Multi‐Modal Knowledge Graph Hybrid Attention Network for Recommendation

In the context of the highly abundant information landscape of modern society, recommender systems are designed to provide personalized information by analyzing user behavior data to uncover latent user preferences. Recent research has demonstrated that recommendation approaches integrating knowledge graphs with attention mechanisms can significantly address prevalent challenges in traditional recommender systems, such as sparse data problems and the cold‐start problem. Furthermore, with the widespread presence of multi‐modal content in real‐world applications, incorporating multi‐modal item information into knowledge graphs has garnered increasing attention, as it enables a deeper insight into user preferences. Despite these advances, conventional techniques largely focus on modeling users' single‐dimensional preferences, overlooking the complex and dynamic nature of user interests. Additionally, the heterogeneity of multi‐modal features is often insufficiently addressed. To tackle these issues, this study proposes a Multi‐modal Knowledge Graph Hybrid Attention Network (MTKGHAT) for recommendation, aiming to improve both accuracy and robustness. The model first constructs a multi‐modal knowledge graph by converting initial data into knowledge triples. Based on this foundation, the CLIP model is employed to capture the textual and visual representations of items and map them into a unified embedding space, thereby reducing discrepancies across modalities. Concerning user modeling, a hybrid attention fusion framework is designed, which integrates time‐sensitive attention to capture users' short‐term dynamic preferences and multi‐head graph attention to learn their multi‐dimensional structural preferences. These two types of preferences are adaptively integrated via a parameterized weighted fusion mechanism that balances their contributions. Experimental findings indicate that the model outperforms existing baseline methods on multiple evaluation metrics, demonstrating robust recommendation effectiveness and accuracy.