🤖 machine learning

Safe-Subspace Pseudo-Label Refinement for Source-Free Graph Domain Adaptation

This paper proposes SafeSubspace Pseudo-Label Refinement (S $^2$ PLR), a source-free graph domain adaptation framework that enhances model robustness by identifying a confidence-consistent safe subspace to apply hard pseudo-label supervision only to samples supported by both semantic and structural evidence, while handling uncertain samples with noise-tolerant soft regularization.

Original authors: Yingxu Wang, Xinwang Liu, Siyang Gao, Nan Yin

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yingxu Wang, Xinwang Liu, Siyang Gao, Nan Yin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a master chef who has spent years perfecting a recipe for a specific type of soup in your home kitchen (the Source Domain). You know exactly how the ingredients should look, smell, and taste. Now, you are hired to cook for a new group of people in a different city (the Target Domain). You have the same recipe book, but you can't go back to your old kitchen to check your notes, and the new kitchen has different lighting, different water quality, and the vegetables look slightly different.

The challenge is: How do you teach your new kitchen staff to cook this soup without being able to taste-test the ingredients yourself or go back to your old kitchen?

This is the problem of Source-Free Graph Domain Adaptation. In the world of AI, "Graphs" are networks of connected things (like social networks or molecules). The paper introduces a new method called S2PLR (Safe-Subspace Pseudo-Label Refinement) to solve this.

Here is how the paper explains the problem and their solution, using simple analogies:

The Problem: The "Confidently Wrong" Mistake

In the past, AI tried to solve this by guessing the labels for the new data (like guessing which vegetable is a carrot) and then teaching itself based on those guesses. This is called Self-Training.

The paper argues this is dangerous. Because the new kitchen is different, the AI might guess with 100% confidence that a potato is a carrot. If the AI trusts this wrong guess and teaches itself, it creates a "domino effect." In a graph (a network), information flows between neighbors. If the AI wrongly labels one node, that error spreads to its neighbors, corrupting the whole network. It's like a rumor starting with one confident liar; soon, the whole town believes the lie.

The Solution: The "Safe Zone" Strategy

Instead of trusting every guess the AI makes, S2PLR says: "Let's only trust the guesses that pass a strict safety check."

The authors propose creating a "Safe Subspace"—a small, secure zone where we are confident the AI's guesses are actually correct. They do this by acting like a detective with two different tools:

1. The "Expert Panel" (Semantic Confidence)

Imagine you don't just have one chef; you have a panel of three experts who all trained in your old kitchen. When they look at a new vegetable:

If all three experts agree it's a carrot, and they are all very confident, that's a good sign.
If one expert says "carrot" but the others are unsure, or if they all disagree, the AI marks this as "Suspicious."
The Goal: Only keep the guesses where the experts agree and are confident.

2. The "Neighborhood Watch" (Structural Consistency)

In a graph, things are connected. A carrot usually grows near other carrots, not next to a rock.

The AI looks at the "neighbors" of the vegetable in the new network.
If the vegetable is surrounded by other things that also look like carrots, the AI feels safer.
If the vegetable is isolated or surrounded by things that look like rocks, the AI says, "Wait, this doesn't fit the pattern," even if the experts were confident.
The Goal: Ensure the guess makes sense within its local neighborhood.

The Process: How S2PLR Works

The paper describes a four-step process to filter the data:

The Committee Check: The AI uses a team of "experts" (source models) to vote on what the data is. It throws out any guess where the experts disagree or aren't confident.
The Map Check: The AI learns a new map of the target data to see how things are connected. It checks if the "carrots" are actually grouped together in this new map.
The Safe Zone Selection: Only the data points that pass both the Expert Check and the Map Check are put into the "Safe Subspace." These are the only ones the AI uses to teach itself with "hard" labels (definite answers).
The Soft Touch for the Rest: What about the data that didn't pass the test? The paper says don't throw them away! Instead of forcing a wrong label on them, the AI uses "soft regularization." Think of this as gently nudging the AI to keep the data organized without forcing it to make a specific, potentially wrong, decision. It's like saying, "Keep these items in the 'maybe' bin and don't let them mess up the 'carrot' bin."

Why This Matters

The paper tested this on various datasets (like images turned into graphs and real-world chemical molecules). They found that:

Old methods often got "confidently wrong" and failed when the data changed.
S2PLR was much more robust. By being picky about which data it trusts, it avoided the "rumor mill" effect where one mistake ruins everything.
It achieved better results than previous methods, proving that quality of trust is more important than quantity of guesses.

The Bottom Line

The paper claims that in a world where we can't go back to our original data, we shouldn't blindly trust our AI's guesses. Instead, we should create a "Safe Subspace" where we only trust the guesses that are backed up by both expert agreement and neighborhood consistency. For the rest, we should be gentle and not force a decision. This prevents the AI from learning the wrong lessons and failing in new environments.

