Evaluating the Transfer of Class Imbalance Correction Methods from Traditional Machine Learning to Graph-Based Drug-Drug Interaction Prediction

Polypharmacy, the concurrent use of multiple medications, increases the risk of adversedrug-drug interactions that can be difficult to detect prospectively. Computational predictionof polypharmacy side effects may be particularly challenging because common and rare sideeffects are highly imbalanced. This study investigates whether commonly used class-imbalancecorrection techniques (SMOTE-style oversampling and class-weighted loss) transfer effectivelyfrom traditional machine learning to graph-based drug-drug interaction prediction using theDecagon graph convolutional network. We compare seven model conditions spanning tradi-tional machine learning (Random Forest), neural network classifiers (multilayer perceptron withand without SMOTETomek), and graph neural networks (LightGCN and Decagon variants),evaluated on 50 polypharmacy side effects from the SNAP Decagon dataset. We find thatconventional SMOTE does not transfer directly to Decagon because SMOTE generates syn-thetic observations in feature space, whereas Decagon represents interactions as graph edges;consequently, feature-space interpolation does not provide a principled way to construct biolog-ically meaningful graph edges. We also find that the definition of class weights substantiallyaffects performance: batch-based weighting produces an approximately uniform weight of 2.0under 1:1 negative sampling, whereas full-graph weighting produces much larger weights forrare side effects and is associated with reduced performance for common side effects. Finally,the common-versus-rare performance gap differs systematically between independently trainedclassifiers and jointly trained graph models, suggesting that shared-parameter architecture maycontribute to the observed gap, although the current experiments do not isolate architecturefrom other methodological differences. These findings highlight the need for graph-aware ap-proaches to imbalance handling and for controlled ablation studies that separate loss-functioneffects from class-weighting effects.