PhD Dissertation Proposal by LIU Zhongzhou | Towards Trustworthy Recommendation Systems: Beyond User-Item Collaborative Filtering

Recommendation systems have permeated every aspect of our lives. As one of the most classical paradigms of recommendation systems, user-item collaborative filtering has been widely employed and proven effective in various domains. However, they are still subject to a series of issues. So far, we have delved into two important limitations. Firstly, a typical collaborative filtering method often falls short in adequately capturing and modeling the fine-grained preferences of individual users. Secondly, traditional collaborative filtering is susceptible to bias, particularly the issue of popularity bias.

In this proposal, we aim to propose the trustworthy recommendation systems by addressing the limitations mentioned above in contrast to traditional collaborative filtering techniques. In uncovering the fine-grained preference for effectiveness, we integrate user's multi-granularity preferences in a dependance aware manner and propose a learning framework to model the preferences adaptively. In uncovering popularity bias for fairness, we identified popularity bias for both user and item sides in the context of traditional collaborative filtering. And we then proposed an adversarial framework that mitigates the two-sided popularity bias simultaneously. Finally, we made some initial explorations on other limitations of traditional user-item collaborative filtering techniques and proposed some promising future works.

LIU Zhongzhou is a Ph.D. candidate in Computer Science at the SMU School of Computing and Information Systems, supervised by Assistant Prof. FANG Yuan. His research aims to investigate the untapped possibilities of recommendation systems that go beyond traditional user-item collaborative filtering, including fine-grained preferences, fairness, causality-based recommendations and more.