Learning Scalable Structural Representations for Link Prediction with Bloom Signatures

Graph neural networks (GNNs) have shown great potential in learning ongraphs, but they are known to perform sub-optimally on link prediction tasks.Existing GNNs are primarily designed to learn node-wise representations andusually fail to capture pairwise relations between target nodes, which provesto be crucial for link prediction. Recent works resort to learning moreexpressive edge-wise representations by enhancing vanilla GNNs with structuralfeatures such as labeling tricks and link prediction heuristics, but theysuffer from high computational overhead and limited scalability. To tackle thisissue, we propose to learn structural link representations by augmenting themessage-passing framework of GNNs with Bloom signatures. Bloom signatures arehashing-based compact encodings of node neighborhoods, which can be efficientlymerged to recover various types of edge-wise structural features. We furthershow that any type of neighborhood overlap-based heuristic can be estimated bya neural network that takes Bloom signatures as input. GNNs with Bloomsignatures are provably more expressive than vanilla GNNs and also morescalable than existing edge-wise models. Experimental results on five standardlink prediction benchmarks show that our proposed model achieves comparable orbetter performance than existing edge-wise GNN models while being 3-200× faster and more memory-efficient for online inference.