Multimodal Recommendation Systems in the LLM Era: A Survey of Feature Representation and Fusion Methods

The surge in multimodal data such as text, images, and sometimes audio, has led to a growing interest in recommendation systems capable of integrating these diverse modalities for enhanced personalization. Multimodal recommendation systems leverage such rich data to better understand user preferences and provide highly tailored recommendations. Traditional approaches employ a variety of techniques, including feature alignment, modality-specific graphs, attention mechanisms, and advanced fusion strategies, to effectively combine multimodal information. Recently, the integration of large language models (LLMs) into multimodal recommenders has transformed the field, enabling advanced preference summarization, context-aware fusion and even personalized content generation. This survey provides a comprehensive overview of both the traditional and LLM-based multimodal recommendation systems, highlighting their architectures, key methodologies, datasets, evaluation benchmarks, and real-world applications. Furthermore, we explore the challenges and opportunities posed by the fusion and processing of multi-modal data by these recommender systems.