Diffusion with Synthetic Features: Feature Imputation for Graphs with Partially Observed Features
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 neural networks, Missing features
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: For graphs with missing features, we propose a new diffusion-based imputation scheme using synthetic features.
Abstract: In this paper, we tackle learning tasks on graphs with missing features, improving the applicability of graph neural networks to real-world graph-structured data. Previous diffusion-based imputation methods overlook the presence of channels with low-variance features, and these channels contribute very little to the performance in graph learning tasks. To overcome this issue, we propose a new diffusion-based imputation scheme using synthetic features in addition to observed features. The proposed scheme first identifies channels with low-variance features via pre-diffusion and generates a synthetic feature for a randomly chosen node in each low-variance channel. Then, our diffusion process spreads the synthetic features widely while considering observed features simultaneously. Extensive experiments on graphs with various rates of missing features demonstrate the effectiveness of our scheme, achieving state-of-the-art performance in both semi-supervised node classification and link prediction.
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: 9301
Loading