Multi-view Graph Representation Learning via Mutual Learning for Drug RecommendationDrug Recommendation via Multi-view Graph Learning

Drug recommendation aims to recommend a combination of drugs to patients with specific symptoms based on the interactions between symptom sets and drug sets. However, most existing methods focus on the interaction between single symptom and single drug, ignoring the set-level semantic information of symptoms and drugs. Moreover, they usually rely on complex medical records (such as history treatment records, diagnosis, lab tests, and procedures) to learn the representation of symptoms and drugs. To address these issues, we propose Multi-view Graph Representation Learning via Mutual Learning for Drug Recommendation (MGRL-DR), a novel method that only requires the symptom condition of patients and can capture the item-level and set-level interaction between symptoms and drugs via collaborative learning. Our framework will construct three graphs: symptom-drug interaction graph, symptom set-drug interaction graph, and drug set-symptom interaction graph, and train three graph-based information extraction modules in a single stage to capture the item-level and set-level semantic information of symptoms and drugs simultaneously. Extensive experiments on benchmark datasets show the significant improvements of MGRL-DR over the state-of-the-art methods.