Unsupervised Graph Domain Adaptation

• Unsupervised Graph Domain Adaptation is a set of methods that transfer knowledge from a label-rich source graph to an unlabeled target graph while managing discrepancies in node attributes and graph structure.
• It employs diverse strategies including metric-based alignment, adversarial training, and generative models to address covariate, structure, and label shifts in graph data.
• Empirical benchmarks and architectural insights demonstrate the effectiveness of these methods and highlight emerging challenges like heterophily, multi-source scenarios, and open-set adaptation.

Unsupervised Graph Domain Adaptation (UGDA), also referred to as Graph Domain Adaptation in the unsupervised target setting, studies how to transfer knowledge from a labeled or label-rich source graph to an unlabeled target graph under domain discrepancies. The task is difficult because graph data are non-IID: mismatch can arise not only in node attributes but also in graph structure, label distributions, and the neighborhood-dependent representations induced by message passing. The literature is centered on node classification, but also includes graph classification, and it has expanded from the standard single-source closed-set setting to multi-source, source-free, open-set, gradual, and noisy-label regimes (Liu et al., 2024, Cai et al., 2021).

1. Formal setting and distribution shifts

In the benchmark formulation, an attributed graph is written as $G=(\mathcal{V},\mathcal{E})$, with node attribute matrix $\mathbf{X}\in\mathbb{R}^{n\times d}$, adjacency matrix $\mathbf{A}\in\mathbb{R}^{n\times n}$, and label matrix $\mathbf{Y}\in\mathbb{R}^{N\times C}$. The node-level UGDA setting uses a labeled source graph and an unlabeled target graph, and seeks a graph model $h:\mathcal{G}\rightarrow\mathcal{Y}$ that generalizes to the target domain despite $\mathbb{P}_S(G)\neq \mathbb{P}_T(G)$ (Liu et al., 2024). A commonly used concrete instantiation is $G_S=(V_S,A_S,X_S,Y_S)$ for the source domain and $G_T=(V_T,A_T,X_T)$ for the target domain, with target node classification as the objective (Wang et al., 29 May 2025).

The benchmark literature distinguishes several kinds of shift. Covariate shift is written as

$$\mathbb{P}_S(G)\neq \mathbb{P}_T(G)\quad \text{with}\quad \mathbb{P}_S(Y\mid G)=\mathbb{P}_T(Y\mid G),$$

and is further decomposed into feature shift,

$$\mathbb{P}_S(X,Y)\neq \mathbb{P}_T(X,Y)\quad \text{with}\quad \mathbb{P}_S(Y\mid G)=\mathbb{P}_T(Y\mid G),$$

structure shift,

$\mathbf{X}\in\mathbb{R}^{n\times d}$0

and label shift,

$\mathbf{X}\in\mathbb{R}^{n\times d}$1

The benchmark emphasizes that real UGDA tasks often involve a mixture of all three shifts, and that structure shift is distinctive to graph data because neighborhood aggregation directly affects learned representations (Liu et al., 2024).

Although node classification is the dominant setting, graph-level formulations are also established. In a graph-classification view, the source and target domains are sets of graphs,

$\mathbf{X}\in\mathbb{R}^{n\times d}$2

and each graph sample is $\mathbf{X}\in\mathbb{R}^{n\times d}$3 (Cai et al., 2021). This broader formulation matters because several later methods address graph-level transfer under structural or semantic shift, and some also incorporate noisy source supervision (Wang et al., 1 Aug 2025).

2. Dominant methodological paradigms

A standard organizing principle is the distinction between metric-based alignment and adversarial alignment. Metric-based approaches minimize discrepancy between source and target latent representations using measures such as MMD or CMD, while adversarial approaches use a domain classifier and train the encoder to fool it in a DANN-style procedure (Liu et al., 2024). The benchmark groups many UGDA methods as classical DA plus deep graph embeddings, including DANE, ACDNE, UDAGCN, ASN, AdaGCN, CWGCN, SAGDA, DMGNN, and DGDA; a second group explicitly targets graph structure shift, including StruRW, JHGDA, KBL, and PairAlign; and a third focuses on GNN transferability and message passing, including GRADE, SpecReg, and A2GNN (Liu et al., 2024).

