协同滤波 Collaborative filtering 《推荐系统实践》 第二章

利用用户行为数据

 

简介:

用户在网站上最简单存在形式就是日志。

原始日志(raw log)------>会话日志(session log)-->展示日志或点击日志

用户行一般分为两种:

1显性反馈:包括用户明确表示对物品喜好的行为(数据量小)

2隐形反馈:网页浏览等(数据量大)

 

 

wps_clip_image-28823

wps_clip_image-25503

 

 

 

用户行为的统一标准如下:

wps_clip_image-15497

 

 

协同滤波与实验设计:

本文参考《推荐系统实践》这本书,但细节和书中略有不同,因为个人把书中代码组合到一起有些小问题,所以自己小修改了一番,可以运行,与大家分享。

 
 
实验数据集:

采用GroupLens提供的MovieLens数据集。下载地址http://www.grouplens.org/node/73。为了提供实验速度,本文采用较小数据集,即m1-100k那个数据集中的u.data文件,其他文件没有用,如果有兴趣,读者可以自己参考readme.

 

实验数据说明

U.data数据包含4列,分别是  UserID::MovieID::Rating::Time  ,本实验关心的是topN推荐,所以只关心用户是否看了某个电影,而不关心用户对电影的评分和看电影的时间。所以取数据前两列。

离线设计如下,将用户行为数据集随机分成M份,取M-1份为训练集,1份为测试集。本文M=8.代码

def SplitData(data,M=8,k=3,seed=1):
    test = {}
    train = {}
    random.seed(seed)
  

    for user, item in data:
        if random.randint(0,M) ==k:
            if user not in test:
                test[user]=set()
            test[user].add(item)
        else:
            if user not in train:
                train[user]=set()
            train[user].add(item)
    return train,test

 

评测指标:

对用户u推荐N个物品(R(u)), 令用户u在测试集喜欢的物品集合为T(u),然后定义

召回率:wps_clip_image-9143

 

 

两种召回率代码如下:

def Recall(train,test,N):
    hit=0
    alls=0
    W=UserSimilarity2(train)

      
    for user  in train.keys():
       try:#有可能test有没user看过的item
            te_user_item = test[user]

            recomRank = Recommend(user,train,W,N)


            for  recom_item,w in recomRank:
                if recom_item in te_user_item:
                    hit+=1
            alls+=len(te_user_item)
       except:
           pass;



    return hit*1.0/alls

#ItemFC_recall     
def ItemRecall(train,test,N):
    hit=0
    alls=0
    W=ItemSimilarity(train)

       
    for user  in train.keys():
       try:#有可能test有没user看过的item
            te_user_item = test[user]
            recomRank = ItemRecommendation(user,train,W,N)
            #pdb.set_trace()
            for  recom_item,w in recomRank:
                if recom_item in te_user_item:
                    hit+=1
            alls+=len(te_user_item)
       except:
          pass;


    return hit*1.0/alls

 

准确率:wps_clip_image-16536

def Precision(train,test,N):
    hit=0
    alls=0
    W=UserSimilarity2(train)
    for user  in train.keys():
        try:#有个能test有没user看过的item
            te_user_item = test[user]
            recomRank = Recommend(user,train,W,N)
            #pdb.set_trace()
            for  recom_item,w in recomRank:
                if recom_item in te_user_item:
                    hit+=1
            alls+=N
        except:
            pass

    return hit*1.0/alls


def ItemPrecision(train,test,N):
    hit=0
    alls=0
    W=ItemSimilarity(train)

       
    for user  in train.keys():
       try:#有可能test有没user看过的item
            te_user_item = test[user]
            recomRank = ItemRecommendation(user,train,W,N)
            #pdb.set_trace()
            for  recom_item,w in recomRank:
                if recom_item in te_user_item:
                    hit+=1
            alls+=N
       except:
          pass;


    return hit*1.0/alls

覆盖率:反应了推荐算法发觉长尾的能力,覆盖率越高,说明推荐算法越能够推荐长尾中的物品给用户。一个简单的定义如下:所有推荐的物品的并集/测试集的所有物品

wps_clip_image-14884

 

 

