PaperlayerView exact revision on arXiv

Source-linked AI summary

Graph Data Augmentation for Graph Machine Learning: A Survey

Tong Zhao, Wei Jin, Yozen Liu, Yingheng Wang, Gang Liu, Stephan Günnemann, Neil Shah, Meng Jiang

arXiv:2202.08871v2cs.LG

TL;DR

Graph data augmentation is less straightforward than conventional augmentation because graph structure is irregular and non-Euclidean, amid graph-learning challenges such as noisy topology and incomplete attributes. This survey organizes existing methods through taxonomies and reviews their methodologies, applications, and challenges. It concludes with a structured account of the field and directions for further research.

  • Problem

    Graph data augmentation remains relatively under-explored because graph structure makes traditional augmentation operations difficult to transfer directly, while graph learning faces data-quality challenges.

    Source: §1 · p. 2

  • Method

    The survey categorizes graph data augmentation by operated data, downstream task, and learnability, then reviews methods, applications, and open challenges.

    Source: §1 · p. 2, §8 · p. 19

  • Results

    The survey presents a structured overview of existing graph data augmentation techniques, their methodologies, applications in self-supervised learning, and current challenges.

    Source: §8 · p. 19

  • Takeaways & Limitations

    The survey aims to guide GML researchers and practitioners in studying and using graph data augmentation and to inspire further work.

    Source: §1 · p. 2, §8 · p. 19

    Existing automated graph augmentation methods mainly target specific datasets and tasks, while transferability across datasets is a desired characteristic.

    Source: §7.1 · p. 18

Abstract

from arXiv · show

Data augmentation has recently seen increased interest in graph machine learning given its demonstrated ability to improve model performance and generalization by added training data. Despite this recent surge, the area is still relatively under-explored, due to the challenges brought by complex, non-Euclidean structure of graph data, which limits the direct analogizing of traditional augmentation operations on other types of image, video or text data. Our work aims to give a necessary and timely overview of existing graph data augmentation methods; notably, we present a comprehensive and systematic survey of graph data augmentation approaches, summarizing the literature in a structured manner. We first introduce three different taxonomies for categorizing graph data augmentation methods from the data, task, and learning perspectives, respectively. Next, we introduce recent advances in graph data augmentation, differentiated by their methodologies and applications. We conclude by outlining currently unsolved challenges and directions for future research. Overall, our work aims to clarify the landscape of existing literature in graph data augmentation and motivates additional work in this area, providing a helpful resource for researchers and practitioners in the broader graph machine learning domain. Additionally, we provide a continuously updated reading list at https://github.com/zhao-tong/graph-data-augmentation-papers.

1 Introduction

Graph machine learning depends on data quality, but graph structure makes it difficult to transfer conventional augmentation operations directly. This survey organizes graph data augmentation methods, reviews representative techniques and applications, and identifies challenges and future directions.

1 Introduction

  • Graph data augmentation addresses graph-learning challenges including sparse structure, noisy topology, incomplete attributes, adversarial attacks, scarce labels, and over-smoothing.These challenges arise in part from graph-data dependencies and GNN message passing.Source: §1 · pp. 1–2
  • Graph data’s irregular, non-Euclidean structure makes direct transfer of image and text augmentation methods non-trivial beyond basic masking, dropping, or cropping.The survey describes graph-specific approaches including structure learning, adversarial training, and rationalization.Source: §1 · p. 2
  • The survey presents three complementary taxonomies based on operated data, downstream task, and whether augmentation involves learning.It also reviews representative methods and applications and outlines challenges and future research directions.Source: §1 · p. 2
  • The paper aims to orient GML researchers and practitioners to the growing area and provide a guide for studying and using graph data augmentation.Its planned coverage includes background, a bird’s-eye view of methods, representative techniques and applications, and research challenges.Source: §1 · p. 2

2 Preliminaries

The preliminaries define graph and GNN notation, describe graph data augmentation as graph transformation, and motivate it through mismatch between observed graphs and learning objectives. They also explain why graph augmentation operations do not transfer straightforwardly from Euclidean data.

2.1 Notations

  • A graph is represented by nodes V, edges E, adjacency matrix A, node features X, and task-dependent labels y; tildes mark GDA-generated data.For example, ˜A and ˜x_i denote augmented adjacency and node features.Source: §2.1 · p. 2

2.2 Graph Neural Networks

  • Message-passing GNN layers aggregate neighboring node embeddings and update each node representation using its previous embedding and the aggregate.The survey notes that GNNs can be implemented with sparse matrix multiplications and uses a GCN as an example.Source: §2.2 · pp. 3–4

2.3 Graph Data Augmentation

  • Graph data augmentation transforms a graph into augmented graph data, potentially changing adjacency, node features, edge features, or labels.Graph tasks are inter-connected and non-i.i.d., so methods typically modify the graph rather than an isolated node or edge.Source: §2.3 · p. 4

2.4 Motivation: Why Augment Graphs?

  • Observed graphs may diverge from intended relationships through noise, partial observation, measurement errors, or human errors, and may not suit a downstream task even when accurately observed.Augmentation can denoise, impute, or enhance graph structure toward an intended modeling process or task objective; unclear label-preserving graph transforms motivate further study.Source: §2.4 · pp. 4–5

3 Taxonomies

The survey offers orthogonal taxonomies for classifying graph data augmentation by manipulated data, downstream task, and learnability. It uses the rule-based versus learned distinction to organize its later discussion of methods.

