Measuring Graph Similarity Using Transfer Cost of Forster Distributions

Lin Li, Bharat Singhal, Wei Zhang, Jr-Shin Li

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.

Keywords: Graph similarity, Foster distributions

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: In recent years, optimal transport-based distance metrics have shown to be effective similarity and dissimilarity measures for tackling learning problems involving network data. Prominent examples range from graph classification and community detection to object matching. However, the high computational complexity of calculating optimal transport costs substantially confines their applications to large-scale networks. To address this challenge, in this paper, we introduce a probability distribution on the set of edges of a graph, referred to as the Foster distribution of the graph, by extending Foster's theorem from electrical to general networks. Then, we represent Foster distributions as probability measures on the real line and estimate the Wasserstein metric between the corresponding probability measures to quantify graph similarity. The applicability of the proposed approach is corroborated on diverse graph-structured datasets, through which we particularly demonstrate the high efficiency of computing the proposed graph distance for sparse graphs.

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: 8671