Several later papers explicitly argue that IID-style UDA is inadequate for graphs. CMPGNN states that most prior methods simply adopt existing UDA techniques developed for independent and identically distributed data to gain domain-invariant node embeddings, which do not fully consider graph structure and the message-passing mechanism of GNNs and can fail when label distribution shift exists among domains. Its central intervention is to modify the graph input itself by predicting and inserting cross-domain edges so that target nodes can directly receive messages from source nodes during GNN propagation. The method combines a shared graph encoder, a link predictor, a classifier, a reconstruction loss for source-graph edges, entropy minimization on target predictions, and a mutual-information-based identity-preserving objective intended to prevent loss of target discriminative information (Wang et al., 29 May 2025).

A distinct line adopts a generative and disentanglement-based view. DGDA assumes that graph-structured data generation is controlled by three independent types of latent variables: semantic latent variables, domain latent variables, and random latent variables. It uses variational graph auto-encoders to recover these latent variables and disentangles them via three supervised learning modules, one for label information, one for domain information, and one for uncertainty or noise induced by graph perturbation (Cai et al., 2021). This formulation treats transfer not merely as representation alignment but as separation of task-relevant semantics from domain nuisance and graph uncertainty.

These paradigms are not mutually exclusive. A plausible implication is that the field has moved from a narrow focus on latent alignment toward a broader set of interventions that include graph rewiring, latent disentanglement, and target-side structural correction.

3. Propagation, architecture, and local dependency

A major recent theme is that UGDA performance depends not only on the alignment objective but also on the behavior of the underlying GNN. A2GNN reevaluates the role of propagation in graph domain adaptation and derives a generalization bound for multi-layer GNNs. Its abstract states that, by formulating GNN Lipschitz for $\mathbf{X}\in\mathbb{R}^{n\times d}$4-layer GNNs, the target risk bound can be tighter by removing propagation layers in the source graph and stacking multiple propagation layers in the target graph (Liu et al., 2024). This directly shifts attention from post hoc alignment losses to propagation design.

The architectural perspective is developed further in unfolded GNNs. “Domain Adaptive Unfolded Graph Neural Networks” models UGNN training as a bi-level optimization problem and reports that, when transferring from the source domain to the target domain, the lower-level objective value significantly increases, which in turn increases the upper-level objective. Its proposed cascaded propagation reuses the first propagation output as the input to a second propagation stage and is proved to be non-increasing for the lower-level objective. Across six citation tasks, the CP-enhanced UGNNs improve over the best baseline by +0.74% Macro-F1 and +0.92% Micro-F1 on average, and on EN $\mathbf{X}\in\mathbb{R}^{n\times d}$5 DE, ElasticGNN$\mathbf{X}\in\mathbb{R}^{n\times d}$6 exceeds UDAGCN by +5.04% Macro-F1 and +4.39% Micro-F1 (Zhang et al., 2024).

A different theoretical argument is advanced by DFT. That work states that conditional shift can be observed only if there exists local dependencies among node features, derives generalization bounds when dependent node features are modeled using Markov chains or dependency graphs, and argues that stronger node dependency worsens generalization through a penalty involving mixing time or forest complexity. It further shows, in a simplified setting, that GCN-style message passing can increase feature correlation with depth, and proposes decorrelated GCN layers together with graph transformer layers to reduce local dependency. Representative reported results include 78.45 micro-F1 and 75.54 macro-F1 on $\mathbf{X}\in\mathbb{R}^{n\times d}$7, 75.77 micro-F1 and 73.05 macro-F1 on $\mathbf{X}\in\mathbb{R}^{n\times d}$8, 39.43 micro-F1 on Blog1 $\mathbf{X}\in\mathbb{R}^{n\times d}$9 Blog2, and 43.87 micro-F1 on Blog2 $\mathbf{A}\in\mathbb{R}^{n\times n}$0 Blog1 (Tai et al., 15 Dec 2025).