两种覆盖率代码如下:

def Coverage(train,N):
    recommend_items = set()
    all_items = set()
    W=UserSimilarity2(train)
    for user in train.keys():
        for item in train[user]:
            all_items.add(item)

        rank =Recommend(user,train,W,N)

        for item in rank[0]:
            recommend_items.add(item)

    return len(recommend_items)/(len(all_items)*1.0)

def ItemCoverage(train,N):
    recommend_items = set()
    all_items = set()
    W=ItemSimilarity(train)
    
    for user in train.keys():
        for item in train[user]:
            all_items.add(item)

        rank =ItemRecommendation(user,train,W,N)

        for item in rank[0]:
            recommend_items.add(item)

    return len(recommend_items)/(len(all_items)*1.0)

 

 

 

 

基于用户的协同滤波User_CF(Collaborative filtering):

算法思路:

1)找到和目标永和兴趣相似的用户集合

2)找到这个集合中的用户喜欢的,且目标用户没有听说过的物品推荐给用户

相似度计算其中N(u)表示用户看过的电影集合。

wps_clip_image-18085

如下公式度量了UserCF算法中用户u对物品i的感兴趣程度:

wps_clip_image-14502

 

 

基于物品的协同滤波Item_CF

思路步骤:

1)计算物品之间的相似度

2)根据物品的像吸毒和用户的历史行为给用户生成推荐列表

N(i)若代表喜欢物品i的用户数目,则物品i和j相似度可以用下面的公式表示:wps_clip_image-21319

USER_CF,ITEM_CF计算物品i,j相似度的代码如下:

 

def UserSimilarity2(train,flag=1):
#第二中计算W的函数
    item_users = dict() #bulid an new empty dicitionary
    for u ,item in train.items():
        for i in item:
            if i not in item_users:
                item_users[i] = set() #生成一个集合
            item_users[i].add(u)

    C = dict()
    N = dict()#N[u]表示拥护u的项目(看电影)个数



    for item ,users in item_users.items():

        for u in users:
            if u not in N:
                N[u]=1#如果用户u不在字典N里面,先创建
            else:
                N[u]+=1
            for v in users :
                if u!=v:
                    if flag==0:#正常情况
                        if (u,v) not in C:
                            C[(u,v)]=1
                        else:
                            C[(u,v)]+=1
                    elif flag==1:
                         if (u,v) not in C:
                            C[(u,v)]=1/log(1+len(users))
                         else:
                            C[(u,v)]+=1/log(1+len(users))

    W = dict()

    for uv in C.keys():
        #pdb.set_trace()
        u=uv[0]
        v=uv[1]
        if u not in W:
            W[u]=set()
        #添加与用户u相关的用户v,第二个意思是他们的权重Wuv
        W[u].add((v,C[uv]/sqrt(N[u] * N[v])))


    return W


def ItemSimilarity(train):
    C = dict() #记录 N(i)并N(j)
    N = dict() #记录 N(i) i表示喜欢物品i的用户数

    for u , items in train.items():
        for i in items:
            if i not in N:
                N[i]=1
            else:
                N[i]+=1

            for j in items:
                if i != j:
                    if (i,j) not in C:

                       C[(i,j)]=1
                    else:
                       C[(i,j)]+=1

    #calculate finial similarity:
    W= dict()

    for ij ,val in C.items():

        i=ij[0]#物品i
        j=ij[1]#物品j

        if i not in W:
            W[i]=set()
        W[i].add((j,val/sqrt(N[i]*N[j])))
    
    
    return W

 

计算用户u对物品j的兴趣公式如下:

wps_clip_image-19431

 

userCF ,Item CF 推荐topN代码如下:

def Recommend(user,train,W,N,K=20):


    rank = dict()
    interacted_items = train[user]



    
    for v,wuv in sorted(W[user], key=lambda x:x[1],reverse=True)[0:K]:


        for i  in train[v]:#v看过的电影
            if i not in interacted_items:#如果电影i不在user已看过的电影里
                if i not in rank:
                    rank[i]=wuv * 1
                else:
                    rank[i]+=wuv * 1

    rank=sorted(rank.items(), key = lambda x:x[1],reverse=True)

