PaperlayerView exact revision on arXiv

Source-linked AI summary

GATTA: Graph Active Learning with Test-Time Augmentation

Zsombor Bánfi, András Gézsi, András Formanek

arXiv:2608.15084v1cs.LGcs.AI

TL;DR

Graph active learning needs reliable uncertainty estimates while graph-structured applications of test-time augmentation remain largely unexplored. GATTA aggregates predictions across augmented graph views with consistency filtering, and shows that simple uncertainty methods benefit most, often matching or outperforming more sophisticated approaches without architectural changes or retraining.

  • Problem

    Test-time augmentation remains largely unexplored for graph-structured data, where active learning also faces costly labeling demands and needs reliable uncertainty estimates.

  • Method

    GATTA is a plug-and-play graph active learning framework that estimates acquisition uncertainty across augmented graph views using score or prediction aggregation with consistency filtering.

  • Results

    +3.03% average: GATTA-enhanced Entropy gains this amount, while simple uncertainty methods benefit most and GATTA matches or exceeds MC Dropout on 4 of 5 datasets.

  • Takeaways & Limitations

    Simple uncertainty strategies augmented with GATTA can achieve competitive active learning performance with lower implementation effort and without architectural modifications or retraining.

  • Takeaways & Limitations

    Evaluation is limited to transductive node classification on five homophilic citation and co-purchase networks, leaving other graph tasks, settings, and domains unexplored.

Abstract

from arXiv · show

Test-time augmentation (TTA) has proven effective for improving model robustness and uncertainty estimation in computer vision, yet its application to graph-structured data remains largely unexplored. We introduce GATTA (Graph Active Learning with Test-Time Augmentation), a framework for enhancing active learning by aggregating predictions across multiple augmented views to produce more reliable uncertainty estimates. To address the challenge of label-preserving graph augmentations, GATTA incorporates a consistency-based filtering mechanism that discards augmented views yielding unreliable predictions. We systematically evaluate GATTA across multiple graph datasets, GNN architectures, and acquisition strategies. Our results show that simple uncertainty-based methods, such as Entropy and Least Confidence, benefit most from TTA, achieving performance competitive with more sophisticated and computationally expensive approaches. GATTA generalizes across architectures, outperforms model-side ensemble methods such as MC Dropout. We further show that GATTA scales efficiently with both ensemble size and graph size. Extensive analysis of augmentation types, strengths, and filtering strategies provides practical guidelines for effective deployment. Our findings demonstrate that augmenting simple methods with TTA offers a more efficient path to strong active learning performance than engineering complex acquisition functions, enabling practitioners to achieve competitive results with lower computational overhead and reduced implementation complexity.

1 Introduction

The introduction identifies graph labeling bottlenecks and the difficulty of uncertainty estimation in structurally dependent data. It presents GATTA as a test-time augmentation framework for graph active learning that uses aggregation and consistency filtering to improve practical acquisition.

  • Motivation: GNNs achieve strong performance across relational-data applications, but many high-performing models require substantial labeled data, creating a significant labeling bottleneck.The challenge is especially acute in scientific and industrial domains.
  • Problem: Active learning reduces labeling demands by selecting informative nodes with strategies such as Least Confidence, Entropy, and BALD.Graph structure and non-i.i.d. dependencies make uncertainty estimation particularly challenging.
  • Motivation: Although test-time augmentation improves uncertainty estimation in computer vision, its application to graph-structured data remains largely unexplored.TTA generates perturbed input views during inference and aggregates their predictions.
  • Contribution: GATTA integrates graph-specific test-time augmentations into graph active learning through Score Aggregation, Prediction Aggregation, and consistency-based filtering.The filtering mechanism discards potentially misleading perturbations while preserving label-relevant properties.
  • Contribution: Experiments show that simple uncertainty methods with GATTA match complex acquisition functions at lower computational cost.The framework also provides deployment guidelines for augmentation type, strength, and ensemble size.

