DuoLink: A Dual Perspective on Link Prediction via Line Graphs

Astrit Tola, Surbhi Kumar, Baris Coskunuzer

Keywords: Link prediction, graph neural networks, heterophilic graphs, line graph transformation, edge-centric learning, graph representation learning, proximity indices.

TL;DR: DuoLink turns link prediction into a node‑classification task on the line graph, letting GNNs and graph transformers directly learn edge representations for improved performance and scalability over classical heuristics and prior GNN methods.

Abstract: Link prediction is a fundamental task in network science with broad applications, yet state-of-the-art Graph Neural Networks (GNNs) consistently underperform simple heuristic methods on established benchmarks. We identify two central limitations: the common two-stage pipeline of node embedding followed by separate edge decoding creates a misalignment between learned representations and the prediction task, and existing GNN models fail to effectively leverage classical heuristic features, particularly under feature heterophily. To overcome these challenges, we propose a novel reformulation of link prediction as node classification on the line graph of the original network. This reformulation enables direct modeling of edge-level structures, seamlessly integrates heuristic-based features, and aligns naturally with the inductive biases of GNN architectures. We introduce DuoLink, a versatile framework utilizing GNNs and graph transformers to learn end-to-end from both structural heuristics and learned representations. Theoretically, we show that this approach achieves enhanced expressivity and improved efficiency in capturing essential edge-level motifs. Empirically, DuoLink significantly outperforms heuristic baselines and leading GNN models across homophilic and heterophilic graphs, often by substantial margins. Our work highlights the importance of architectural alignment and bridges the long-standing gap between classical heuristics and modern graph learning.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 6413