GraphFLEx: Structure Learning $\underline{\text{F}}$ramework for $\underline{\text{L}}$arge $\underline{\text{Ex}}$panding $\underline{\text{Graph}}$s

Mohit Kataria, Nikita Malik, Sandeep Kumar, Jayadeva Jayadeva

TL;DR: GraphFLEx is a unsupervised graph structure learning framework designed to handle large and expanding graphs effectively. It confines the number of relevant nodes for potential connections by leveraging clustering and coarsening techniques.

Abstract: Graph structure learning is a fundamental problem critical for interpretability and uncovering relationships in data. While graphical data is central to information representation, inferring graph structures remains challenging. Existing methods falter with expanding graphs, requiring costly relearning of the entire structure for new nodes, and face severe computational and memory demands on large graphs. To overcome these challenges, we propose $\textbf{GraphFLEx}$: a unified framework for structure learning in Large and Expanding Graphs. GraphFLEx efficiently limits potential connections to relevant nodes by leveraging clustering and coarsening techniques, significantly reducing computational costs and enhancing scalability. $\textbf{GraphFLEx}$ provides 48 flexible methods for graph structure learning by integrating diverse learning, coarsening, and clustering approaches. Extensive experiments with various GNN models demonstrate its effectiveness. Our code is available at https://anonymous.4open.science/r/Scaling_Graph_Learning-5644.

Primary Area: Deep Learning->Graph Neural Networks

Keywords: Graph Neural Networks, Unsupervised Graph Structure Learning, Graph Coarsening, Graph Clustering

Submission Number: 12117