GraphReach: Locality-Aware Graph Neural Networks using Reachability Estimations

Analyzing graphs by representing them in a low dimensional space using Graph Neural Networks (GNNs) is a promising research problem, with a lot of ongoing research. In this paper, we propose GraphReach, a position-aware GNN framework that captures the global positioning of nodes with respect to a set of fixed nodes, referred to as anchors. The model incorporates the reachability of nodes from anchors and vice versa, captured through multiple random walks. We use random walk since it captures the locality better. An extensive experimental evaluation comparing different variants of the model provides an insight into its ability to perform well on the downstream prediction tasks. GraphReach significantly outperforms the current state-of-the-art models and achieves an improvement of up to 31 and 11 percentage points, measured in ROC AUC, for the standard pairwise node classification and link prediction tasks, respectively.