GTAP: Graph Topology-Aware Pre-Training for Graph Classification
TL;DR
GTAP (Graph Topology-Aware Pre-training), a self-supervised initialization framework for within-dataset graph classification, improves over a matched GCN trained from scratch and achieves competitive accuracy against published baselines on most datasets with available results.
Abstract
Graph Convolutional Networks (GCNs) are widely used for graph-structured data, but graph pre-training still lacks simple structural objectives that can be applied when semantic annotations are unavailable. This paper proposes GTAP (Graph Topology-Aware Pre-training), a self-supervised initialization framework for within-dataset graph classification. GTAP constructs derived high-order neighborhood graphs from powers of the adjacency matrix and groups them by parity, using odd-order derived neighborhood graphs as the positive class and even-order derived neighborhood graphs as the negative class. The objective trains a lightweight GCN encoder to distinguish multi-order structural views before supervised downstream training. GTAP pre-training uses only unlabeled graph structure and node features from the benchmark dataset. The downstream classifier is then evaluated with fixed 10-fold cross-validation, with label supervision restricted to the training split of each fold. Experiments on seven public graph classification benchmarks show that GTAP improves over a matched GCN trained from scratch and achieves competitive accuracy against published baselines on most datasets with available results. The results support parity-based multi-order structural pre-training as a within-dataset unlabeled signal, while fully inductive pre-training and cross-dataset transfer remain important future work.