2 Related Work

Related work highlights the structural difficulty of uncertainty estimation in graph active learning, the computational cost of sophisticated uncertainty methods, and the unresolved challenge of transferring test-time augmentation to graphs. GATTA addresses this gap by integrating graph-specific consistency filtering into test-time augmentation for graph active learning.

  • Graph Active Learning: Graph active learning is challenging because node predictions depend on multi-hop neighborhoods, violating the independence assumption behind classical uncertainty sampling.This interconnected structure complicates uncertainty estimation.
  • Graph Active Learning: Early graph-specific methods adapted classical strategies using graph embeddings, uncertainty, centrality, bandits, or expected error reduction.AGE combined embeddings with uncertainty and centrality; ANRMAB formulated selection as a multi-armed bandit; GEEM targeted expected error reduction.
  • Uncertainty Quantification: More recent approaches quantify uncertainty with jackknife estimators or epistemic-uncertainty approximations, but sophisticated methods remain computationally demanding.JuryGCN uses jackknife estimators, while Multiple Pseudo-Labels and Expected Single Pseudo-Label provide practical approximations.
  • Uncertainty Quantification: Bayesian methods, ensembles, Monte Carlo Dropout, and temperature scaling improve graph uncertainty estimation but often require architectural modifications or additional training procedures.These requirements can limit applicability to existing models.
  • Test-Time Augmentation: Test-time augmentation aggregates predictions across transformed views to capture prediction variance, and prior active-learning work showed improved uncertainty estimation for medical image annotation.Whether this mechanism transfers to graph neural networks remains open because graph predictions jointly depend on node attributes and relational structure.
  • Graph Augmentation: Graph augmentations risk changing semantic evidence because they modify node features and local connectivity, making label preservation more difficult than for common image transformations.GATTA is presented as the first framework to systematically integrate test-time augmentation into graph active learning and uses a graph-specific consistency mechanism for non-label-invariant perturbations.

3 Method

GATTA is a plug-and-play graph active learning module that estimates acquisition uncertainty across stochastic graph augmentations rather than a single graph realization. It combines augmentation, score or prediction aggregation, and consistency filtering to focus estimates on model-consistent perturbations.

  • Framework: GATTA enhances transductive node-classification active learning without architectural changes by aggregating predictions from perturbed graph views.The framework iteratively acquires labels for unlabeled nodes to improve node classification performance.
  • Framework: GATTA estimates uncertainty-based acquisition scores over a stochastic augmentation distribution, favoring uncertainty that persists under plausible local graph perturbations.Because the expectation is generally intractable, Monte Carlo samples from augmented views and the original graph approximate it.
  • Aggregation strategies: GATTA-S averages acquisition scores computed independently for each augmented view, whereas GATTA-P applies the acquisition function after averaging predictive distributions.GATTA-P requires only one acquisition-function evaluation but is more sensitive to disagreement from structurally unstable, non-label-preserving augmentations.
  • Graph augmentations: The method uses feature masking, feature noising, and edge dropout to expose uncertainty from feature reliance, controlled feature perturbations, and structural dependencies.Feature masking zeros node features with probability pmask, feature noising adds Gaussian noise N(0, σ^2_noise), and edge dropout removes edges with probability pdrop.
  • Consistency filtering: Consistency filtering excludes augmentation-induced semantic drift by restricting uncertainty estimation to views whose predictions remain consistent with the original graph.The filtered scores approximate a conditional expectation over model-consistent perturbations by averaging only consistent augmented views.

4 Active Learning Protocol

The paper follows a standard transductive graph active learning protocol, retraining the GNN after each acquisition round and evaluating methods under controlled experimental conditions. It uses a compact labeling budget and shows that filtering is important for effective augmentation strengths, especially for GATTA-P.

