# -*- coding: utf-8 -*-
"""
Created on Thu Nov 1 10:29:52 2018
@author: AZ
"""
# 代码说明：
# 基于用户的协同过滤算法的具体实现
import math
import numpy as np
import pandas as pd
import os
os.chdir('E:/广电大数据营销推荐项目案例/数据清洗/电视节目信息数据预处理')
# 借助pearson相关系数进行修正后的余弦相似度计算公式，计算两个用户之间的相似度
# 记 sim(user1, user2) = sigma_xy /sqrt(sigma_x * sigma_y)
# user1和user2都表示为[[节目名称,隐性评分], [节目名称,隐性评分]],如user1 = [['节目一', 3.2], ['节目四', 0.2], ['节目八', 6.5], ...]
def calCosDistByPearson(user1, user2):
x = 0.0
y = 0.0
sigma_xy = 0.0
sigma_x = 0.0
sigma_y = 0.0
for item in user1:
x += item[1]
# user1对其看过的所有节目的平均评分
average_x = x / len(user1)
for item in user2:
y += item[1]
# user2对其看过的所有节目的平均评分
average_y = y / len(user2)
for item1 in user1:
for item2 in user2:
if item1[0] == item2[0]: # 对user1和user2都共同看过的节目才考虑进去
sigma_xy += (item1[1] - average_x) * (item2[1] - average_y)
sigma_x += (item1[1] - average_x) * (item1[1] - average_x)
sigma_y += (item2[1] - average_y) * (item2[1] - average_y)
if sigma_x == 0.0 or sigma_y == 0.0: # 若分母为0，相似度为0
return 0
return sigma_xy/math.sqrt(sigma_x * sigma_y)
# 创建所有用户的观看信息（包含隐性评分信息）,“从用户到节目”
# 格式例子：users_to_items = {用户一:[['节目一', 3.2], ['节目四', 0.2], ['节目八', 6.5]], 用户二: ... }
def createUsersDict(df):
(m, n) = df.shape
data_array = np.array(df.iloc[:m + 1, 1:])
users_names = np.array(df.iloc[:m + 1, 0]).tolist()
items_names = np.array(df.columns)[1:]
users_to_items = {}
for i in range(len(users_names)):
user_and_scores_list = []
for j in range(len(items_names)):
if data_array[i][j] > 0:
user_and_scores_list.append([items_names[j], data_array[i][j]])
users_to_items[users_names[i]] = user_and_scores_list
return users_to_items
# 创建所有节目被哪些用户观看的字典，也就是创建“从节目到用户”的倒排表items_and_users
# items_to_users = {节目一: [用户一, 用户三], 节目二: ... }
def createItemsDict(df):
(m, n) = df.shape
data_array = np.array(df.iloc[:m + 1, 1:])
users_names = np.array(df.iloc[:m + 1, 0]).tolist()
items_names = np.array(df.columns)[1:]
items_to_users = {}
for i in range(len(items_names)):
users_list = []
for j in range(len(users_names)):
if data_array[j][i] > 0:
users_list.append(users_names[j])
items_to_users[items_names[i]] = users_list
return items_to_users
# 找出与用户user_name相关的所有用户（即邻居）并依照相似度排序
# neighbors_distance = [[用户名, 相似度大小], [...], ...] = [['用户四', 0.78],[...], ...]
def findSimilarUsers(users_dict, items_dict, user_name):
neighbors = [] # neighbors表示与该用户看过相同节目的所有用户
for items in users_dict[user_name]:
for neighbor in items_dict[items[0]]:
if neighbor != user_name and neighbor not in neighbors:
neighbors.append(neighbor)
# 计算该用户与其所有邻居的相似度并降序排序
neighbors_distance = []
for neighbor in neighbors:
distance = calCosDistByPearson(users_dict[user_name], users_dict[neighbor])
neighbors_distance.append([neighbor, distance])
neighbors_distance.sort(key=lambda item: item[1], reverse=True)
return neighbors_distance
# 基于用户的协同过滤算法
# K为邻居个数，是一个重要参数，参数调优时使用
def userCF(user_name, users_dict, items_dict, K, all_items_names_to_be_recommend):
