=4,lwd=4,main="Loss",xlab="epoch",ylab="Loss")
+lines(err_valid,type="l",col=3,lwd=3)
+#lines(err_random-1,type="l",col=2,lwd=2)
+abline(h=err_random,col=2,lwd=2)
+legend("topright",c("train","test","baseline"),lty=1,col=c(4,3,2),lwd=c(4,3,2))
+
+
+```
+
+```{r}
+#展示一下预测和真实的对比
+gap=40:50
+rbind (rating[gap],pred_out[gap])
+```
+
+
+```{r}
+# TOP-N 推荐
+#输出要为用户i推荐的j部电影 及其预测评分
+i=5
+j=6
+user_i_rating=w1_P1[i,]%*%t(w1_M1)+mean_rating
+used=data[data$V1==i,2]
+user_i_rating[used]=0
+order(user_i_rating,decreasing=TRUE)[1:j]
+user_i_rating[order(user_i_rating,decreasing=TRUE)[1:j]]
+```
+
+```{r}
+#计算Pre@10 Re@10 其中实际评分大于4的，我们认为是需要推荐的物品
+# TOP-N 推荐
+
+j=10
+txt=numeric(user_num)
+for(i in 1:user_num){
+  user_i_rating_real=probe_vec[(probe_vec$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)))
+
+```
+
+```{r}
+#在新样本上进行预测
+test=read.csv("test.txt", header = FALSE,sep=' ')
+# 读取test中的用户 一一进行top10推荐
+j=10
+record=rep(0,0)
+for(i in test$V1){
+  user_i_rating=w1_P1[i,]%*%t(w1_M1)+mean_rating
+  used=data[data$V1==i,2]
+  user_i_rating[used]=0
+  record=rbind(record,order(user_i_rating,decreasing=TRUE)[1:j])
+}
+final_result=cbind(test,data.frame(record))
+write.csv(final_result,file = 'result.csv')
+```
+