Graph Decoding via Generalized Random Dot Product Graph

Alvaro Ciudad Serrano, Zinnera Tariq, Stelina Tarasi, Alexis Molina

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 autoencoders, inner dot product decoder, generalized random dot product, link prediction, node clustering, molecular graph

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

TL;DR: This papers present a novel graph decoding approach for improving graph autoencoder architectures.

Abstract:

Graph Neural Networks (GNNs) have established themselves as the state-of-the-art methodology for a multitude of graph-related tasks, including but not limited to link prediction, node clustering, and classification. Despite their efficacy, the performance of GNNs in encoder-decoder architectures is often constrained by the limitations inherent in traditional decoders, particularly in the reconstruction of adjacency matrices.

In this paper, we introduce a novel graph decoding approach through the use of the Generalized Random Dot Product Graph (GRDPG) as a generative model for graph decoding. This novel methodology enhances the performance of encoder-decoder architectures across a range of tasks, owing to GRDPG's better capability to capture structures embedded within adjacency matrices.

To evaluate our approach, we design a benchmark focused on graphs of varying sizes, thereby enriching the diversity of existing benchmarks for link prediction and node clustering tasks. Our experiments span a variety of tasks, encompassing both traditional benchmarks and specialized domains such as molecular graphs.

The empirical results show the capability of GRDPG on faithfully capturing properties of the original graphs while simultaneously improving the performance metrics of encoder-decoder architectures. By addressing the subtleties involved in adjacency matrix reconstruction, we elevate the overall performance of GNN-based architectures, rendering them more robust and versatile for a wide array of real-world applications, with special regard on molecular 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: 3375