基于用户协同过滤python

# -*- coding: utf-8 -*- ''' Created on 2015��5��2�� @author: Joeyqzhou ''' import random import numpy as np import math #Input:'ratings.dat' #A example #1::1193::5::978300760# #UserId::Moive::Ratings::Timestamp #Output:userMoiveData,dictionary,key:userId,value:[moiveid...] def readMoiveIdFromRatings(filename='smallData_u.data',splitSymbol="\t"): fileRating = open(filename) userMoiveData = dict() for line in fileRating: piece = line.split(splitSymbol) id = piece[0] moiveId = piece[1] if id not in userMoiveData: userMoiveData[id] = [] userMoiveData[id].append(moiveId) return userMoiveData#######返回以用户为主键，以电影为值字典集合。 #To split data into (M-1) piece of training and 1 piece of testing data #k is a random seclected number to between (1,M) #seed is to generate the random number with certain mode (seed) def splitData(data,M=8,k=3,seed=1): test = dict() train = dict() random.seed(seed)###如果使用相同的seed( )值，则每次生成的随即数都相同，如果不设置这个值，则系统根据时间来自己选择这个值，此时每次生成的随机数因时间差异而不同 for user, items in data.items(): for itemj in items: if random.randint(0,M) ==k:###在python中的random.randint(a,b)用于生成一个指定范围内的整数，其中参数a是下限，参数b是上限，生成的随机数n: a <= n <= b if user not in test: test[user] = list() test[user].append(itemj) else: if user not in train: train[user] = list() train[user].append(itemj) return train,test #To compute the user similarity of all data #input #data:dict(),key:user,value:[] list that contains items #type: choose different similarity mode, type=0 is the default choice def userSimilarity(data,type='0'): #Build the inverse table: item_users(key:item,value:list) item_users = dict()#####--训练集中的以电影为主键，以用户为列表的集合。 for useri , items in data.items():#######data为训练集 if type == 'rating': items = items[0] for itemj in items:####--遍历当前用户看过的所有电影 if itemj not in item_users: item_users[itemj] = list() item_users[itemj].append(useri)###-----训练集中的以电影为主键，以用户为列表的集合。 #Calculate the absolute relative time C = dict()###--以用户为主键，其他用户及两个用户的共同打分物品数为值的嵌套字典 #key:useri #value:dict,key:userj,value:the number of movie both useri and uerj have seen N = dict()###--以用户为主键，以每个用户看过的电影数为值的字典集合。 #key:userId #value:the total item number of userId for itemi , users in item_users.items():####item_users---训练集中的以电影为主键，以用户为列表的集合。 for useri in users:###遍历当前电影的所有看过的用户 if useri not in N: N[useri] = 0 N[useri] += 1 for userj in users:###遍历当前电影的所有看过的用户 if useri != userj: if useri not in C: C[useri] = dict() if userj not in C[useri]: C[useri][userj] = 1 else: C[useri][userj] += 1 W = dict() #Normalizing C to get W #key:useri #value:dict,key:userj,value:simlarity(useri,userj) for useri, related_user in C.items():###-C--以用户为主键，其他用户及两个用户的共同打分物品数为值的嵌套字典 W[useri] = dict() for userj in related_user:#####当前用户的其他用户及两个用户的共同打分物品数的字典 W[useri][userj] = C[useri][userj]*1.0/math.sqrt(N[useri]*N[userj])####### N -----以用户为主键，以每个用户看过的电影数为值的字典集合。 return W #To recommend user a item list # trainData: as the name #data:dict(),key:user,value:[] list that contains items #W is similarity dictionary #key:useri #value:dict,key:userj,value:simlarity(useri,userj) #K: choose the nearest K neighbor of the user #output: #recommendationList:type(dict),key(item),val(how much you want to recommend) def userCF_Recommend(user,trainData,W,K=6):###trainData为训练及数据--k为当前用户的近邻个数 recommendationList = dict()#####前提是在训练集中，通过当前用户的近邻用户，来找出近邻用户看过的电影，而当前用户没有看过的电影。以未看过的电影为主键，以与其他用户的累计相似度为值的集合。 alreadyUsedItem = trainData[user] for useri , similar_useri in sorted(W[user].items(), \ key = lambda x:x[1],\ reverse=True)[0:K]:#######张利---找出当前用户的k个近邻用户及相似度 for itemj in trainData[useri]:###当前用户的近邻用户所看过的电影 if itemj not in alreadyUsedItem:####若当前用户没有看过电影itemj if itemj not in recommendationList:#####若itemj不在推荐列表里 recommendationList[itemj] = 0 recommendationList[itemj] += similar_useri return sorted(recommendationList.items(), key = lambda x:x[1], reverse=True) def itemSimilarity(trainData): C = {} #key:item #value:key(itemj),value(absolute number seen by two users) N = {} #key:item #value: len(item) for user,items in trainData.items(): for item in items: if item not in N: N[item] = 0 N[item] += 1 if item not in C: C[item] = {} for itemj in items: if itemj not in C[item]: C[item][itemj] = 0 C[item][itemj] += 1 #key:item #value:key(itemj),value(simi(item,itemj)) W = {} for item, relatedItem_val in C.items(): W[item] = {} for itemi in relatedItem_val: W[item][itemi] = C[item][itemi]*1.0/math.sqrt(N[item]*N[itemi]) return W def itemCF_Recommend(user,trainData,W,K=6): itemsAlreadyUsed = trainData[user] recomList = {} for itemi in itemsAlreadyUsed: for itemj, Wij in sorted(W[itemi].items(), key = lambda x:x[1], reverse = True)[0:K]: if itemj not in itemsAlreadyUsed: if itemj not in recomList: recomList[itemj] = 0 recomList[itemj] += Wij return sorted(recomList.items(), key = lambda x:x[1], reverse=True) #Calculate the recall of the CF #train:training data #test:testing data #Recommend at most N items for each user def Evaluation_userCF(train, test, N): hit = 0 totalTestItemNum = 0.0 totalRecommendationNum = 0.0 W = userSimilarity(train) print "Evaluation" for user in train.keys(): if user in test.keys(): test_user_item = test[user] totalTestItemNum += len(test_user_item) recommendationList = userCF_Recommend(user,train,W) #Only get the movie ID recommendationList = [recomItem[0] for recomItem in recommendationList] if len(recommendationList)>N: #If the recommendation list's length is larger than N recommendationList = recommendationList[0:N] totalRecommendationNum += N else: