A multimodal outfit recommendation Q&A system based on LLMs and KGs

Abstract An outfit acts as a silent language, not only reflecting a person’s personality, taste, and lifestyle but also showcasing careful coordination and creativity. Outfit recommendation systems are widely applied in scenarios such as e-commerce platforms and virtual styling services, where personalized and dynamic suggestions are essential. However, existing fashion recommendation approaches suffer from notable limitations, including heavy reliance on purchase history, an emphasis on individual items rather than holistic outfit coordination, and limited adaptability to rapidly changing fashion trends. To address these, we propose a multimodal outfit recommendation(MOR) Q&A framework that integrates large language models (LLMs), knowledge graphs (KGs), and retrieval-augmented generation. Unlike conventional systems, MOR enables a Q&A-based interaction mode for more personalized and context-aware recommendations. We introduce a graph-driven reranker that leverages a custom scoring function to improve recommendation accuracy. The LLM is used to construct a fashion KG and a multimodal knowledge base, supporting the retrieval and ranking of top-n outfit candidates. The final output combines text and images for a richer recommendation experience. Extensive experiments demonstrate that MOR outperforms baselines in both effectiveness and personalization, offering a more intelligent and practical solution for real-world fashion recommendation.