The article presents a study on the use of Graph Neural Networks (GNNs) for link prediction in heterophilic graphs. The authors introduce GRAFF-LP, an extension to the physics-inspired GRAFF model, to improve link prediction performance. Tested on a collection of heterophilic graphs, GRAFF-LP surpasses previous methods, showcasing flexibility in various contexts and achieving relative AUROC improvements of up to 26.7%. The study suggests the potential impact of link prediction under heterophily in several applications, including recommender systems.
Publication date: 23 Feb 2024
Project Page: https://doi.org/XXXXXXX.XXXXXXX
Paper: https://arxiv.org/pdf/2402.14802
