Structure-Preserving Projection for Mitigating Modality Bias in LLM-Based Sequential Recommendation

2026-08-09

AI Summary

This work addresses a critical challenge in integrating large language models (LLMs) into recommender systems: the projection of collaborative embeddings into the LLM’s semantic space often introduces modality-induced bias, which distorts the intrinsic collaborative structure between users and items and consequently degrades performance. To mitigate this issue, the paper proposes the first explicit structure-preserving projection approach, introducing a novel structure-preserving loss function that effectively maintains the geometric relationships of collaborative embeddings during cross-modal mapping. By seamlessly unifying collaborative filtering signals with LLM semantics, the method significantly enhances recommendation accuracy across multiple benchmark datasets, demonstrating its effectiveness and novelty in alleviating collaborative signal distortion while preserving essential structural information.

Abstract

Recent LLM-based recommenders integrate textual and collaborative signals by projecting collaborative embeddings into the embedding space of the LLM. However, this projection can introduce modality bias that distorts the underlying collaborative structure and limits the usefulness of projected embeddings. To address this issue, we propose a novel structure-preserving projection approach that maintains the relational geometry of collaborative embeddings through dedicated structure-preserving losses. Comprehensive experiments demonstrate that our approach consistently improves recommendation performance, providing a more reliable path for LLM-based recommendation.

Problem

• modality bias
• LLM-based recommendation
• collaborative embeddings
• embedding projection
• sequential recommendation

Innovation

• structure-preserving projection
• modality bias
• LLM-based recommendation
• collaborative embeddings
• relational geometry