Empowering GNNs for Domain Adaptation via Denoising Target Graph

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 with -distance, demonstrating such auxiliary task can imposes a constraint which tightens the bound and thereby improves generalization. The experimental results demonstrate superior performance compared to the existing baselines in handling both time and regional domain graph shifts.

💡 Research Summary

The paper addresses the problem of graph domain adaptation (GDA), where a Graph Neural Network (GNN) trained on a source graph with labels must generalize to a target graph that lacks labels and may exhibit structural shifts due to temporal or geographic changes. Existing domain adaptation techniques from computer vision cannot be directly applied because graphs are non‑grid data and node labels are heavily influenced by connectivity patterns. The authors propose a simple yet powerful framework called GraphDeT that augments the standard node‑classification training with an auxiliary edge‑denoising task on the target graph.

In GraphDeT, random “fake” edges are added to the adjacency matrix of the target graph, producing a noisy graph (\tilde A_T = A_T + A’T). The noisy graph and its node features are fed through a shared GNN to obtain node embeddings (\tilde h_T). For each node pair ((u,v)) the model computes a similarity score (\tilde s{uv} = \sigma(\phi(\tilde h_u \circ \tilde h_v))) using a Hadamard product and a small MLP. A binary cross‑entropy loss is applied to encourage high scores for true edges and low scores for the injected fake edges, forming the edge‑denoising loss (\ell_{\text{DeT}}). Simultaneously, the source graph is processed with the same GNN and a classifier head, and a standard node‑classification loss (\ell_{\text{cls}}) (KL divergence between predictions and true labels) is computed. The total objective is a weighted sum (\ell_{\text{total}} = \ell_{\text{cls}} + \lambda \ell_{\text{DeT}}).

The theoretical contribution links this auxiliary task to the classic domain adaptation bound based on the (\mathcal{H}\Delta\mathcal{H}) (A‑distance) metric introduced by Ben‑David et al. The bound states that target error is upper‑bounded by source error plus a term proportional to the A‑distance between source and target distributions. The authors prove (Proposition 3.1) that the edge‑denoising task imposes two constraints: (1) an (\ell_2) bound (\xi_1) on the distance between embeddings of adjacent nodes, and (2) a bound (\xi_2) on the disagreement probability between two classifiers—one trained only on the source ((g_1)) and a hypothetical one trained on both domains ((g_2)). Under these constraints, the A‑distance term can be bounded by a quantity that scales with (\xi_1) and (\xi_2), the size of the largest connected component, and the Lipschitz constants of the classifiers. Consequently, the auxiliary edge task effectively reduces the A‑distance, tightening the generalization bound and leading to better target performance.

Empirically, the authors evaluate GraphDeT on two large‑scale citation and academic‑graph datasets: the Arxiv citation network and the Microsoft Academic Graph (MAG). They construct temporal domain shifts (e.g., papers from 1950‑2007 vs. 2014‑2016) and regional shifts, creating ten distinct source‑target pairs. Baselines include standard empirical risk minimization (ERM), adversarial methods (DANN, IWDAN), graph‑specific adaptation methods (UDAGCN, SpecReg, StruRW, PA‑BOTH), and variants that use graph auto‑encoders or link‑prediction as auxiliary tasks. Across all settings, GraphDeT achieves the highest accuracy, with improvements ranging from 7.30 % to 21.83 % absolute over the best prior method on the Arxiv splits and 3.45 % to 26.75 % on MAG. The edge‑denoising auxiliary task consistently outperforms other edge‑related tasks, indicating that exposing the model to the full graph plus carefully crafted noise forces it to learn robust structural representations.

The paper’s contributions can be summarized as follows: (1) Identification of structural domain shift as a critical bottleneck for GNN generalization; (2) Introduction of a lightweight auxiliary edge‑denoising loss that can be seamlessly integrated with any GNN architecture; (3) Theoretical analysis that directly connects the auxiliary task to a reduction in A‑distance, thereby providing a principled justification for the observed empirical gains; (4) Extensive experiments demonstrating state‑of‑the‑art performance on realistic temporal and geographic graph shifts. GraphDeT offers a practical recipe for practitioners who need to deploy GNNs in dynamic environments where labeling new graphs is costly or impossible.