Go to ICLR 2025 Conference homepageGraph Vision Networks for Link Prediction
The potential of the vision modality for enhancing graph structural awareness has long been overlooked in the mainstream graph neural network (GNN) community. In this paper, we propose a simple yet effective framework called Graph Vision Networks (GVN), which first incorporates vision awareness into Message Passing Neural Network (MPNN) and achieves effective performance for link prediction, highlighting this unexplored but promising direction. Specifically, GVNs transform graph structures into images and extract Visual Structural Features (VSFs) from those images, where VSFs are considered a novel type of structural feature. Similar to previous structural features, VSFs also mitigate the limitations of traditional MPNNs in expressive power and substructure awareness. Additionally, unlike most previous heuristic-based structural features (e.g., common-neighbor-based and path-based ones), which typically depend on fixed structural priors, VSFs are adaptive and capable of capturing varying structural insights to better suit different scenarios. Extensive experiments across seven commonly used benchmark datasets demonstrate that GVNs and their variants can significantly enhance MPNNs in link prediction tasks. Additionally, the straightforward design of the framework makes it highly compatible with current methods, providing additional performance gains to achieve new state-of-the-art performance.