Denoising Graph Dissipation Model Improves Graph Representation Learning
Primary Area: representation learning for computer vision, audio, language, and other modalities
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 representation learning, diffusion generative models
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Graph-structured data are considered non-Euclidean as they provide superior representations of complex relations or interdependency. Many variants of graph neural networks (GNNs) have emerged for graph representation learning which is essentially equivalent to node feature embedding, since an instance in graph-structured data is an individual node. GNNs obtain node feature embedding with a given graph structure, however, graph representation learning tasks entail underlying factors such as homophilous relation for node classification or structure-based heuristics for link prediction. Existing graph representation learning models have been primarily developed toward focusing on task-specific factors rather than generalizing the underlying factors. We introduce Graph dissipation model that captures latent factors for any given downstream task. Graph dissipation model leverages Laplacian smoothing and subgraph sampling as a noise source in the forward diffusion process, and then learns the latent factors by capturing the intrinsic data distribution within graph structure in the denoising process. We demonstrate the effectiveness of our proposed model in two distinct graph representation learning tasks: link prediction tasks and node classification tasks, highlighting its capability to capture the underlying representational factors in various graph-related tasks.
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: 5365
Loading