信息管理与商业智能系学术讲座

时间：2024-10-22 13:30-14:30

地点：思源楼724室

题目：Mitigating Compounding Self-Selection Biases in Recommender System Training: A Joint Generative Modeling Approach

主讲人：石岩松信息管理与商业智能系青年副研究员

内容摘要:

Conventional recommender systems rely on consumers’ feedback, such as product ratings, to train machine learning models for personalized recommendations. However, this approach is prone to data biases stemming from consumers’ self-selection behaviors. When machine-learning models are trained on such biased data, the recommendations they produce may be skewed and cannot accurately align with consumer preferences. This study delves into consumer conversion funnel, which includes the exposure, purchase, and evaluation stages, and identifies three types of data biases—exposure, acquisition, and under-reporting—that arise from consumers’ compound self-selection behaviors. To address these biases in training recommender systems, we propose a novel joint generative modeling approach that incorporates a rating data observation process depicting consumer behavioral patterns through the conversion funnel. We provide a theoretical proof of the identifiability of the proposed model with respect to its key parameters, ensuring that the three types of biases can be precisely inferred from observed biased data. To rigorously evaluate the performance of the proposed approach, we design a comprehensive evaluation framework for cases both with and without bias-free ratings and include three real-world datasets for evaluation. The comprehensive evaluation demonstrates that our proposed approach outperforms state-of-the-art methods in both rating prediction and rating observation process modeling. Furthermore, we extend our model to accommodate biased implicit feedback captured through parallel conversion funnels and to integrate deep learning model structures. Our research contributes to the literature and practice of RSs by introducing an innovative approach to mitigate compounding self-selection biases in training recommender systems.