Probing Digital Footprints and Reaching for Inherent Preferences: A Cause-Disentanglement Approach to Personalized Recommendations

• 成果类型: Article
• 署名作者: Wang, Cong; Shi, Yansong; Guo, Xunhua; Chen, Guoqing
• 署名单位: Peking University; Fudan University; Tsinghua University; Tsinghua University
• 刊物名称: INFORMATION SYSTEMS RESEARCH
• ISSN/ISSBN: 1047-7047
• DOI: 10.1287/isre.2023.0181
• 发表日期: 2025
• 关键词: social-influence; systems; BIAS; commerce; MODEL

摘要

The abundance of multiple types of consumer digital footprints recorded on e-commerce platforms has fueled the design of personalized recommender systems for decision support. However, capturing consumers' inherent preferences for effective recommendations based on consumer digital footprints can be challenging because of the multitude of factors driving consumer behaviors. Model training and recommendation outcomes may become biased if other factors are inappropriately recognized as consumers' inherent preferences in the learning process. Drawing on consumer behavior theories, we tease out various factors that drive consumers' digital footprints at different consumption stages. We develop a novel recommendation approach, namely, DISC (Disentangling consumers' Inherent preferences, item Salience effect, and Conformity effect), which leverages disentangled representation learning with a causal graph to derive the effect of each factor driving consumer behaviors. This approach provides personalized and interpretable recommendations based on the inference of consumers' normative inherent preferences. The DISC model's identifiability is demonstrated through theoretical analysis, enabling rigorous causal inference based on observational data. To evaluate DISC's performance, extensive experiments are conducted on real-world data sets with a carefully designed protocol. The results reveal that DISC outperforms state-of-the-art baselines significantly and possesses good interpretability. Moreover, we illustrate the potential impact of different marketing strategies' by intervening on the disentangled causes through follow-up counterfactual analyses based on the causal graph. Our study contributes to the literature and practice by causally unpacking the behavioral mechanism behind consumers' digital footprints and designing an interpretable personalized recommendation approach anchored in their inherent preferences.