GAT-TCM: Optimizing pediatric diarrhea treatment with Graph Attention Networks for personalized Traditional Chinese Medicine recommendations

As the application of Traditional Chinese Medicine (TCM) recommendation systems for pediatric diarrhea treatment grows, existing systems often lack personalized optimization when modeling the complex relationships between symptoms and herbs, resulting in lower recommendation accuracy. Specifically, in the recommendation of traditional formulas like Ginseng-Mume Decoction, existing methods fail to fully capture the intrinsic relationships between herbs, leading to suboptimal personalized recommendations for different symptom combinations. To address this, we propose a novel TCM treatment recommendation system (GAT-TCM). The system leverages Graph Attention Networks to model the complex relationships between symptoms and herbs, using self-attention to dynamically learn the importance of each node, optimizing the relevance of herb recommendations. We also introduce contrastive pre-training to improve the model’s understanding of symptom-herb similarities, enhancing learning capability and generalization in data-scarce scenarios. Additionally, Hierarchical Structure Network effectively addresses the imbalance in learning rare symptoms and herbs, strengthening the model’s ability to learn from less common data. Experimental results show that GAT-TCM significantly outperforms baseline models, with P@5 reaching 0.6025 and F1@5 at 0.5247, demonstrating its effectiveness in personalized TCM recommendations, especially for Ginseng-Mume Decoction.