Disentangling Node Attributes from Graph Topology for Improved Generalizability in Link Prediction

A primary problem in link prediction lies in models memorizing training network configurations instead of learning meaningful node features. We study how to correctly disentangle topological constraints (like path length or degree biases) from innate node attributes. This paper provides experimental observations proving the lack of generalizability caused by observational biases and introduces novel formulations to achieve unbiased inductive link prediction over unseen environments.