Technical Summary: Safe-Subspace Pseudo-Label Refinement for Source-Free Graph Domain Adaptation

Problem Definition
Source-Free Graph Domain Adaptation (SF-GDA) addresses the challenge of adapting a graph neural network trained on a labeled source domain to an unlabeled target domain when the source data is inaccessible due to privacy or storage constraints. A central obstacle in this setting is the reliability of pseudo-labels generated by a fixed source hypothesis. Under domain shifts affecting both node attributes and graph topology, source-induced predictions can become "confidently wrong." Indiscriminate self-training on such pseudo-labels can amplify systematic errors through graph message passing, leading to unbounded noise regimes where the model diverges rather than adapts. Existing methods often lack a theoretical justification for which target samples are safe to use for hard supervision, relying instead on global confidence filtering or topology alignment without explicitly characterizing the reliability region.

Methodology: S2PLR
The authors propose Safe-Subspace Pseudo-Label Refinement (S2PLR), a framework grounded in a selective-risk perspective. Instead of assuming pseudo-label noise is globally bounded, S2PLR identifies a "confidence-consistent safe subspace" where noise can be controlled and applies hard supervision only to this subset. The framework consists of four integrated modules:

Multi-Expert Uncertainty Quantification: To estimate semantic reliability, S2PLR utilizes a committee of $K_e$ pretrained source hypotheses (experts). It computes an ensemble consensus prediction and a predictive variance. Samples are filtered into a "semantic candidate set" only if they exhibit high consensus (low variance) and high confidence, mitigating the risk of overconfident errors from a single unstable model.
Target-Intrinsic Structure Learning: To address structural shifts, S2PLR learns a target-specific structural manifold using a dedicated branch trained via graph contrastive learning on unlabeled target data. This creates a metric space where local neighborhoods are meaningful and compact, independent of source-specific topology biases.
Dual-View Nested Pseudo-Label Refinement: This module intersects the semantic candidates with structural evidence. It constructs a $k$ -nearest-neighbor graph in the target-intrinsic embedding space and calculates a local confidence-consistency score. A sample enters the "safe subspace" ( $H_{\zeta,\rho}$ ) only if it passes both the semantic confidence/variance checks and the structural neighborhood consistency check.
Noise-Tolerated Regularization: For the remaining "uncertain set" ( $U$ ), which contains valuable distributional information but lacks verified labels, S2PLR avoids hard pseudo-labeling. Instead, it applies soft regularization (entropy minimization and KL divergence against the source ensemble) to control the residual risk without propagating erroneous hard labels.

Theoretical Analysis
The paper provides a theoretical analysis revisiting pseudo-label noise in SF-GDA.

Unbounded Noise: The authors acknowledge existing theory showing that under global domain shifts, source classifiers can be systematically and confidently wrong on non-empty regions of the target domain, rendering full-domain self-training unsafe.
Safe Subspace Guarantee: The authors derive conditions under which pseudo-label noise is bounded. They define a safe subspace $H_{\tau,\rho}$ where samples satisfy high source confidence ( $\tau$ ) and local neighborhood consistency ( $\rho$ ). Under a "restricted posterior discrepancy" assumption and target-manifold smoothness, they prove that the pseudo-label noise on this subspace is bounded and learnable.
Risk Decomposition: The target risk is decomposed into three components: the fitting error on the safe subspace, the intrinsic noise of the safe subspace, and the risk on the uncertain set. This decomposition theoretically justifies the strategy of using hard labels only on the safe subspace and soft regularization elsewhere.

Experimental Results
S2PLR was evaluated on image-derived graph benchmarks (MNIST, CIFAR-10) and real-world graph datasets (Mutagenicity, DD, PROTEINS, COX2, BZR, ogbg-molhiv) under node-density, edge-density, and feature-domain shifts.

Performance: S2PLR achieved state-of-the-art or competitive performance across diverse transfer settings, outperforming source-only baselines, general source-free domain adaptation methods, and existing graph-specific SF-GDA methods.
Diagnostics: Post-hoc analysis revealed a consistent precision-coverage trade-off. The final safe subspace achieved significantly higher pseudo-label precision than the initial semantic candidate set, confirming that structural verification effectively filters out unreliable samples.
Ablation: Removing any of the four modules (multi-expert quantification, confidence filtering, target structure learning, or soft regularization) led to performance degradation, validating the necessity of the integrated approach.
Visualizations: t-SNE visualizations showed that S2PLR produces more compact and well-separated target representations compared to baselines, which often exhibited class mixing.

Significance and Claims
The paper claims that S2PLR moves beyond purely empirical pseudo-label refinement or topology alignment by connecting selective bounded-noise analysis with structure-aware graph adaptation.

Theoretical Justification: It provides the first theoretical characterization of a "safe subspace" in SF-GDA where pseudo-label noise is provably bounded, distinguishing between population-level noise bounds and target-only structural certificates.
Reliability over Quantity: The framework prioritizes pseudo-label reliability over quantity, selecting a smaller but highly trustworthy subset for hard supervision while leveraging uncertain samples through noise-tolerant regularization.
Robustness: The results demonstrate that the performance gains stem from the reliability-guided safe-subspace selection mechanism rather than simply increasing the number of source hypotheses or model capacity.

The authors note a limitation: under severe structural shifts, safe-subspace selection may become overly conservative or exhibit class bias, suggesting future work on adaptive coverage control and class-aware reliability estimation.