Search papers, labs, and topics across Lattice.
This paper introduces a structure-preserving projection method to mitigate modality bias in LLM-based sequential recommendation systems, which often distort collaborative embeddings when integrating textual and collaborative signals. By employing dedicated structure-preserving losses, the proposed approach maintains the relational geometry of collaborative embeddings, enhancing their utility in recommendation tasks. Experimental results show significant improvements in recommendation performance, underscoring the effectiveness of this method in preserving the integrity of collaborative structures.
Maintaining the relational geometry of collaborative embeddings can significantly enhance the performance of LLM-based recommendation systems, countering the detrimental effects of modality bias.
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.
