Distilling Transitional Patterns to Multimodal Large Language Models for Session-based Recommendation

This is the authors' abstract. We don't add key points for this paper.

Session-based recommendation (SBR) predicts the next item based on anonymous sessions. Traditional SBR explores user intents based on ID collaborations or auxiliary content. To further alleviate data sparsity and cold-start issues, recent Multimodal SBR (MSBR) methods utilize simplistic pre-trained models for modality learning but have limitations in semantic richness. Considering the semantic reasoning abilities of multimodal Large Language Models (LLMs) and the practical request for high inference efficiency, we focus on the LLM-enhanced MSBR scenario in this paper, which leverages LLM cognition for comprehensive multimodal representation generation, to enhance downstream MSBR. Tackling this problem faces two challenges: i) how to obtain LLM cognition on both inherent multimodal knowledge and transitional patterns, ii) how to align transitional patterns and inherent knowledge into one unified LLM without increasing inference complexity. To this end, we propose a multimodal LLM-enhanced framework TPAD, which extends a teacher-student distillation paradigm to decouple and align transitional patterns for promoting MSBR. TPAD establishes parallel Knowledge-MLLM and Transfer-MLLM, where the former as student model interprets item knowledge-reflected features and the latter as teacher model extracts transition-aware features. Transitional pattern alignment harnessing mutual information estimation unites two MLLMs, distilling transitional patterns into representations. Extensive experiments on real-world datasets demonstrate the effectiveness of TPAD. Implementation codes are available at https://github.com/TiliaceaeSU/TPAD .