CFALR merges collaborative filtering with LLMs for personalized fashion outfits

Personalized outfit recommendation has long challenged e-commerce and social platforms, requiring systems that balance user preferences with aesthetic compatibility. Traditional collaborative filtering (CF) struggles with sparse data and complex user-item-outfit relationships, while template-based approaches are too rigid. To address these gaps, researchers from Singapore, China, and Australia introduced CFALR (Collaborative Filtering-Augmented Large Language Model for Recommendation). This framework synergizes CF with LLMs by describing user-outfit interactions in natural language, allowing the LLM to capture fashion semantics. Simultaneously, CF-enhanced embeddings bridge the semantic space and collaborative interaction spaces, enabling the system to handle sparse scenarios and generate cohesive outfits.

CFALR's technical contributions are threefold: it is the first LLM-based architecture specifically designed for personalized outfit recommendation; it features a CF-augmented generative mechanism that efficiently navigates the exponential combination space of fashion items; and it uses trainable projection layers to optimally integrate relational and content features. Evaluated on the Polyvore and IQON benchmarks, CFALR outperforms both traditional CF-based and existing LLM-based methods across personalized fill-in-the-blank and personalized outfit generation tasks. This research, published on arXiv and led by Yujuan Ding, marks a significant step toward AI-powered fashion assistants that truly understand individual style.