CS Events Monthly View
PhD DefenseTowards Human-centered Recommender Systems |
|
||
Thursday, September 16, 2021, 09:30am - 11:30am |
|||
Speaker: Zuohui Fu
Location : Remote via Zoom
Committee:
Prof. Gerard de Melo (Advisor)
Prof. Yongfeng Zhang
Prof. Hao Wang
Prof. Qingyao Ai (External member)
Event Type: PhD Defense
Abstract: Recently, there has been extensive interest in developing intelligent human-centered AI (artificial intelligence)) systems that support human participation so as to facilitate cooperation between humans and machines. Recommender systems are a particularly pervasive form of AIsystem that can aid in decision-making in the face of ever-growing amounts of information. Modern deep learning based recommender systemshave made great strides in accuracy and effectiveness, but raise a number of important challenges: 1) How can we actively incorporate humanparticipation into the decision-making procedure of recommender systems? 2) How can we ensure that explanations are provided such that users can better understand why particular items are being recommended? 3) How can we alleviate biases in recommender systems? This thesis proposes several novel methods to fill these gaps. In particular, for improved human understanding, it introduces an adversarial semantic learning framework for cross-lingual settings. For human integration, a human-in-the-loop conversational recommender system with external graph structure is introduced. To ensure fair explanations, we mitigate the unfairness within graph-based explainable reasoning in the recommender system. Finally, for human-system cooperation, we present a popularity debiasing framework to integrate user interaction and debiased dialogue state management in a conversational recommender system. We not only extensively evaluate our proposed approaches on multiple real-world recommendation datasets, but also contribute open public datasets to the community. The experimental results demonstrate the effectiveness of the proposed methods in achieving satisfying prediction accuracy, mitigating bias, and providing users with faithful understandable explanations.
: