Spectrum-guided Multi-view Graph Fusion
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
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: unsupervised learning, multi-view graph, spectral graph theory, embedding, clustering
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: Unsupervised learning on multi-view graphs by approximating a graph fusion with desired spectral properties.
Abstract: Multi-view graphs capture diverse relations among entities through graph views and individual characteristics via attribute views, presenting a challenge for unsupervised learning due to potential conflicts across views. Existing approaches often lack efficacy, efficiency, and the ability to explicitly control view contributions. In this paper, we present SMGF, a novel graph fusion framework that approximates underlying entity connections by aggregating view-specific graph structures. We construct a multi-view Laplacian $\mathcal{L}$ from normalized Laplacian matrices representing all views. View weights are determined through the optimization of two objectives derived from $\mathcal{L}$'s spectral properties, which exploit the eigenvalue gap and enhance connectivity.
Comprehensive experiments on six real-world datasets showcase the superior performance of SMGF in node embedding and clustering results, along with its efficiency and scalability. SMGF offers a promising solution for unsupervised learning on multi-view graphs, addressing the challenge of interpretably combining diverse and potentially conflicting information from both graph and attribute views.
The source code of SMGF is available at \url{https://anonymous.4open.science/r/SMGF-E903/}.
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: 5026
Loading