Syndrome-Aware Herb Recommendation with Heterogeneous Graph Neural Network

The herb recommender system usually induces the implicit syndrome representations based on TCM prescriptions to generate related herbs as a treatment to cure a given symptom set. Previous methods primarily focus on modeling the interaction between symptoms (or diseases) and herbs without explicitly considering the syndrome information. As a result, these methods only capture the coarse-grained syndrome information. In this paper, we propose a new method to incorporate the explicit syndrome information for herb recommendation. To model the coarse-grained interaction between diseases and herbs within a specific syndrome class, we employ clustering algorithms to obtain the syndrome class, and apply the graph convolution network (GCN) on multiple disease-herb bipartite subgraphs. Next, we model the fine-grained interaction upon the syndrome-herb graph. Further, we propose a syndrome-aware heterogeneous graph neural network architecture, which integrates the syndrome information into the GCN message propagation process by combining the coarse-grained and fine-grained information of the interactions. The experimental results on the real TCM dataset demonstrate the improvements over state-of-the-art herb recommendation methods, further validate the effectiveness of our model.