3 Taxonomies

  • The survey’s three orthogonal taxonomies classify methods by manipulated data, downstream task, and whether augmentation involves learning.It focuses on the rule-based versus learned taxonomy in later sections.Source: §3 · p. 5

3.1 Operated Data Modality

  • Data-modality categories are structure augmentation, which changes connectivity; feature augmentation, which changes node features; and label augmentation, which changes labels.Some methods combine structure and feature changes, while Mixup and counterfactual methods can generate new labels.Source: §3.1 · p. 5

3.2 Downstream Tasks

  • Task-based categories are node-level, edge-level, and graph-level, but some methods are task-specific while generic methods such as DropEdge can apply broadly.For example, CFLP generates counterfactual links for link prediction, whereas molecular substructure methods may not suit large graph datasets.Source: §3.2 · p. 5

3.3 Rule-based vs. Learned Augmentations

  • Rule-based methods apply predefined rules without learnable augmentation parameters, whereas learned methods use learnable parameters to generate augmented examples.Learned modules may train independently or end-to-end; examples include graph structure learning, rationalization, and automated augmentation.Source: §3.3 · pp. 5–6

4 Rule-based Approaches for GDA

Rule-based GDA is presented as a commonly used, simple, and efficient family organized into data removal, addition, and manipulation.

4.1 Data Removal

  • DropEdge resamples edges each epoch; the survey reports improved generalization, gains on deeper GNNs, and mitigation of over-smoothing.It replaces the original graph structure for message passing during each training epoch.Source: §4.1 · p. 6
  • Removal methods also drop nodes or features, shuffle node features, mask aggregated messages, or crop subgraphs while aiming to preserve graph semantics or topology.Node dropping can remove a node’s features or the node and its incident edges; cropping methods include random-walk sampling and diffusion-based selection.Source: §4.1 · pp. 6–8

4.2 Data Addition

  • For graph-level Mixup, Graph Mixup mixes features and hidden states without rebuilding neighborhoods, while G-Mixup interpolates class graphons to generate synthetic graphs.ifMixup instead mixes graph data using index-based node matching, and Graph Transplant uses selected substructures as mixing units.Source: §4.2 · pp. 8–9
  • Graph addition also covers minority-class synthetic nodes, counterfactual links or examples, generated node attributes, and propagated labels for unlabeled nodes.GraphSMOTE generates edges for synthetic nodes; label propagation iteratively spreads labels along graph edges.Source: §4.2 · p. 9

4.3 Data Manipulation

  • Diffusion creates a weighted adjacency that can expose multi-hop information to GNNs, but its often-dense output is commonly sparsified.The diffused adjacency can also serve as an augmented view for self-supervised objectives such as contrastive learning.Source: §4.3 · p. 10
  • Substructure-based manipulation uses motifs for edge additions or removals and replaces molecular substructures such as functional groups.M-Evolve selects motifs using Resource Allocation-based sampling weights; MoCL uses biomedical domain knowledge.Source: §4.3 · p. 10

5 Learned Approaches for GDA

Learned graph data augmentation (GDA) methods learn augmentation strategies from data, addressing the limits of rule-based operations that may not use downstream-task information. The survey organizes these methods into graph structure learning, adversarial training, rationalization, and automated augmentation.

5 Learned Approaches for GDA

  • Learned GDA methods address rule-based approaches’ potential failure to use downstream-task information by learning augmentation strategies in a data-driven manner.The survey groups them into structure learning, adversarial training, rationalization, and automated augmentation.Source: §5 · p. 10

5.1 Graph Structure Learning

  • Graph structure learning refines graph structure alongside model parameters; its generalization methods learn either continuous or discrete adjacency matrices.Continuous methods avoid requiring discrete structures during optimization, while other methods sample structures from targeted distributions or edit edges.Source: §5.1 · pp. 10–11
  • GLCN learns a sparse, continuous graph using feature similarities and a loss balancing smoothness, sparsity, and closeness to the original graph.TO-GCN adds label similarity; IDGL jointly learns a similarity metric and prediction model; SLAPS adds feature-denoising self-supervision.Source: §5.1 · p. 11
  • Learned discrete structures include GAug [131] and LDS, while other methods prune or add edges according to task relevance, graph spectra, or predictions.GAug variants differ in whether edge manipulation is feasible at inference; PTDNet, NeuralSparse, TADropEdge, and AdaEdge exemplify edge editing.Source: §5.1 · pp. 11–12
  • Test-time augmentation such as GTrans [49] optimizes a self-supervised loss on the test graph and has been shown to improve GNN generalization on out-of-distribution data.Source: §5.1 · p. 12
  • For robustness, ProGNN regularizes learned structure to preserve graph properties, but its robustness can decline under adaptive attacks.ProGNN has O(N^3) time and O(N^2) space complexity; LRGNN decouples low-rank and sparse components to speed structure learning.Source: §5.1 · pp. 12–13

