Beyond Homophily: Class imbalance graph classification via Rewiring Graph of Graphs.

Graph Neural Networks (GNNs) have gained prominence as a leading paradigm for graph encoding, achieving notable success in graph classification tasks. This success, however, heavily relies on the assumption of the balanced class distribution in the training data, which often does not align with real-world scenarios. In the face of imbalanced class distributions, the classification results tend to be suboptimal. Previous research have shown that Graph of Graphs(GoG) can effectively capture inter-graph supervisory signals, thereby aiding in the representation of the minority graphs. We argue that existing GoG strategies rooted in the assumption of homophily provide reliable supervision primarily for majority class graphs, while remaining unreliable for minority classes. To address this issue, we introduce a novel framework called GraphBHR (Beyond Homophily Rewiring Graph of Graphs). GraphBHR supplements the GoG with additional heterophily perspectives, allowing for the provision of reasonable supervisory signals for minority classes. To further enhance the network reliability, we have introduced a graph rewiring strategy that optimizes the initial inter-graph relationships. This is followed by GoG propagation for representation learning. We also employ consistency contrastive loss and focal loss to optimize graph representation. Extensive experiments on multi-scale datasets have shown the effectiveness of GraphBHR in handling imbalanced graph classification tasks.