Interrogating Paradigms in Self-supervised Graph Representation Learning

Graph contrastive learning (GCL) is a newly popular paradigm for self-supervised graph representation learning and offers an alternative to reconstruction-based methods.However, it is not well understood what conditions a task must satisfy such that a given paradigm is better suited. In this paper, we investigate the role of dataset properties and augmentation strategies on the success of GCL and reconstruction-based approaches. Using the recent population augmentation graph-based analysis of self-supervised learning, we show theoretically that the success of GCL with popular augmentations is bounded by the graph edit distance between different classes. Next, we introduce a synthetic data generation process that systematically controls the amount of style vs. content in each sample- i.e. information that is irrelevant vs. relevant to the downstream task- to elucidate how graph representation learning methods perform under different dataset conditions. We empirically show that reconstruction approaches perform better when the style vs. content ratio is low and GCL with popular augmentations benefits from moderate style. Our results provide a general, systematic framework for analyzing different graph representation learning methods and demonstrate when a given approach is expected to perform well.