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

Area of Science:

• Computational biology
• Pharmacology
• Artificial intelligence in medicine

Background:

• Traditional Chinese Medicine (TCM) herb recommendations personalize treatments using compatibility principles.
• Existing methods struggle with complex multi-relational interactions between herbs, effects, and symptoms.
• This limits feature fusion and recommendation performance.

Purpose of the Study:

• To propose a novel framework, Multi-Relational Hierarchical Attention with Hybrid Knowledge Fusion (MRHAF), for enhanced TCM herb recommendation.
• To improve predictive accuracy and interpretability by modeling latent relationships and the material basis of herbal efficacy.
• To address limitations in current approaches to hybrid feature fusion.

Main Methods:

• Developed MRHAF with three core components: Herb-Efficacy-Symptom Knowledge Graph (HESKG), Herb-Symptom Interaction Graph (HSIG), and Herb-Attribute-Component Knowledge Graph (HACKG).
• Utilized multi-head and self-attention mechanisms for capturing global semantic and direct therapeutic associations.
• Integrated global semantic and local interaction features via a dual-branch attention architecture.

Main Results:

• MRHAF outperformed state-of-the-art baselines, showing improvements of 9.75% and 22.3% in Precision@10 on two benchmark datasets.
• Clinical evaluations confirmed MRHAF's ability to capture TCM formulation principles and provide reliable recommendations.
• Network pharmacology analyses validated the rationality of the recommended herb combinations.

Conclusions:

• MRHAF offers a novel perspective on herbal compatibility in TCM.
• The framework enhances predictive accuracy and interpretability in herb recommendations.
• MRHAF provides valuable guidance for clinical decision-making in TCM.