      #  rank=[(key,val) for key,val in rank.items()]#字典转换为list
    rank=rank[:N]
    return rank
    
    
    
    
def ItemRecommendation(user,train,W,N,K=10):
    rank = dict()

    user_items =train[user]
    for i in user_items:
        for j , wij in sorted(W[i], key = lambda x:x[1],reverse =True)[0:K]:
            if j not in user_items:
                if j not in rank:
                    rank[j] = wij*1
                else:
                    rank[j]+=wij*1
    
    rank=sorted(rank.items(), key = lambda x:x[1],reverse=True)

    rank=rank[:N]    
    return rank

 

 

 

参数M=8,N=10,k=10时候,输出结果如下:

 

可以通过调节参数获得其他结果

 

 

全部代码如下

 

ItemCoverage:  0.601796407186
ItemRecall: 0.172728085068
ItemPrecision: 0.208972972973
Recall  0.165132695916
Precision  0.199783783784
Coverage 0.698203592814

 

 

 

 

# -*- coding: utf-8 -*-
'''

Created on 2014��4��16��

@author: Administrator
'''
import random
import pdb
from math import *
import traceback


def SplitData(data,M=8,k=3,seed=1):
    test = {}
    train = {}
    random.seed(seed)
  

    for user, item in data:
        if random.randint(0,M) ==k:
            if user not in test:
                test[user]=set()
            test[user].add(item)
        else:
            if user not in train:
                train[user]=set()
            train[user].add(item)
    return train,test












#USER_FC_recall

def Recall(train,test,N):
    hit=0
    alls=0
    W=UserSimilarity2(train)

      
    for user  in train.keys():
       try:#有可能test有没user看过的item
            te_user_item = test[user]

            recomRank = Recommend(user,train,W,N)


            for  recom_item,w in recomRank:
                if recom_item in te_user_item:
                    hit+=1
            alls+=len(te_user_item)
       except:
           pass;



    return hit*1.0/alls

#ItemFC_recall     
def ItemRecall(train,test,N):
    hit=0
    alls=0
    W=ItemSimilarity(train)

       
    for user  in train.keys():
       try:#有可能test有没user看过的item
            te_user_item = test[user]
            recomRank = ItemRecommendation(user,train,W,N)
            #pdb.set_trace()
            for  recom_item,w in recomRank:
                if recom_item in te_user_item:
                    hit+=1
            alls+=len(te_user_item)
       except:
          pass;


    return hit*1.0/alls

     # pdb.set_trace()
     
     
def Precision(train,test,N):
    hit=0
    alls=0
    W=UserSimilarity2(train)
    for user  in train.keys():
        try:#有个能test有没user看过的item
            te_user_item = test[user]
            recomRank = Recommend(user,train,W,N)
            #pdb.set_trace()
            for  recom_item,w in recomRank:
                if recom_item in te_user_item:
                    hit+=1
            alls+=N
        except:
            pass

    return hit*1.0/alls


def ItemPrecision(train,test,N):
    hit=0
    alls=0
    W=ItemSimilarity(train)

       
    for user  in train.keys():
       try:#有可能test有没user看过的item
            te_user_item = test[user]
            recomRank = ItemRecommendation(user,train,W,N)
            #pdb.set_trace()
            for  recom_item,w in recomRank:
                if recom_item in te_user_item:
                    hit+=1
            alls+=N
       except:
          pass;


    return hit*1.0/alls
    
#计算覆盖率
#USER_CF
def Coverage(train,N):
    recommend_items = set()
    all_items = set()
    W=UserSimilarity2(train)
    for user in train.keys():
        for item in train[user]:
            all_items.add(item)

        rank =Recommend(user,train,W,N)

        for item in rank[0]:
            recommend_items.add(item)

    return len(recommend_items)/(len(all_items)*1.0)

def ItemCoverage(train,N):
    recommend_items = set()
    all_items = set()
    W=ItemSimilarity(train)
    
    for user in train.keys():
        for item in train[user]:
            all_items.add(item)

        rank =ItemRecommendation(user,train,W,N)

        for item in rank[0]:
            recommend_items.add(item)

    return len(recommend_items)/(len(all_items)*1.0)

def UserSimilarity2(train,flag=1):
#第二中计算W的函数
    item_users = dict() #bulid an new empty dicitionary
    for u ,item in train.items():
        for i in item:
            if i not in item_users:
                item_users[i] = set() #生成一个集合
            item_users[i].add(u)

