SlopOne推荐算法

在开源框架taste中有SlopOne的Java实现，效果不错。使用movielens的数据，代码如下

代码

#coding:utf-8

import re
import math

#读取数据，并生成矩阵
def getMatrix():
	mat = {}
	f = open("u.data", "r")
	for line in f:
		list = map(int, line.strip("\n").split("\t"))
		if list[0] in mat:
			mat[ list[0] ][ list[1] ] = list[2]
		else:
			mat[ list[0] ] = { list[1] : list[2] }
	f.close()
	return mat

#计算某个用户的平均分
def getAvg(usr):
	res = 0.0
	for i in usr:
		res += usr[i]
	return res / len(usr)

#预测分数, 返回矩阵mat中用户usr对item的评分
def getSlopOne(mat, user, item):

	#用户user的所有item的列表
	list = mat[user]
	#分子
	mole = 0.0
	#分母
	demo = 0.0
	#对于每一个物品，都计算它和物品item的差值，最终算出相对它item的score
	for it in list:
		diff_sum = 0.0
		user_num = 0
		for us in mat:
			us_it = mat[us]
			#如果该user同时评价过这两个item，则采纳他的评分
			if item in us_it and it in us_it:
				diff_sum += us_it[item] - us_it[it]
				user_num += 1
		#如果item被人评价过
		if user_num:
			diff_avg = diff_sum / user_num
			mole += (list[it] + diff_avg) * user_num
			demo += user_num

	#如果没有人评价过，则取这个人的平均分
	if user_num == 0:
		return getAvg(list)

	#否则返回最终的score
	return mole / demo

def main():
	mat = getMatrix()
	rf = open("u.data", "r")
	wf = open("o.data", "w")
	for line in rf:
		list = map(int, line.strip("\n").split("\t"))
		score = getSlopOne(mat, list[0], list[1])
		output = str(list[0]) + "\t" + str(list[1]) + "\t" + str(list[2]) + "\t" + str(score) + "\n"
		wf.write(output)
	rf.close()
	wf.close()

if __name__ == "__main__":
	main()