User-based Collaborative Filtering
User-based CF 包含下列步骤:
1. Selecting Neighborhoods
1.1 Select all neighbors
1.2 Threadhold similarity or distance
1.3 Random neighbors
1.4 Top-*N* neighbors by similarity or distance
理论上,当具备良好的相似度度量方式时,选择越多user作为近邻,user的兴趣面就越广.
而实际上,选择越多user,越容易引进噪声.
一般选择 25 - 100 个user作为近邻
本节实验部分采用 Top-30
2. Scoring Items from Neighborhoods
如何根据近邻user对item的评分预测给定user对item的的评分
可以考虑如下方法:
- Average
- Weighted Average
- Multiple linear regression
其中,Weighted Average方法常用且有效,本节实验部分采用该方法.
3. Normalizing Data
user的习惯各不相同,有的user偏向于打高分,有的偏向于打低分. 因此有必要对user的打分做归一化.
可以采用Mean-centering方法,即Substract user mean prior to computing
也可以采用z-score normization方法:
a. Mean-center, and divide by standard deviation
b. Normalizes for the spread across the scale
c. small additional gain in prediction accuracy over mean-centering
本节实验采用Mean-center方法,结合Weighted Average,定义如下:
Pu,i=μu+∑v∈N(u;i)S(u,v)(rv,i−μv)∑v∈N(u;i)∣s(u,v)∣
That is, compute the weighted average of each neighbor v′ offset from average (rv,j−μv), then add the user's average rating muu. N(u;i)
is the neighbors of μ for item i
4. Computing Similarities
如何选取user的近邻users?
4.1 使用user的rating作为特征,按cosine求距离,选择Top-*30*
4.2 使用Pearson correction coefficient 定义如下:
4.3 使用聚类方法寻找近邻users
然后pick user's cluster to generate predictions.
本节实验部分采用4.1的方法.
本节课程材料
vedio
slides
code
推荐阅读:
- Herlocker J L, Konstan J A, Borchers A, et al. An algorithmic framework for performing collaborative filtering[C]//Proceedings of the 22nd annual international ACM SIGIR conference on Research and development in information retrieval. ACM, 1999: 230-237.
- Herlocker J L, Konstan J A, Riedl J. Explaining collaborative filtering recommendations[C]//Proceedings of the 2000 ACM conference on Computer supported cooperative work. ACM, 2000: 241-250.