    C = dict()
    N = dict()#N[u]表示拥护u的项目(看电影)个数



    for item ,users in item_users.items():

        for u in users:
            if u not in N:
                N[u]=1#如果用户u不在字典N里面,先创建
            else:
                N[u]+=1
            for v in users :
                if u!=v:
                    if flag==0:#正常情况
                        if (u,v) not in C:
                            C[(u,v)]=1
                        else:
                            C[(u,v)]+=1
                    elif flag==1:
                         if (u,v) not in C:
                            C[(u,v)]=1/log(1+len(users))
                         else:
                            C[(u,v)]+=1/log(1+len(users))

    W = dict()

    for uv in C.keys():
        #pdb.set_trace()
        u=uv[0]
        v=uv[1]
        if u not in W:
            W[u]=set()
        #添加与用户u相关的用户v,第二个意思是他们的权重Wuv
        W[u].add((v,C[uv]/sqrt(N[u] * N[v])))


    return W


def ItemSimilarity(train):
    C = dict() #记录 N(i)并N(j)
    N = dict() #记录 N(i) i表示喜欢物品i的用户数

    for u , items in train.items():
        for i in items:
            if i not in N:
                N[i]=1
            else:
                N[i]+=1

            for j in items:
                if i != j:
                    if (i,j) not in C:

                       C[(i,j)]=1
                    else:
                       C[(i,j)]+=1

    #calculate finial similarity:
    W= dict()

    for ij ,val in C.items():

        i=ij[0]#物品i
        j=ij[1]#物品j

        if i not in W:
            W[i]=set()
        W[i].add((j,val/sqrt(N[i]*N[j])))
    
    
    return W

#给出要推荐的物品item,(并且存储于rank中)
#rank是一个字典,rank[item]=推荐力度
#返回前N个推荐
def Recommend(user,train,W,N,K=10):


    rank = dict()
    interacted_items = train[user]



    
    for v,wuv in sorted(W[user], key=lambda x:x[1],reverse=True)[0:K]:


        for i  in train[v]:#v看过的电影
            if i not in interacted_items:#如果电影i不在user已看过的电影里
                if i not in rank:
                    rank[i]=wuv * 1
                else:
                    rank[i]+=wuv * 1

    rank=sorted(rank.items(), key = lambda x:x[1],reverse=True)

      #  rank=[(key,val) for key,val in rank.items()]#字典转换为list
    rank=rank[:N]
    return rank
    
    
    
    
def ItemRecommendation(user,train,W,N,K=10):
    rank = dict()

    user_items =train[user]
    for i in user_items:
        for j , wij in sorted(W[i], key = lambda x:x[1],reverse =True)[0:K]:
            if j not in user_items:
                if j not in rank:
                    rank[j] = wij*1
                else:
                    rank[j]+=wij*1
    
    rank=sorted(rank.items(), key = lambda x:x[1],reverse=True)

    rank=rank[:N]    
    return rank



f = open('u.data')

data=[]#存储数据
for line in f:
 
    data.append(line.split('\t')[:2])

train,test=SplitData(data)




print 'ItemCoverage:  %s' % ItemCoverage(train,10)
print 'ItemRecall: %s' % ItemRecall(train,test,10)
print 'ItemPrecision: %s' %   ItemPrecision(train,test,10)

print 'Recall  %s' %  Recall(train,test,10)
print 'Precision  %s' %Precision(train,test,10)
print 'Coverage %s' % Coverage(train,10)

 

 

总结:

本人愚笨,不太清楚怎么把《推荐系统实践》里面的代码整合在一起,自己改了改,希望可以跟他家分享做个参考吧。PS:数据量还不是很小,运行需要一段时间。

 

参考书目:推荐系统实践

转载请标注:http://www.cnblogs.com/Dzhouqi/p/3668919.html                       

 

posted @ 2014-04-16 16:29 joey周琦 阅读(...) 评论(...) 编辑 收藏