Research Seminar by Guandong Xu | Causality-inspired Recommendation: Robustness, Transparency and Fairness

Recommendation System (RS) as an information filtering tool to alleviate the information explosion has gained prominence in academia and industry. The functions of an RS in real-world services largely rely on its performance, thereby the importance to build trustworthy RS. Robustness, transparency and fairness are the three-layer hierarchy toward trustworthy RS, which directly impact user satisfaction, recommendation persuasiveness and stakeholder reputation. In this talk, the speaker will discuss how causality helps to fulfil the three-layer hierarchy of trustworthy RS by using causal inference approaches. He will first give the evidence that causality is richer in information than statistical dependencies, then note the necessity of causal structural information for enhancing recommendations based on three rungs. He will provide his explorations in causality-inspired recommendations and discuss his major findings. Specifically, he will discuss recommendation robustness facing low-quality data bias scenarios with social exposure, label noise and distribution shift. He will also show how causality-based explanations can enhance recommendation transparency and help to build fairness-aware recommendation algorithms. He will finally highlight some future directions and open questions.