Feature Driven Graph Coarsening for Scaling Graph Representation Learning

Manoj Kumar, Sumit Kumar, Vipul Kumar Singh, Sandeep Kumar

Keywords: Graph Coarsening, Graph Neural Networks

Abstract:

Graphical modelling for structured data analysis has gained prominence across numerous domains. A significant computational challenge lies in efficiently capturing complex relationships within large-scale graph structures. Graph coarsening, which reduces graph size by merging nodes and edges into supernodes and superedges, enhances scalability and is crucial for graph neural networks (GNNs). However, current methods either construct graphs from large-scale attribute data or assume a pre-existing graph before coarsening, limiting their applicability, especially in domains like healthcare and finance where graph structure is often unavailable. In this paper, we present a novel framework that directly learns a coarsened graph from attribute information, reducing computational complexity and enhancing robustness against adversarial attacks, which commonly target vulnerabilities in graph structures. By integrating label information, our framework also enables semi-supervised learning, leading to improved performance on downstream tasks. Extensive experiments show that our method outperforms state-of-the-art coarsening techniques in both accuracy and computational efficiency.

Primary Area: learning on graphs and other geometries & topologies

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 12838