Link Prediction with Untrained Message Passing Layers

In this work, we explore the use of untrained message passing layers in graph neural networks for link prediction. The untrained message passing layers we consider are derived from widely used graph neural network architectures by removing trainable parameters and nonlinearities in their respective message passing layers. Experimentally, we find that untrained message passing layers can lead to competitive and even superior link prediction performance compared to fully trained message passing layers while being more efficient, especially in the presence of high-dimensional features. We also provide a theoretical analysis of untrained message passing layers in the context of link prediction and show that the inner product of features produced by untrained message passing layers relate to common neighbour and path-based topological measures which are widely used for link prediction. As such, untrained message passing layers offer a more efficient alternative to trained message passing layers in link prediction tasks, with clearer theoretical links to classical path-based heuristics.