PRISM framework disentangles user preference from relevance in e-commerce search

E-commerce search systems rely on modeling user behavior to estimate item relevance and preference, but these signals are often entangled due to exposure mechanisms, feedback loops, and semantic matching. This creates confounding effects that degrade ranking robustness. Researchers propose PRISM, a Preference-Relevance Interaction Semantic Modeling framework that explicitly models the interaction between user preference and item relevance rather than treating them as independent components. PRISM introduces three key modules: a preference rectification module that iteratively refines user preference under relevance-aware constraints to reduce behavioral confounding; an LLM-driven semantic anchoring mechanism that leverages positive and negative prototypes to calibrate relevance representations, ensuring semantic consistency; and a preference-conditioned evidence routing module that adaptively aggregates multi-source behavioral signals for context-aware relevance estimation.

Extensive experiments on two public e-commerce benchmarks demonstrate that PRISM consistently outperforms strong baselines, validating the effectiveness of explicitly modeling preference-relevance interaction. The framework particularly excels in scenarios with dynamic user behavior and high data entanglements, offering a robust solution for real-world e-commerce search. By decoupling preference and relevance while leveraging LLMs for semantic grounding, PRISM paves the way for more accurate and bias-resistant search personalization. The researchers highlight that this approach can be extended to other recommendation tasks where user behavior signals are similarly entangled.