5.2 Graph Adversarial Training

  • Adversarial training augments training graphs with adversarial perturbations to node features or structure, aiming for models that tolerate perturbations at test time.Methods include edge perturbations, such as PGD-based structure changes, and gradient-based feature perturbations such as FLAG.Source: §5.2 · pp. 13–14

5.3 Rationalization

  • Rationalization learns representative subgraphs that support predictions; methods use them for interpretation and prediction, while environment replacement creates virtual augmented examples.GIB [121] learns an informative compressed subgraph, and GSAT selects task-relevant subgraphs; causal-rationale methods also target out-of-distribution generalization.Source: §5.3 · p. 14

6 GDA for Self-supervised Learning

The survey describes graph data augmentation in self-supervised learning, where views of graph data support representation learning. It reviews contrastive, non-contrastive, and consistency-based schemes, along with the augmentations they use.

6 GDA for Self-supervised Learning

  • Self-supervised graph learning commonly uses rule-based augmentations to corrupt or perturb graphs, unlike learned GDA methods that seek task-relevant information.The section covers contrastive, non-contrastive, and consistency-training schemes.Source: §6 · p. 16

6.1 Contrastive Learning

  • Contrastive graph learning treats augmented views of an example as positive samples and other examples in the batch as negatives; common augmentations include edge dropping and feature masking.Methods also use node dropping, edge perturbation, attribute masking, subgraph sampling, and graph diffusion; stochastic operations may fail to induce useful task-relevant invariance.Source: §6.1 · pp. 16–17

6.2 Non-contrastive Learning

  • Non-contrastive methods avoid negative examples, comparing views of the same sample and using mechanisms such as prediction heads or stop-gradient to avoid collapse.CCA-SSG, GBT, BGRL, and T-BGRL [84] use random edge dropping and node-feature masking.Source: §6.2 · p. 17

6.3 Consistency Training

  • Consistency training compares representation distributions across views and is commonly combined with supervised loss in semi-supervised settings.Examples include NodeAug, GRAND, NASA, and MV-GCN, which use different graph augmentations and consistency regularization.Source: §6.3 · p. 17

7 Challenges and Directions

The survey identifies open needs in transferable and domain-aware augmentation, scalable implementation, broader evaluation, and theoretical explanation.

7.1 Domain Adaptation and Regularization

  • Automated GDA methods mainly target specific datasets and tasks; transferable methods usable across datasets, ideally cross-domain or OOD, remain missing.The survey also calls for domain-based regularization because common augmentations can change molecule semantics; only MoCL is noted as considering local-substructure semantics.Source: §7.1 · p. 18
  • Molecule augmentations can change a graph’s underlying semantics, while domain-based regularization remains under-explored beyond MoCL’s use of local-substructure semantics.Dropping a carbon atom from aspirin’s phenyl ring changes it into an alkene chain, a different compound.Source: §7.1 · p. 18

7.2 Scalability for Large-Scale Graphs

  • Global operations can impede scale: GAug-M selects top K from O(N^2) node-pair logits, while GAug-O’s full adjacency backpropagation creates massive memory overhead.TADropEdge also requires precomputing an edge score before GNN training [29].Source: §7.2 · p. 18

7.3 Comprehensive Evaluation Criteria and Standards

  • Evaluation often focuses on prediction performance on specific datasets, leaving time and resource costs, transferability, and scalability less examined.The survey calls for benchmarks spanning datasets of different scales and domains to assess transferability–scalability tradeoffs.Source: §7.3 · pp. 18–19

7.4 Theoretical Foundation

  • Rigorous understanding of how GDA improves generalization and alleviates over-smoothing remains limited, especially for (semi-)supervised learning.The survey notes limited proofs or theoretical bounds for these relationships.Source: §7.4 · p. 19
  • Potential theoretical perspectives include certified robustness, topology bottlenecks and over-squashing, and causal analysis of counterfactual augmentation.Theoretical analyses of image augmentation cannot be directly applied to graph data’s irregular structure.Source: §7.4 · p. 19

8 Conclusion

The survey organizes graph data augmentation through three taxonomies and reviews approaches, applications, challenges, and future directions.

8 Conclusion

  • The survey categorizes GDA techniques using three taxonomies, reviews methods and self-supervised learning applications, and outlines challenges and future research directions.It aims to guide GML researchers and practitioners and encourage further work on GDA.Source: §8 · p. 19
Loading 2202.08871v2…