Self-Supervised Graph Representation Learning via Topology Transformations

We present the Topology Transformation Equivariant Representation learning, a general paradigm of self-supervised learning for node representations graph data to enable wide applicability Graph Convolutional Neural Networks (GCNNs). formalize proposed model from an information-theoretic perspective, by maximizing mutual information between topology transformations and before after transformations. derive that such can be relaxed minimizing cross entropy applied transformation its estimation representations. In particular, we seek sample subset pairs original flip edge connectivity each pair transform topology. Then, self-train representation encoder learn reconstructing feature transformed graphs. experiments, apply downstream classification, classification link prediction tasks, results show method outperforms state-of-the-art unsupervised approaches.