[Python知识库]推荐算法-完整代码实现ItemCF

用python实现了基础的 Item-based collaborative filtering方法，
更多的内容可以访问我的项目：添加链接描述

import numpy as np
import pandas as pd
import json
from pandas import Series, DataFrame
import math

tdata=pd.read_csv(unicode(r'D:/推荐系统数据集/训练集与测试集/traindata.csv','utf-8'))
tdata.drop(columns='timestamp',inplace=True)
tdata['userId']=tdata['userId'].astype('str')
tdata['movieId']=tdata['movieId'].astype('str')
tdata['rating']=tdata['rating'].astype('float')
#tdata['rating']=tdata['rating'].apply(lambda x: x/5)
df1=tdata.groupby(by=['userId'])
types=tdata['userId'].unique()
info1={}
for type in types:
    df3={}
    df2=df1.get_group(type)
    for i in range(len(df2)):
        df3[df2.iloc[i,1]]=df2.iloc[i,2]
    info1[type]=df3
    info_json=json.dumps(info1,indent=4)
#print(info_json)

####利用itemcf求出物品间的相似度
train1=info1
C1=dict()
N1=dict()    
for u,items in train1.items():
    for i,rui in items.items():
        if i not in N1.keys():
            N1[i] =0
        N1[i] +=1
        for j,ruj in items.items():
            if i == j:
                continue
            if i not in C1.keys():
                C1.update({i:{j:0}})
			
            if j not in C1[i].keys():
                C1[i].update({j:0}) 
                
            C1[i][j] += 1/((math.log(1+len(items)*1.0))**2)

W1=dict()
for i,related_items in C1.items():#u为一级坐标
	for j,cij in related_items.items():#v为二级坐标，cuv即C[u][v]的值，即|N(u)∩N(v)|
		if i not in W1.keys():
			W1.update({i:{j:cij/math.sqrt(N1[i]*N1[j])}})
		else:
			W1[i].update({j:cij/math.sqrt(N1[i]*N1[j])})
            
Tdata=pd.read_csv(unicode(r'D:/推荐系统数据集/训练集与测试集/traindata.csv','utf-8'))
Tdata.drop(columns='rating',inplace=True)
Tdata['userId']=Tdata['userId'].astype('str')
Tdata['movieId']=Tdata['movieId'].astype('str')
Tdata['timestamp']=Tdata['timestamp'].astype('float')
print(Tdata)
min1=min(Tdata['timestamp'])
print(min1)
db1=Tdata.groupby(by=['userId'])
types=Tdata['userId'].unique()
info2={}
for type in types:
    db3={}
    db2=db1.get_group(type)
    for i in range(len(db2)):
        db3[db2.iloc[i,1]]=db2.iloc[i,2]
    info2[type]=db3
    info2_json=json.dumps(info2,indent=4)
#print(info_json)

train2=info2
Testdata=pd.read_csv(unicode(r'D:/推荐系统数据集/训练集与测试集/testdata.csv','utf-8'))
Testdata.drop(columns='timestamp',inplace=True)
Testdata['userId']=Testdata['userId'].astype('str')
Testdata['movieId']=Testdata['movieId'].astype('str')
Testdata['rating']=Testdata['rating'].astype('float')
Testdata
de1=Testdata.groupby(by=['userId'])
types=Testdata['userId'].unique()
info4={}
for type in types:
    de3={}
    de2=de1.get_group(type)
    for i in range(len(de2)):
        de3[de2.iloc[i,1]]=de2.iloc[i,2]
    info4[type]=de3
    info4_json=json.dumps(info4,indent=4)

def GetRank (user,K,train,w):
    rank1=dict()
    ru=train[str(user)]
    list4=list()
    for j,ruj in ru.items():
        list4.extend([j])
    for j,ruj in ru.items():
        for i,wij in sorted(w[j].items(),key=lambda x:x[1], reverse=True)[0:10]:
            if i in ru.keys():
                continue
            if str(i) not in rank1.keys():
            	rank1.update({i:0})
            rank1[i]+=wij
    dict5 = sorted(rank1.items(), key=lambda x:x[1], reverse=True)
    dict6=dict(dict5[0:K])
    return dict6       

def Getlist(data,testuser):            
    user_items=list()
    for user,related in data.items():
        #print(user,related)
        if str(user) == str(testuser):
            for i in related.keys():
                #print(i)
                user_items.extend([i])   
    return user_items

def Recall(c1,c2,K):
    hit=0
    all=0
    for user in c1.keys():
        tu=Getlist(c2,user)
        rank=GetRank(user,K,train1,W1)
        for i in rank:
            if i in tu:
                hit +=1
        all +=len(tu)
    return hit / (all * 1.0)

def Precision(c1,c2,K):
    hit=0
    all=0
    for user in c1.keys():
        tu=Getlist(c2,user)
        rank=GetRank(user,K,train1,W1)
        for i in rank:
            if i in tu:
                hit +=1 
        all += K
    return hit / (all * 1.0)

def Coverage(train,test,K):
    recommend_items = set()
    all_items = set()
    for user in train.keys():
        for item in train[user].keys():
            all_items.add(item)
        rank = GetRank(user,K,train1,W1)
        for item in rank:
            recommend_items.add(item)
    return len(recommend_items) / (len(all_items)*1.0)


def con(t1,t2,w,K):    
    hit=0
    all=0
    t0=893286639.0
    for user in t1.keys():
        rank=GetRank(user,K,train1,W1)
        for i in rank:
            for j,tuj in sorted(t2[user].items(),key=lambda 				           x:x[1],reverse=True)[0:10]:
                for m,wim in w[i].items():
                    if j==m:
                        hit +=wim/(0.01*(abs(t0-tuj)))
        all+=K    
    return hit / (all * 1.0)
for i in range(10,61,10):
    con1=con(train1,train2,W1,i)
    print("{}".format(con1))

def Popularity(train):
    item_popularity = dict()
    for user,items in train.items():
        for item in items.keys():
            if item not in item_popularity:
                item_popularity[item] = 0
            item_popularity[item] +=1
    return item_popularity
popularlity_all=Popularity(train1)

def novel(train,popularlity,K):
    net=0
    n=0
    for user in train.keys():
        rank = GetRank(user,K,train1,W1)
        for item in rank.keys():
            net += math.log(1+popularlity[item],2)
            n+=1
    net /=n*1.0
    return net
for i in range(10,51,10):
    novel1=novel(train1,popularlity_all,i)
    print("{}".format(novel1))

本文只是实现了简单的item-based协同过滤算法，因为前面有数据的处理，以及后面也编写了评价方法的部分，所以显得稍微长，不过也是为了让初学者能少走点弯路。
如果能够对你有一点帮助，那会令我感到非常荣幸，祝你进步，实现理想。
愿你在这个世界永远被温柔相待，也愿我的姐姐能早日康复。