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

This article presents DISC (Disentangling consumers' Inherent preferences, item Salience effect, and Conformity effect), a novel personalized recommendation approach that uses disentangled representation learning and causal graph modeling to interpret and improve e-commerce recommendations. DISC models consumers' behaviors across three shopping stages—need recognition, prepurchase evaluation, and purchase—by disentangling inherent preferences from external influences such as item salience and social conformity. The approach introduces a latent decision path variable to capture flexible consumer decision mechanisms, enabling rigorous causal inference and addressing challenges like ambiguous negative samples in behavioral data. Extensive experiments on real-world datasets from Beibei and Taobao demonstrate that DISC significantly outperforms state-of-the-art baselines in both in-sample purchase prediction and inherent preference-oriented recommendation tasks, while providing interpretable insights into consumer behavior. The study also validates the causal graph structure, discusses managerial implications for targeted marketing, and highlights the model's extensibility to more complex consumer behavior scenarios.