One Graph, Every Node, Hidden Task: HIDE, a Protocol for Graph Class-Incremental Learning

Graph class-incremental learning (GCIL) methods are compared by their accuracy under the CGLB benchmark protocol. CGLB supplies the task identity the class-incremental setting is meant to withhold, in two ways: test nodes are scored one task at a time, and each node is scored on the subgraph of its own task. We measure what each of them, and each other evaluation choice, contributes to reported accuracy. We introduce HIDE (Hidden-task IDentity Evaluation), which scores every test node in one batch, on the full graph, over every class seen so far, with accuracy averaged over nodes, and we vary each choice independently. We evaluate six methods on eight datasets, using one shared graph per dataset, alongside two reference classifiers with no continual-learning mechanism, one of them trained jointly on all tasks. Every method that does not already collapse under CGLB loses accuracyunder HIDE; the one ranked first under CGLB on seven of eight datasets ranks last under HIDE on seven of eight, with its per-node accu racy falling from 97% to 32% on ogbn-arxiv and from 87% to 12% at the widest. On ogbn-arxiv, no method beats either reference classifier. We release the shared graphs, the splits, the reference classifiers, and every per-seed result. We propose assessing novel GCIL methods on HIDE, where they are evaluated under realistic, more challenging hidden-task node-by-node class-incremental learning.