Node-CwR: Node Classification with Reject Option

Uday Bhaskar Kuchipudi, Charu Sharma, Naresh Manwani

Keywords: Node classification, graph attention networks, reject option, label noise, label smoothing, robust learning

Abstract: Graph attention networks (GAT) have been state-of-the-art GNN architecture used as the backbone for various graph learning problems. One of the key tasks in graph learning is node classification. While several works cover multiple aspects of node classification, there has yet to be an attempt to understand the behaviour of GAT models for node classification with a reject option. This paper proposes a new approach called Node-CwR, which models node classification with a reject option using GAT. We offer both cost-based and coverage-based models to include the reject option in the node classification task. Cost-based models find the optimal classifier for a given cost of rejection value. Such models are trained by minimizing the rejection and misclassification rates on unrejected samples. Coverage-based methods take coverage as input and find the optimal model for a given coverage rate. We empirically evaluate our approaches on three benchmark datasets and show their effectiveness in learning efficient reject option models for node classification tasks. We observe that, in general, cost-based methods outperform coverage-based models for reject option. Additionally, our results include robust learning of node classifiers using label smoothing in the presence of label noise. We observe that label smoothing works well to handle label noise in cost-based models, while it works adversely in coverage-based models.

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.

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

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: 9404