Graph Contrastive Learning with Implicit Augmentations

Graph Neural Networks (GNN) and their variants To generate different views (augmentations) for the [1, 2, 3, 4] have achieved state-of-the-art performance input graphs, existing GCL methods at both node-level on both graph-level and node-level tasks such as social [10, 11, 12] and graph-level [13, 14, 15] rely on augmentations network analysis [5], molecular interactions [6] and based on random perturbations, including recommender systems [7]. In many scenarios, training randomly adding or dropping the graph's edges or nodes, end-to-end GNN models with supervision is impractical and shuffling or masking node attributes. These augmentation as label information is frequently unavailable or difficult approaches are based on a strong assumption to retrieve. Meanwhile, Self-Supervised Learning (SSL), that a small random perturbation will not alter the semantic an unsupervised method that first trains model on auxiliary property of the original graph [13]. Nevertheless, tasks without label information and then uses the this assumption involves two critical issues. First, manipulations learned embeddings for downstream tasks, has gradually of edges and nodes on graphs in specific gained popularity in recent literature [8]. As one domains can ruin the data.