  • Protocol: Initially, one node per class is labeled; a GNN is retrained from scratch, unlabeled nodes are scored, and the highest-scoring node is added each round.This protocol ensures that performance improvements arise from the acquisition strategy rather than continued model training.
  • Protocol: Unless stated otherwise, one node is acquired per iteration with a budget of 4C, where C denotes the number of classes.The protocol supports direct comparison with prior graph active learning work and keeps the final labeled set substantially smaller than standard training splits.
  • Experimental controls: Results average 25 independent trials, combining 5 random initial labeled pools with 5 random model initializations.Training hyperparameters remain identical between each baseline and its GATTA-enhanced counterpart, with GATTA modifying only inference-time acquisition scores.
  • Augmentation strength: Improvements concentrate at higher augmentation strengths, approximately 0.3–0.5, across noise variance and dropout probability settings.The evaluated ranges are noise variance σ2_noise ∈[0, 0.5] and dropout probability pdrop ∈[0, 0.5].
  • Augmentation strength: GATTA-S benefits without filtering, whereas filtering broadens its effective-strength region; GATTA-P requires filtering to achieve any improvement.These findings are reported for FN+ED using heatmaps of accuracy gains (%) across augmentation strengths.

5 Sensitivity and Configuration Analysis

GATTA’s configuration study finds that combined, stronger augmentations with consistency filtering produce the most informative uncertainty estimates. Moderate ensemble sizes capture most gains, while runtime remains efficient for simple acquisition strategies and GATTA-P avoids the scaling problems of expensive ones.

  • Augmentation Types: Feature Noising improves performance more than Feature Masking, yielding 0.69 ± 0.67% versus 0.03 ± 0.83%.Additive noise preserves feature scale during neighborhood aggregation, whereas masking creates information voids that propagate through message passing.
  • Augmentation Types: Combined augmentations outperform single augmentations, with FN+ED achieving the largest average improvement of +1.12 ± 1.05%.Multi-modal perturbations produce more informative uncertainty signals when both augmentation types contribute meaningfully to the ensemble.
  • Augmentation Strength and Filtering: With filtering, stronger perturbations perform best, with noise ∈[0.3, 0.5] and pdrop ∈[0.3, 0.5] achieving improvements up to 3.0 over baseline.Stronger perturbations expose more informative uncertainty signals when label-inconsistent augmentations are filtered.
  • Augmentation Strength and Filtering: GATTA-P is highly sensitive to filtering, declining by up to −0.5% without filtering but improving by 1.96 ± 1.27% with filtering.Without filtering, GATTA-P fails overall at 0.12 ± 1.04%; GATTA-S is more robust to augmentation strength and consistency filtering.
  • Ensemble Size and Runtime: Ensemble gains scale logarithmically, reaching +2.65±0.93% around N ≈200, while each additional 100 views beyond N = 500 adds less than 0.1%.The reported relationship is Performance ≈baseline+0.37×log(N+1), indicating diminishing returns after moderate ensemble sizes.
  • Ensemble Size and Runtime: For Entropy, GATTA-P and GATTA-S both require ∼2× baseline time at N = 500, whereas GATTA-P remains efficient for expensive AGE acquisition.GATTA-S scales poorly with AGE because it requires N+1 AGE evaluations; runtime scaling remains consistent across CoraML and PubMed.

6 Results

Across five graph datasets and multiple GNN architectures and acquisition strategies, GATTA most strongly improves simple uncertainty-based active learning. Its gains emerge early, persist throughout learning, generalize across architectures, and generally match or exceed MC Dropout.

  • R1. Selective benefits for uncertainty methods: Least Confidence gains +2.87% average and Entropy gains +3.03%, with GATTA-S exceeding +5% on multiple datasets.GATTA-S outperforms GATTA-P for these methods by preserving per-view uncertainty through score-level aggregation.
  • R2. Early and persistent benefits: GATTA improves sample efficiency from the earliest acquisition rounds, and its performance gap over baselines remains consistent throughout learning.The gains therefore extend beyond merely accelerating early learning.
  • R3. Dataset-dependent performance: GATTA’s effectiveness varies substantially within citation and co-purchase networks, with Entropy-S and LC-S exceeding +5% on AmazonComputers using SGC.The variation suggests that dataset-specific graph properties matter more than broad network categories.
  • R4. Architectural generalization: GATTA generalizes across GCN, SGC, GAT, and GraphSAGE, with consistent improvements on most datasets.GATTA-S remains stronger for GAT, whereas GraphSAGE shows more dataset-dependent behavior.
  • R5. Comparison with MC Dropout: GATTA matches or exceeds MC Dropout on 4 of 5 datasets, with gains up to +3.73% on PubMed.Combining GATTA with MC Dropout provides no consistent improvement and can reduce AmazonComputers performance by 8.39%.

