题　目：Jointly Modeling Cascading Self-selection Behaviors for Mitigating Data Biases in Training Recommendation Systems

嘉　宾：Yansong Shi, PhD candidate, Tsinghua University

主持人：魏煊  助理教授  77779193永利官网

时　间：2023年11月15日（周三）13:30-15:00

地　点：77779193永利官网 徐汇校区安泰楼A305室

内容简介:

Conventional recommender systems (RSs) rely on consumers’ feedback, such as product ratings, to train models for personalized recommendations. However, this approach is prone to data sampling biases stemming from consumers’ self-selection behaviors. Machine-learning models trained on such biased data may result in biased systems that struggle to accurately predict consumer preferences. This study examines the comprehensive process of purchase and rating and identifies three types of data biases—exposure, acquisition, and under-report biases—that arise from consumers’ cascading self-selection behaviors. To address these biases in training RSs, we propose a novel sample selection model that incorporates consumer behavioral patterns in the purchase and rating stages. We theoretically prove the identifiability of the proposed model with respect to key parameters to guarantee that the three types of biases can be precisely inferred from observed data. To rigorously evaluate the performance of the proposed approach, two bias-free datasets are used as testbeds. Experimental results demonstrate that our proposed approach outperforms state-of-the-art methods in both rating prediction and rating observation process modeling. Furthermore, we illustrate the heterogeneous effects of consumers with different rating disclosure patterns. Our research contributes to the literature and practice of RSs by introducing an innovative debiasing approach that accounts for rating observation mechanisms when handling biased training data.

演讲人简介：

Yansong Shi is currently a Ph.D. candidate in Management Science and Engineering at the School of Economics and Management, Tsinghua University. His primary research interests lie in the area of machine learning, causal inference, and data mining, with a particular focus on consumer behavior data biases. His works were presented/accepted at a variety of leading conferences in Information Systems and Computer Science fields, including ICIS, WITS, and WSDM. At ICIS 2021, one of his works was selected as Kauffman Best Student Paper (Runner-up). Moreover, one of his works was a nominee for the CNAIS 2022 Best Paper Award. Several of his papers are currently under review/revision in top-tier journals.

欢迎广大师生参加！