【机器学习】7.1.余弦相似度 协同过滤

包含全部示例的代码仓库见GIthub

1 导入库

import pandas as pd
import matplotlib.pyplot as plt
%matplotlib inline
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity

2 数据准备

data = pd.DataFrame({'one':[4, np.nan, 2, np.nan],
                     'two':[np.nan, 4, np.nan, 5],
                     'three':[5, np.nan, 2, np.nan],
                     'four':[3, 4, np.nan, 3],
                     'five':[5, np.nan, 1, np.nan],
                     'six': [np.nan, 5, np.nan, 5],
                     'seven':[np.nan, np.nan, np.nan, 4]},
                     index = list('ABCD'))
data
# output
    one	two	three	four	five	six	seven
A	4.0	NaN	5.0	3.0	5.0	NaN	NaN
B	NaN	4.0	NaN	4.0	NaN	5.0	NaN
C	2.0	NaN	2.0	NaN	1.0	NaN	NaN
D	NaN	5.0	NaN	3.0	NaN	5.0	4.0

3 找到和A趣味相投的人

3.1 余弦相似度

sim_AB = cosine_similarity(data.loc['A',:].fillna(0).values.reshape(1, -1),
                           data.loc['B',:].fillna(0).values.reshape(1, -1))
# 新版本 Serise 没有 reshape
sim_AB
# output
array([[0.18353259]])

sim_AC = cosine_similarity(data.loc['A',:].fillna(0).values.reshape(1, -1),
                           data.loc['C',:].fillna(0).values.reshape(1, -1))
sim_AC
# output
array([[0.88527041]])

3.2 去中心化

可能有些用户对所有物品的评分都比较高，所以对行取平均值

data_center = data.apply(lambda x: x-x.mean(), axis=1)  # 对行取平均值
sim_AB = cosine_similarity(data_center.loc['A',:].fillna(0).values.reshape(1, -1),
                           data_center.loc['B',:].fillna(0).values.reshape(1, -1))
sim_AB
# output
array([[0.30772873]])

sim_AC = cosine_similarity(data_center.loc['A',:].fillna(0).values.reshape(1, -1),
                           data_center.loc['C',:].fillna(0).values.reshape(1, -1))
sim_AC
# output
array([[-0.24618298]])

sim_AD = cosine_similarity(data_center.loc['A',:].fillna(0).values.reshape(1, -1),
                           data_center.loc['D',:].fillna(0).values.reshape(1, -1))
sim_AD
# output
array([[0.56818182]])

4 预测A对two的评分

(sim_AD*data.loc['D', 'two'] + sim_AB*data.loc['B', 'two'])/(sim_AD+sim_AB)
# output
array([[4.64867562]])