Recent Advances in Multimodal Representation Learning: A Survey

Modern recommendation systems play a crucial role in helping users discover items of interest by modeling user preferences and item attributes through implicit interactions, such as purchases and clicks. Researchers have developed multimodal recommender systems capable of interpreting and integrating diverse data sources, inspired by the way humans perceive the world through multiple sensory modalities (e.g., audio, text, and images). These models are designed to uncover hidden relationships among modalities and recover complementary information that single-modal systems might miss. A core challenge in the multimodal domain is bridging the heterogeneity gap — the differences which are present in the representation across modalities. Multimodal representation learning, particularly deep learning-based methods, plays a crucial role in addressing this challenge by learning robust representations across multiple abstraction levels. A comprehensive overview of recent advances in deep multimodal representation learning for recommendation systems is presented in this survey, further categorizing the approaches into three major frameworks: joint representation , coordinated representation , and encoder-decoder models , based on how they integrate data from various modalities.