OG-SNR : Open set graph learning with structural noise robustness

Qin Zhang, MaoRiqiang, Jiexin Lu

Keywords: GNN, Open Set, Structural noise, robustness

TL;DR: We propose a robust graph learning framework that integrates structure refinement, curriculum learning, and entropy maximization to achieve noise-resistant open-set node classification on graph data.

Abstract:

Open-set graph learning aims to training graph-based models that can accurately classify known in-distribution classes while identifying and handing previously unknow classes during inference. It is critical for high-stakes, real-world applications where models frequently encounter unexpected data, including finance, security and healthcare etc. Though utilizing the message-passing mechanism,
Graph Neural Networks (GNNs) have demonstrated outstanding performance in this area by focusing on preserving structural properties, the robustness against noises is generally ignored, especially the structural noise which is inevitable in open-world scenarios of real-world applications. In this paper, we propose a novel framework to achieve open-set node classification with structure-noise robustness, which is specifically tailored for open-set graph data with structural noise and out-of-distribution (OOD) classes. Specifically, our approach refines graph structures by leveraging both node features and structural information. Furthermore, we mitigate the impact of noisy edges using a curriculum learning framework and dynamically select a subset of samples as ”pseudo-OOD nodes” during training. By incorporating an entropy maximization loss, the method achieves open-set node classification guided by node confidence scores. To the best of our knowledge, this is the first study to explore this problem. Extensive experiments validate the superiority of our proposed approach.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 16108