Interpolating Graph Pair to Regularize Graph Classification

We present a simple and yet effective interpolation-based regularization technique, aiming to improve the generalization of Graph Neural Networks (GNNs) on supervised graph classification. leverage Mixup, an regularizer for vision, where random sample pairs their labels are interpolated create synthetic images training. Unlike with grid-like coordinates, graphs have arbitrary structure topology, which can be very sensitive any modification that alters graph's semantic meanings. This posts two unanswered questions Mixup-like schemes: Can we directly mix up pair inputs? If so, how well does such mixing strategy regularize learning GNNs? To answer these questions, propose ifMixup, first adds dummy nodes make same input size then simultaneously performs linear interpolation between aligned node feature vectors edge representations graphs. empirically show schema effectively classification learning, resulting in superior predictive accuracy popular augmentation GNN methods.