Edge contrastive learning for link prediction

Link prediction is a critical task within the realm of graph machine learning. While recent advancements mainly emphasize node representation learning, the rich information encapsulated within edges, proven advantageous in various graph-related tasks, has been somewhat overlooked. To bridge the gap, this paper explores the potential of incorporating edge representation learning for link prediction and identifies three inherent challenges associated with this approach. We introduce the Edge Contrastive Learning for Link Prediction (ECLiP) framework to tackle these challenges. ECLiP integrates edge information into node representations through edge-level contrastive learning, with a distinctive perspective on treating edges, rather than nodes, as the units of instance discrimination. We first illustrate the implementation of this framework using an established edge representation learning method. However, it incurs significant additional training overhead when the number of edges is huge. To mitigate this issue, we present a computationally efficient variant employing a multi-layer perceptron (MLP) for direct edge representation learning. Conducting rigorous experiments across eight distinct datasets with node counts spanning from 2k to 235k, we demonstrate a noteworthy improvement of over 10% on certain datasets, validating the efficacy of our proposed methodology.