The benchmark perspective is consistent with these architecture-centered results. GDABench reports that, with appropriate neighbourhood aggregation mechanisms, simple GNN variants can even surpass state-of-the-art UGDA baselines. It also finds that multi-hop neighbors are not always beneficial, that GCN-style sum aggregation often outperforms mean or max aggregation, and that more expressive aggregators such as GAT and GIN are not always better in UGDA because they can be source-biased (Liu et al., 2024). Taken together, these works reject the common assumption that adaptation quality is determined only by an alignment loss.

4. Structural shift, heterophily, and large-gap transfer

Structural discrepancy is repeatedly identified as a bottleneck when source and target graphs differ substantially. GDABench states that, when the source and target graphs face significant distribution shifts, it is imperative to formulate strategies to effectively address and mitigate graph structural shifts (Liu et al., 2024). Several recent methods make this claim central.

RSGDA targets the case in which existing GDA methods implicitly assume homophily and therefore perform poorly when heterophily is present. Its homophily-agnostic design reconstructs both source and target into a homophilic variant and a heterophilic variant, filters and encodes each branch separately, and aligns source and target in the corresponding reconstructed spaces. The total objective combines correlation reduction, reconstruction, and alignment losses, and the paper reports state-of-the-art accuracy across all reported cross-domain tasks, including an improvement of up to 8.6% on WI $\mathbf{A}\in\mathbb{R}^{n\times n}$1 TX over the second-best method (Fang et al., 7 Feb 2026). This directly addresses the practical difficulty that the homophily level of an unlabeled target graph cannot be assessed beforehand.

Large source–target gaps motivate intermediate-domain or trajectory-based strategies. GGDA frames graph adaptation as gradual domain adaptation, generates knowledge-preserving intermediate graphs over the Fused Gromov-Wasserstein metric, and then constructs domains via vertex-based progression using close vertex selection and adaptive domain advancement. Its theoretical contribution is to concretize the intractable inter-domain distance $\mathbf{A}\in\mathbb{R}^{n\times n}$2 through implementable upper and lower bounds, and its experiments report that the best GGDA model improves over the best baseline by about +5.02% on average (Lei et al., 29 Jan 2025). DiffGDA replaces discrete intermediate steps with a continuous-time generative process based on stochastic differential equations, jointly modeling structural and semantic transitions and steering the reverse diffusion with a domain-aware guidance network. It reports consistent gains across 14 graph transfer tasks across 8 real-world datasets, and more than 50% lower total runtime than GraphAlign (Chen et al., 11 Feb 2026).

This body of work suggests that severe graph shift is not well approximated by a single global alignment step. Instead, homophily-aware reconstruction, compact domain sequences, or continuous trajectories are used to approximate the source-to-target transformation more faithfully.

5. Expanded settings beyond the standard closed-set single-source formulation

The original single-source closed-set setup has diversified substantially. In multi-source unsupervised domain adaptation on graphs, models trained on annotated source domains are transferred to an unsupervised target graph for node classification. “Selective Multi-source Adaptation for Graph” introduces a graph-modeling-based domain selector, a sub-graph node selector, and a bi-level alignment objective. Its abstract states that graph similarity is disentangled and measured with the transferability of a graph-modeling task set for source selection, node transferability is modeled within each source domain, the target is aligned to selected source data by minimizing the optimal transport distance in embedding space and by distilling the label function at the classification level, and the modules are learned in virtual training splits with a meta-learning strategy (Zhao et al., 2024).

A stricter variant is multi-source-free UGDA, where source data are inaccessible and only source pre-trained models remain. GraphATA studies this setting and performs adaptation at node granularity rather than at the model level. Each target node is parameterized with its own graph convolutional matrix, automatically aggregated from multiple source models according to local context; sparsemax is used for sparse source-model selection; and adaptation uses pseudo labels built from a representation bank and a prediction bank. The paper claims to be the first to study multi-source-free unsupervised graph domain adaptation and reports the best accuracy on all reported tasks in the main table (Zhang et al., 5 Feb 2025).