# recommend_items = {节目名：某个看过该节目的该用户user_name的邻居与该用户的相似度, ...}
recommend_items = {}
# 将上面的recommend_items转换成列表形式并排序为recommend_items_sorted = [[节目一, 该用户对节目一的感兴趣程度],[...], ...]
recommend_items_sorted = []
# 用户user_name看过的节目
items_user_saw = []
for item in users_dict[user_name]:
items_user_saw.append(item[0])
# 找出与该用户相似度最大的K个用户(邻居)
similar_users = findSimilarUsers(users_dict, items_dict, user_name)
if len(similar_users) < K:
k_similar_user = similar_users
else:
k_similar_user = similar_users[:K]
# 得出对该用户的推荐节目集
for user in k_similar_user:
for item in users_dict[user[0]]:
# 该用户user_name没有看过的节目才添加进来，才可以推荐给该用户
if item[0] not in items_user_saw:
# 而且该节目必须是在备选推荐节目集中
if item[0] in all_items_names_to_be_recommend:
if item[0] not in recommend_items:
# recommend_items是一个字典。第一次迭代中，表示将第一个邻居用户与该用户的相似度加到节目名上，后续迭代如果有其他邻居用户也看过该节目，
# 也将其与该用户的相似度加到节目名上，迭代的结果就是该用户对该节目的感兴趣程度
recommend_items[item[0]] = user[1]
else:
# 如果某个节目有k个邻居用户看过，则将这k个邻居用户与该用户的相似度相加，得到该用户对某个节目的感兴趣程度
recommend_items[item[0]] += user[1]
for key in recommend_items:
recommend_items_sorted.append([key, recommend_items[key]])
# 对推荐节目集按用户感兴趣程度降序排序
recommend_items_sorted.sort(key=lambda item: item[1], reverse=True)
return recommend_items_sorted
# 输出推荐给该用户的节目列表
# max_num:最多输出的推荐节目数
def printRecommendItems(recommend_items_sorted, max_num):
count = 0
for item, degree in recommend_items_sorted:
print("节目名：%s， 推荐指数：%f" % (item, degree))
count += 1
if count == max_num:
break
# 主程序
if __name__ == '__main__':
# all_users_names = ['A','B','C']
all_users_names = [3,13,23]
df1 = pd.read_csv('./wordsbag/dataprocess/data/week/mydata/temp_movies_01mat.csv',sep=',',encoding='gbk',header='infer',error_bad_lines=False)
(m1, n1) = df1.shape
# 按照"备选推荐节目集及所属类型01矩阵"的列序排列的所有用户观看过的节目名称
items_to_be_recommended_names = np.array(df1.iloc[:m1 + 1, 0]).tolist()
df2 = pd.read_csv('./wordsbag/dataprocess/data/week/mydata/temp_user_scores_mat2.csv',sep=',',encoding='gbk',header='infer',error_bad_lines=False)
# users_dict = {用户一:[['节目一', 3.2], ['节目四', 0.2], ['节目八', 6.5]], 用户二: ... }
users_dict = createUsersDict(df2)
# items_dict = {节目一: [用户一, 用户三], 节目二: [...], ... }
items_dict = createItemsDict(df2)
for user in all_users_names:
print("为用户id为 %s 的推荐节目如下：" % user)
recommend_items = userCF(user, users_dict, items_dict, 2, items_to_be_recommended_names)
printRecommendItems(recommend_items, 5)
print('该用户的推荐任务完成。')
print()
数据下载地址：https://download.csdn.net/download/qq_38281438/10757266
转自：https://blog.csdn.net/WuchangI/article/details/80160566