SAMRec: Self-Attention Based Multimodal Recommendation

With the development of multimedia platforms, traditional recommendation systems that rely on historical user-item interactions are no longer sufficient for most platforms' demand for precise recommendations, as they fail to leverage item modality information. Multimodal recommendation is a method that combines historical user-item interaction data with multimodal item information to better capture user preferences. Existing multimodal recommendation methods often treat item modality information in a simplistic manner, typically using pretrained models to extract features, without deeply exploring the relationships between various modality information. To address this issue, we propose a multimodal recommendation model based on the self-attention mechanism, which effectively captures the global relationships among the multiple item modality features that users have interacted with, thereby more accurately modeling user preferences. Additionally, to prevent the loss of relatively important information when using the self-attention mechanism, we incorporate skip connect to linearly combine the original modality features, ensuring that all critical information is retained while improving the global feature representation. The experimental results show that the self-attention mechanism greatly improves the model's capacity to better capture user preferences, resulting in increased recommendation accuracy. We evaluate the suggested model on three publicly available datasets.