7 Discussion

GATTA mainly improves the predictive distribution used by acquisition functions, so its benefits are greatest for methods that rely heavily on predictive uncertainty. The discussion recommends uncertainty-focused methods, consistency-aware aggregation choices, moderate-to-high augmentation strengths, and validation on representative target-graph subsets.

  • Interpretation: GATTA benefits Entropy and Least Confidence most because these acquisition functions depend almost entirely on predictive uncertainty.Its primary effect is improving the predictive distribution rather than the acquisition objective itself.
  • Interpretation: GATTA-P is sensitive to semantic drift because inconsistent predictions directly affect aggregation, whereas GATTA-S is more robust by averaging independently computed acquisition scores.Consistency filtering restricts GATTA-P aggregation to model-consistent perturbations; GATTA-S performs well even without filtering.
  • Practical guidelines: Practitioners should prioritize simple uncertainty methods and use GATTA-P with filtering when acquisition functions are expensive.The discussion presents these as practical deployment guidelines based on the differing behavior of acquisition strategies and aggregation variants.
  • Practical guidelines: Recommended settings are noise ∈[0.4, 0.5], pdrop ∈[0.3, 0.5], and ensemble size N = 500.These settings consistently provide a favorable trade-off between performance and runtime; augmentation choices should be validated on a representative target-graph subset.

8 Conclusions and Limitations … A.3 Model Details

GATTA integrates test-time augmentation into graph active learning, with simple uncertainty-based acquisition benefiting most, while the study identifies important scope limitations and establishes a standardized experimental protocol. The evaluation uses five benchmark networks, fixed literature hyperparameters, and iterative label acquisition under a constrained budget.

  • 8 Conclusions and Limitations: GATTA systematically integrates test-time augmentation into graph active learning and improves uncertainty estimation, with effectiveness depending on acquisition-process design.The framework’s contribution is both methodological and analytical: it integrates TTA and characterizes design choices governing its effectiveness.
  • 8 Conclusions and Limitations: Simple uncertainty-based acquisition strategies benefit most from improved predictive uncertainty, often matching or outperforming substantially more sophisticated approaches.This finding supports using TTA to strengthen simpler acquisition methods rather than engineering more complex task-specific strategies.
  • 8 Conclusions and Limitations: The evaluation is limited to transductive node classification on five homophilic citation and co-purchase networks, leaving other tasks, settings, and graph types unexplored.Unexamined areas include link prediction, graph classification, inductive settings, and heterophilic or dynamic graphs.
  • 8 Conclusions and Limitations: Future work should develop adaptive augmentation strategies and clarify how homophily, structural sparsity, and feature informativeness affect TTA-based uncertainty estimation.The paper calls for perturbation strengths that adapt to model confidence or graph structure, alongside deeper theoretical analysis.
  • 8 Conclusions and Limitations: GATTA offers a general graph-specific mechanism for improving uncertainty estimation while making graph active learning more accessible to practitioners.Rather than adding another task-specific acquisition strategy, it integrates TTA and systematically characterizes the choices that govern performance.
  • A Experimental Setup: The experimental implementation follows Fuchsgruber et al. (2024), including its training procedures, datasets, and active learning protocols.The authors state that their GATTA implementation is publicly available in the cited repository.
  • A.1 Active Learning Protocol: The active learning protocol initializes one labeled node per class, trains to convergence, acquires labels iteratively, and reports final held-out test accuracy after retraining.The classifier is retrained from scratch after each iteration; unless stated otherwise, one label is acquired per iteration with budget 4C.
  • A.2 Datasets: Experiments use three citation networks and two co-purchase networks: CoraML, Citeseer, PubMed, AmazonComputers, and AmazonPhotos.Citation-network edges represent citations, whereas co-purchase-network edges indicate frequent product co-purchases.

