PhD Dissertation Proposal by TRAN Nhu Thuat | Disentangling User Preferences Towards Self-Interpretable Recommender Systems

Oftentimes, there have been multiple hidden factors governing user behaviors when they interact with recommender systems. Uncovering these hidden explanatory factors underlying user interests would not only achieve higher recommendation accuracy but also improve the interpretability of user preferences, which in turn gains user trust into recommender systems. This dissertation proposal, therefore, aims at designing more advanced methodologies to disentangle the complex patterns behind user preferences.

Disentangling user preferences faces two main challenges. For one, users might not explicitly mention their specific interests when interacting with items, requiring disentangled recommendation models to work in unsupervised settings. For another, users only interact with a small subset of interested items, which causes their interaction records sparse, preventing learning high quality disentangled representations. This dissertation introduces the solutions for the forementioned challenges. First, a novel disentangled recommendation model is designed to unsupervisedly discover multiple factors of user interests via iterative latent attention and implicit differentiation, which generalizes existing methods to demonstrate stronger performance. Second, a text-aware disentanglement approach to deal with data sparsity problem is proposed, which aligns discovered interest factors from textual content and user adoptions to learn more expressive representations. Extensive experiments on real world datasets show the stronger performance of the proposed methods than existing baselines and demonstrate the potential of the proposed methods in interpreting user preferences in unsupervised settings. Thirdly, this proposal discusses future research directions to improve existing disentangled recommendation models and enhance the interpretability of user preferences.

TRAN Nhu Thuat is a PhD Candidate in Computer Science at the SMU School of Computing and Information Systems, supervised by Prof. Hady W. Lauw. The central theme of his research is interpretability in recommender systems.