Empowering GNNs for Domain Adaptation via Denoising Target Graph
Haiyang Yu, Meng-Chieh Lee, Xiang song, Qi Zhu, Christos Faloutsos
TL;DR
This paper introduces GraphDeT, a framework that improves GNN performance on target graphs in domain adaptation by incorporating an auxiliary denoising edge loss, supported by theoretical analysis and superior experimental results.
Contribution
Proposes GraphDeT, a novel GNN training framework that integrates an auxiliary edge denoising task to enhance domain adaptation performance, backed by theoretical and empirical evidence.
Findings
Auxiliary edge denoising significantly improves GNN accuracy on target graphs.
GraphDeT outperforms existing baselines in time and regional domain shifts.
Theoretical analysis links the auxiliary task to tighter generalization bounds.
Abstract
We explore the node classification task in the context of graph domain adaptation, which uses both source and target graph structures along with source labels to enhance the generalization capabilities of Graph Neural Networks (GNNs) on target graphs. Structure domain shifts frequently occur, especially when graph data are collected at different times or from varying areas, resulting in poor performance of GNNs on target graphs. Surprisingly, we find that simply incorporating an auxiliary loss function for denoising graph edges on target graphs can be extremely effective in enhancing GNN performance on target graphs. Based on this insight, we propose our framework, GraphDeT, a framework that integrates this auxiliary edge task into GNN training for node classification under domain adaptation. Our theoretical analysis connects this auxiliary edge task to the graph generalization bound…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
1. The results are significant and demonstrate clear performance gains. 2. The authors provide a theoretical foundation, showing that auxiliary edge tasks effectively tighten the domain adaptation bounds. This theoretical insight offers a deeper understanding of their role in mitigating domain shifts.
1. The Introduction section lacks clarity and persuasive motivation, and the proposed method does not exhibit sufficient novelty. 2. The chosen baselines are outdated, with only one from 2024 studies, additional recent baselines should be included for a fair comparison. 3. The Related Work section on graph domain adaptation only covers developments up to 2024 and cites merely one paper, making it incomplete and lacking in reference value. A more comprehensive literature review is needed. 4. No c
1. The proposed method is simple and easy to implement yet yields notable performance improvement in both time and regional domain adaptation scenarios. 2. The authors establish a connection between the auxiliary edge task and generalization bound and provide a proof sketch suggesting that enforcing embedding similarity across edges can, under assumptions though, reduce classifier disagreement and thus tighten the A-distance term in the classical domain adaptation bound.
1. The proposed method leverages only label supervision from the source domain but does not transfer or model source structural patterns. Thus, the method does not explicitly learn how source structural priors can generalize to target graphs. This raises questions about whether the method truly performs domain adaptation, i.e., leveraging knowledge from the source domain, or simply applies target side self-supervised regularization. 2. There remains some disconnection between the theoretical ana
* **Empirical performance** GraphDeT yields large and consistent performance gains across both temporal (Arxiv) and regional (MAG) adaptation benchmarks. The ablation on various edge-related tasks (GAE, link prediction, denoising) provides solid empirical grounding. * **Practical simplicity** The proposed approach is simple to implement and can be readily integrated with existing GNN architectures. * **Connection to theory** The authors attempt to bridge empirical improvements and theoretical in
1. **Motivation for denoising remains insufficient.** The paper does not convincingly explain *why* the target graph requires denoising in the first place. While Section 3.1 assumes the existence of “noisy edges,” the source of such noise, its statistical characteristics, or its relation to domain shift are never justified. Without concrete motivation, the denoising task risks appearing as an *ad-hoc* regularizer rather than a principled GDA mechanism. 2. **Theoretical novelty is limited.**
1. The proposed method is simple to implement—it essentially adds a link prediction-style loss on the target graph—yet it yields large and consistent performance improvements over a range of strong baseline. 2. The experimental setup is comprehensive. The paper uses standard, challenging datasets for time-based (Arxiv) and regional (MAG) domain shifts. The comparison includes recent and relevant baselines like SPECREG and Pairwise Alignment (PA). The ablation study in Table 4 also effectively
1. The method introduces a new loss term ($l_{DeT}$), which must be balanced with the primary classification loss ($l_{cls}$), presumably with a weighting hyperparameter (e.g., $\lambda$). This is a critical detail for reproducibility and understanding the method's robustness. The paper does not mention this hyperparameter, how it was selected, or how sensitive the model's performance is to its value. A sensitivity analysis is a key missing piece. 2. The connection between the practical loss fu
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Graph Neural Networks · Domain Adaptation and Few-Shot Learning · Topic Modeling