KDD 2022 | characterization alignment and uniformity are considered in collaborative filtering

Paper title ：

Towards Representation Alignment and Uniformity in Collaborative Filtering

Thesis link ：

https://arxiv.org/pdf/2206.12811.pdf

Code ：

https://github.com/THUwangcy/DirectAU

This paper mainly focuses on the data representation in collaborative filtering methods , The existing research mainly focuses on designing more powerful encoder （ for example , Figure neural network ） To learn a better representation of . And lack of right CF The expected attributes of the representation in , This is important for understanding the existing CF The basic principle of the method and the design of new learning objectives are very important . In this paper, we measure from the point of view of alignment and uniformity on the hypersphere CF Characterization quality in .

• Firstly, it reveals theoretically BPR Loss and these two attributes （ Alignment and uniformity ） The connection between .
• From the point of view of quantitative alignment and uniformity CF The learning dynamics of the method are analyzed , Show that better alignment or uniformity will help improve recommended performance .
• According to the analysis , A learning objective is proposed to directly optimize these two attributes , be called DirectAU.