Graph Representation Learning for Drug Package Recommendation: Discriminative and Generative Approaches

The presentation introduces the urgent need for intelligent drug recommendation within the broader context of smart healthcare, highlighting the shortage and uneven distribution of medical resources in China and the potential of AI and large‑scale electronic health records to improve treatment quality.

Two major challenges are identified: (1) drug recommendation is a package recommendation problem where multiple medicines must be suggested simultaneously, and (2) drug–drug interactions (cooperative, antagonistic, or toxic) are highly personalized and must be modeled.

Graph representation learning is proposed as a solution. The discriminative approach (DPR) uses a pre‑training stage based on NCF with BPR loss to obtain patient and drug embeddings, then constructs interaction‑aware drug graphs (weighted or attributed) and applies GNNs to produce package representations for scoring.

Key components of DPR include: (i) pre‑training with MLP for structured features and LSTM for textual clinical notes; (ii) conversion of drug interaction matrices into heterogeneous graphs, retaining only annotated or frequent edges; (iii) two graph variants—DPR‑WG (weighted graph with mask‑guided edge updates) and DPR‑AG (attributed graph with edge embeddings and self‑attention); and (iv) training with BPR loss and negative sampling.

The generative approach (DPG) reformulates the task as sequence generation using RNNs, enhanced with interaction‑aware vectors and a mask to capture patient‑specific effects. Two training strategies are employed: maximum‑likelihood estimation (MLE) for initial learning and policy‑gradient reinforcement learning (with SCST baseline) to optimize an order‑independent F‑value reward, ensuring generated packages respect drug interaction constraints and avoid duplicate items.

Experimental results on real electronic medical record data from a large tertiary hospital show that both DPR and DPG outperform existing baselines, with the reinforcement‑learning‑based generative model achieving the highest F‑value. Ablation studies confirm the importance of interaction graphs, mask vectors, and RL components.

The work concludes that interaction‑aware graph learning effectively addresses personalized drug package recommendation, and that generative models provide a more flexible and powerful framework for future healthcare recommendation systems.