A.4 Training and Evaluation Details · B Information-Theoretic Interpretation of GATTA Aggregation · C Filtering variants

The paper specifies reproducible training and evaluation protocols, interprets GATTA aggregation through entropy and augmentation-induced disagreement, and compares hard and confidence-weighted filtering. The analysis concludes that consistency-aware uniform weighting is preferable to soft weighting alone or Firm filtering in the reported experiments.

  • A.4 Training and Evaluation Details: All models use binary cross-entropy with Adam, learning rate 10−3, weight decay 10−3, and validation-loss early stopping after 100 iterations of patience.These settings define the training protocol across experiments.
  • A.4 Training and Evaluation Details: Each configuration is evaluated over five dataset splits and five independent initializations, totaling 25 runs, with fixed validation and test node sets.The test set contains 20% of nodes and is inaccessible to acquisition strategies; each split also reuses a 20% validation set.
  • A.4 Training and Evaluation Details: Test accuracy is the primary metric, and every reported accuracy gain is measured against the matching configuration without test-time augmentation.The baseline preserves the same dataset, model, and acquisition function.
  • B Information-Theoretic Interpretation of GATTA Aggregation: For entropy acquisition, GATTA-S estimates expected entropy across augmented views, whereas GATTA-P computes entropy after averaging their predictive distributions.The framework applies to arbitrary uncertainty-based acquisition functions, while this decomposition explains the differing sensitivity of GATTA-P and GATTA-S.
  • B Information-Theoretic Interpretation of GATTA Aggregation: GATTA-P additionally includes I(Y ; φ | G), the mutual information between predicted class and sampled augmentation conditioned on the observed graph, representing cross-view prediction variability.GATTA-S estimates only the average uncertainty of individual augmented graph realizations.
  • B Information-Theoretic Interpretation of GATTA Aggregation: Because graph perturbations may introduce semantic drift, augmentation disagreement can combine meaningful prediction instability with non-label-preserving noise and cause GATTA-P to overestimate uncertainty.Consistency filtering mitigates this effect, while GATTA-S is less sensitive to disagreement between otherwise confident augmented-view predictions.
  • C Filtering variants: Hard filtering enforces strict prediction consistency with binary masks, while Firm filtering discards inconsistent views and confidence-weights the consistent ones.Soft weights alone were ineffective because spuriously confident wrong-class predictions introduced noise.
  • C Filtering variants: Firm filtering did not outperform hard filtering, suggesting that uniform weighting of consistent views is sufficient; its results are reported only for completeness.The comparison was conducted in the reported experiments and summarized in Table 6.

D Impact of Data Augmentation

The augmentation-strength analysis evaluates Feature Noising and Edge Drop across CoraML and PubMed with Entropy and Least Confidence, reporting gains against matched non-TTA baselines. Stronger augmentations generally improve performance, while filtering benefits depend on the aggregation strategy across SGC and GCN.

  • Experimental setup: The experiments analyze Feature Noising and Edge Drop across CoraML and PubMed using Entropy and Least Confidence acquisition strategies.Figures 11–14 report accuracy gains relative to the corresponding dataset-model-acquisition configuration without test-time augmentation.
  • Augmentation strength: Stronger augmentations generally yield better performance across the evaluated settings.
  • Filtering strategy: Consistency-based filtering is essential for GATTA-P, particularly at higher augmentation strengths, whereas GATTA-S performs optimally without filtering.These filtering trends are consistent across both SGC and GCN architectures.

