SMRGAT: A traditional Chinese herb recommendation model based on a multi-graph residual attention network and semantic knowledge fusion

Ethnopharmacological relevance: Traditional Chinese Medicine (TCM) prescriptions are a product of the Chinese medical theory's distinct thinking and clinical experience. TCM practitioners treat diseases by enhancing the efficacy of TCM prescriptions and reducing their poisonous effects. Some TCM herb recommendation methods have been provided for curing the given symptoms to generate a group of herbs according to the TCM principles. However, they ignored the symptoms' semantic characteristics and herbs' different effects on symptoms. Aim of the study: We aim to recommend TCM herbs by considering symptoms' semantic information and the strength of different herbs in curing symptoms. Materials and methods: We propose a herb recommendation model named Multi-Graph Residual Attention Network and Semantic Knowledge Fusion (SMRGAT) to address these problems. Concretely, it uses a multi-head attention mechanism to focus on herbs' different effects on symptoms. Meanwhile, it augments entities' features with a residual network structure while incorporating symptoms' semantic information and external knowledge of herbs. We will verify the effect of SMRGAT on the existing public datasets and the datasets that we have collected and cleaned. Results: Compared with the current best TCM herb recommendation model, on the public dataset, SMRGAT were increased by 15.11%, 20.60%, and 18.25% in Precision@5, Recall@5, and F1 - score@5, respectively; on ours, respectively increased by 9.72%, 9.03%, 9.24%. Conclusions: Our experimental results on two datasets indicate that SMRGAT is capable of recommending herbs with greater precision and outperforms several comparison methods. It can provide a basis for assisting TCM clinical prescriptions.