Source-free UGDA is formulated explicitly by SOGA, where the only information retained from the source domain is a well-trained source model. SOGA is model-agnostic and adapts on the target graph with Information Maximization and Structure Consistency, the latter encouraging connected or structurally similar target nodes to have similar predictions. On four cross-domain tasks, GCN-SOGA is reported to achieve a maximum improvement of about 2.1% in Macro-F1 and about 4% in Macro-AUC over baselines (Mao et al., 2021).

The closed-set assumption is also relaxed. GraphRTA addresses unsupervised open-set graph domain adaptation, in which the target label space is $\mathbf{A}\in\mathbb{R}^{n\times n}$3 and all labels outside $\mathbf{A}\in\mathbb{R}^{n\times n}$4 are collapsed into a single unknown class. Its “dual reprogramming” modifies both the target graph and the model, prunes domain-specific parameters, extends the classifier with an extra output dimension for the unknown class, and uses entropy modeled by a mixture of two Beta distributions to estimate target-known versus target-unknown membership. The paper reports that GraphRTA is consistently best or near-best across almost all tasks, especially on H-score (Zhang et al., 21 Oct 2025).

Graph-level adaptation and noisy-label adaptation further widen the scope. DGDA studies unsupervised graph domain adaptation in graph classification through latent disentanglement (Cai et al., 2021). NeGPR studies graph-level domain adaptation with noisy source labels, uses dual semantic and topology branches, a nested pseudo-label refinement mechanism in which one branch selects high-confidence target samples to guide the other, and a noise-aware regularization that is theoretically argued to mitigate pseudo-label noise even under source overfitting. Its abstract reports gains of up to 12.7% in accuracy under severe label noise (Wang et al., 1 Aug 2025).

6. Benchmarks, empirical regularities, and open directions

The empirical infrastructure for UGDA has been shaped by GDABench, described as the first comprehensive benchmark for unsupervised graph domain adaptation. It evaluates 16 algorithms across 5 datasets with 74 adaptation tasks, standardizes data processing and evaluation, and introduces the PyGDA library for training and evaluating existing UGDA methods (Liu et al., 2024). The datasets span Airport, Blog, ArnetMiner, Twitch, and MAG; the benchmark uses Micro-F1 on Airport, Blog, and ArnetMiner, Macro-F1 on MAG, and AUROC on Twitch (Liu et al., 2024).

The benchmark’s main finding is that current UGDA models vary significantly across datasets and adaptation scenarios, with no universally best method. It also reports that structural discrepancy becomes crucial under significant domain shift, while simple GNN variants with well-chosen aggregation and target-side unsupervised learning can rival or exceed many specialized baselines (Liu et al., 2024). This empirical conclusion aligns with later method-specific claims: CMPGNN argues that marginal embedding alignment can fail under label shift because it does not guarantee alignment of class-conditional distributions (Wang et al., 29 May 2025); RSGDA argues that a homophily assumption can be harmful when heterophily is present (Fang et al., 7 Feb 2026); and GraphRTA shows that closed-set assumptions are often unrealistic when the target contains unseen classes (Zhang et al., 21 Oct 2025).

Several open directions are explicit in the surveyed papers. GraphRTA points to source-free open-set GDA, semi-supervised open-set GDA, and out-of-distribution detection on graphs (Zhang et al., 21 Oct 2025). GraphATA mentions extensions to open-set graph domain adaptation and graph domain out-of-distribution settings (Zhang et al., 5 Feb 2025). More broadly, the literature indicates that future progress is likely to depend on joint treatment of structural shift, label shift, propagation behavior, accessibility constraints on source data, and robustness to target-specific phenomena such as heterophily or unknown classes. A plausible implication is that UGDA is no longer a single technique family but a collection of transfer problems linked by the same core difficulty: learning task-relevant graph representations that remain valid when both semantics and topology change across domains.