E Learning Dynamics and Performance Comparison · F Batch Active Learning and Extended Labeling Budgets · G Computational complexity

GATTA most benefits simple uncertainty-based active-learning strategies, improving accuracy and reducing variance, while gains weaken with larger acquisition batches and scarce-label effects diminish over time. Computationally, GATTA-P is preferable for complex acquisition functions, whereas GATTA-S is viable for simple ones.

  • E Learning Dynamics and Performance Comparison: GATTA variants substantially improve simple strategies, closing the gap with GEEM and sometimes surpassing it on Citeseer, CoraML, and PubMed.Benefits are consistent across GCN and SGC, with GATTA-S often stronger in later iterations, particularly on citation networks.
  • E Learning Dynamics and Performance Comparison: Dataset properties strongly affect GATTA: Amazon Computers gains at least +5% with GATTA-S, whereas Citeseer and Amazon Photos show more modest improvements.The passage attributes this variation to graph characteristics such as feature informativeness and homophily.
  • E Learning Dynamics and Performance Comparison: GATTA provides modest or negative gains for complex strategies, with GEEM showing consistent slight degradation after augmentation-based refinement.Sophisticated acquisition functions may already account for prediction uncertainty, while structure-aware acquisition can interfere with input-level perturbations.
  • E Learning Dynamics and Performance Comparison: 18 of 20 simple-method dataset-architecture combinations show reduced standard deviation with GATTA-S, except PubMed with SGC.Aggregating predictions across augmented views stabilizes node selection and improves outcome consistency across initializations.
  • F Batch Active Learning and Extended Labeling Budgets: Larger acquisition batches consistently reduce performance and diminish GATTA’s benefit because decisions depend on more of the uncertainty ranking.Sequential or small-batch acquisition gains more from refined uncertainty, while larger batches may contain increasingly similar uncertain nodes.
  • F Batch Active Learning and Extended Labeling Budgets: The GATTA-baseline performance gap is largest early in active learning and gradually narrows as more labels are acquired.Uncertainty estimation is most valuable when labeled data are scarce, whereas acquisition strategies converge as the labeling budget increases.
  • G Computational complexity: GATTA-P and GATTA-S have similar runtime scaling for simple strategies, but GATTA-S becomes substantially more expensive with complex strategies such as AGE.GATTA-P is therefore preferred for complex acquisition functions, while GATTA-S suits simple strategies where repeated evaluations add limited overhead and may provide marginal benefits.
  • G Computational complexity: GATTA-P has complexity O((N × I) + Q), while GATTA-S has complexity O(N × (I + Q)) because it reevaluates the acquisition function for each augmented view.When acquisition cost dominates, GATTA-S incurs multiplicative overhead; at 10,000 augmentations with AGE on PubMed, it requires over 5,000 seconds versus approximately 50 seconds for GATTA-P.

H Confidence Analysis

GATTA produces more concentrated confidence distributions than the baseline, while consistency filtering further removes low-confidence predictions and shifts the distribution upward. These changes indicate more reliable uncertainty estimates for active-learning node selection.

  • Confidence distributions: GATTA averaged over 500 augmented views concentrates prediction confidences in the mid-range region (0.4-0.6), unlike the broadly dispersed baseline.The concentration suggests that test-time augmentation reveals underlying prediction uncertainty.
  • Confidence distributions: Mean confidence remains above the most concentrated region for both baseline and GATTA, indicating distributions skewed toward higher confidences.This pattern is consistent with uncertainty-based active learning, which queries uncertain nodes and leaves more high-confidence predictions unlabeled.
  • Consistency filtering: Consistency-based filtering reduces the density of low-confidence predictions (<0.4) and shifts GATTA confidence distributions upward.Filtering discards augmented views with inconsistent predictions, retaining more stable, higher-confidence predictions and producing a more concentrated distribution.
  • Active-learning implications: The concentration and elevation effects suggest that GATTA provides more reliable uncertainty estimates, potentially improving node selection and label efficiency.These distributional changes directly affect active-learning decisions by making confidence estimates more dependable.
Loading 2608.15084v1…