<p>Multimodal recommendation has become a popular research direction because it can capture user preferences more finely, thereby delivering more accurate and personalized recommendations and, in turn, improving user satisfaction and engagement. Most existing studies typically focus on leveraging modal features or modality-related graph structures to learn feature representations. However, these methods suffer from two limitations: They often ignore interaction time information (e.g., interaction timestamps), making it difficult for the recommendation system to capture the dynamic evolution of user interests; When propagating both modal information and behavioral information in the same view (e.g., a user–item graph), noise in the modal information is amplified and intertwined with behavioral signals, thus weakening the discriminability of the modal embeddings. To address these issues, we propose a novel Deep Interaction-Aware Multimodal Recommendation model (DIA) that enhances recommendation performance by deeply exploring user interaction sequences. Specifically, we introduce a temporal graph embedding module that incorporates user interaction time information into graph convolutions via the Multitaper method to capture time-aware collaborative patterns. Furthermore, we design an interaction-adaptive encoder that leverages users’ behavioral interactions to reinforce the consistency and discriminability of modal embeddings in a multi-view setting. Extensive experiments on three benchmark datasets demonstrate that DIA achieves significant advantages over various state-of-the-art recommendation methods, validating the effectiveness of the proposed model.</p>