MMP-Refer: Multimodal Path Retrieval-augmented LLMs For Explainable Recommendation

Researchers Xiangchen Pan and Wei Wei have introduced MMP-Refer (Multimodal Path Retrieval-augmented LLMs for Explainable Recommendation), a novel framework that addresses two critical gaps in current AI recommendation systems. While existing LLM-based recommenders often incorporate collaborative filtering data, they typically ignore multimodal information like product images and descriptions. More importantly, these systems lack transparency—they can't explain why they recommend certain items, making them untrustworthy 'black boxes' for users.

MMP-Refer solves this by creating retrieval paths that combine visual, textual, and interaction data through a heuristic search algorithm. The system uses a sequential recommendation model with joint residual coding to generate multimodal embeddings, then employs a trainable lightweight collaborative adapter to map user interaction patterns into the LLM's semantic space. This allows the language model to understand not just what items users interact with, but why those patterns exist, enabling it to generate natural language explanations alongside recommendations.

The framework represents a significant advancement in making AI recommendations both accurate and interpretable. By bridging the gap between collaborative signals and multimodal content understanding, MMP-Refer can explain recommendations like 'We suggest this hiking backpack because you've shown interest in outdoor gear, and this model has similar features to the camping equipment you previously purchased.' This transparency builds user trust while maintaining or improving recommendation quality through better data integration.