MRHAF: Multi-Relational Hierarchical Attention with Hybrid Knowledge Fusion for Explainable Herb Recommendations

Traditional Chinese medicine (TCM) herb recommendations aim to personalize herb combinations for specific symptom profiles in accordance with TCM compatibility principles, thereby ensuring optimal therapeutic efficacy. However, existing approaches often fail to account for the varying intensity of complex multi-relational interactions among herbs, therapeutic effects, and symptoms. This limitation hampers effective hybrid feature fusion across multi-source herbal knowledge and constrains overall recommendation performance. To address these challenges, we propose Multi-Relational Hierarchical Attention with Hybrid Knowledge Fusion (MRHAF), a novel framework designed to improve both predictive accuracy and interpretability by modeling latent relationships among therapeutic effects, herbs, and symptoms, as well as the material basis underlying herbal efficacy. MRHAF consists of three core components: (1) a global Herb-Efficacy-Symptom Knowledge Graph (HESKG), which applies multi-head attention to capture global semantic information; (2) a Herb-Symptom Interaction Graph (HSIG), which leverages self-attention to model direct therapeutic associations; and (3) a molecular-level Herb-Attribute-Component Knowledge Graph (HACKG), which integrates explicit attributes and implicit biochemical information to establish the material basis of efficacy. Additionally, we integrate global semantic features and local interaction features through a dual-branch attention architecture. Extensive experiments on two benchmark datasets demonstrate that MRHAF outperforms state-of-the-art baselines, achieving improvements of 9.75% and 22.3% in Precision@10, respectively. Clinical evaluations confirm that MRHAF effectively captures TCM formulation principles and delivers reliable recommendation outcomes, while network pharmacology analyses further validate the rationality of the recommended herbs. Overall, this study provides a new perspective on herbal compatibility and offers valuable guidance for clinical decision-making in TCM.