MSFNet: Multi-Modal Semantic Fusion Network for Herb Recommendation

Graph Neural Networks (GNNs) have been extensively utilized in the field of traditional Chinese medicine (TCM) herb recommendation. Nevertheless, the current TCM herb recommendation algorithms have not adequately harnessed the semantic information of symptoms, constraining the system's precision and individualization capabilities. In response, we introduced the multimodal semantic fusion network for herb recommendation (MSFNet), designed to profoundly explore and integrate the semantic characteristics of symptoms. The Semantic Fusion Decoder (SFDec) incorporates a dual-attention mechanism, blending the semantic features of symptoms with the interaction dynamics of TCM herbs. Concurrently, the Symptom Feature Enhancer (SFEnh) guarantees the retention and augmentation of semantic information throughout network dissemination. Experiments conducted on the authentic dataset from Xi'an Traditional Chinese Medicine Hospital have demonstrated that MSFNet outperforms the optimal baseline method, KDHR, by 3.6% in Precision@10, validating the model's efficacy and superiority.