Noise Robust Graph Learning under Feature-Dependent Graph-Noise

Yeonjun In, Kanghoon Yoon, Kibum Kim, Sungchul Kim, Chanyoung Park

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 Neural Networks, Robust Graph Neural Networks, Graph Noise

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

Abstract: In real-world scenarios, node features frequently exhibit noise due to various factors, making GNNs vulnerable. Various methods enhance robustness, but they make an unrealistic assumption that the noise in node features is independent of the graph structure of node labels, restricting their practicality. To this end, we introduce more realistic noise scenario, called feature-dependent graph-noise (FDGN), where noisy node features may entail both structure and label noise, and propose a generative model to capture these causal relationships. Our proposed method, PRINGLE, outperforms baselines on commonly used benchmark datasets and newly introduced real-world graph datasets that simulate FDGN in e-commerce systems.

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: 6941