Go to ICLR 2026 Conference homepageETA: Dual Evidence-Aware Uncertainty Learning for Open-Set Graph Domain Adaptation
Keywords: Graph Domain Adaption, Open-Set Learning
TL;DR: The paper introduces a dual evidence-enhanced uncertainty-aware learning framework for the problem of open-set graph domain adaptation.
Abstract: Graph Neural Networks (GNNs) have shown great promise in node classification tasks, but their performance is often hindered by the scarcity of labeled nodes. Recently, graph domain adaptation has emerged as a promising solution to transfer knowledge from a labeled source graph to an unlabeled target graph. However, most existing methods typically rely on a closed-set assumption, which fails when unknown classes exist in the target domain. Toward this end, in this paper, we investigate the challenging open-set graph domain adaptation problem and propose a dual evidence-aware uncertainty learning framework ETA that simultaneously identifies unknown target nodes and enhances knowledge transfer under the evidential learning theory. Specifically, we adopt a dual-branch encoder to capture both implicit local structures and explicit global semantic consistency within the graph, and leverage evidential deep learning to integrate the evidence from both branches, where the resulting evidence is parameterized by a Dirichlet distribution to estimate class probabilities and enable uncertainty quantification. Based on the identified unknown target node, we further construct cross-domain neighborhoods and perform MixUp-based virtual sample generation in the latent space. Then, we introduce evidential adjacency-consistent uncertainty to evaluate uncertainty consistency across neighborhoods, which serves as auxiliary guidance for robust domain alignment. Extensive experiments on benchmark datasets demonstrate that ETA significantly outperforms state-of-the-art baselines in open-set graph domain adaptation tasks.
Primary Area: learning on graphs and other geometries & topologies
Submission Number: 5363
Loading