PaperlayerView exact revision on arXiv

Source-linked AI summary

Data Augmentation for Deep Graph Learning: A Survey

Kaize Ding, Zhe Xu, Hanghang Tong, Huan Liu

arXiv:2202.08235v3cs.LG

TL;DR

Deep graph learning needs augmentation methods suited to noisy, scarce, non-Euclidean, and multimodal graph data. The survey formalizes GraphDA, proposes a modality-based taxonomy, and reviews its use in reliable and low-resource graph learning, while identifying semantic-preservation, heterophily, and automation as open challenges.

  • Problem

    Graph data are non-Euclidean and multimodal, while deep graph learning depends on input quality and labeled data that may be noisy, scarce, or incomplete.

  • Method

    The survey formally formulates GraphDA, categorizes methods by feature-wise, structure-wise, and label-wise augmentation, and reviews applications across two data-centric learning problems.

  • Results

    The survey synthesizes representative GraphDA techniques and their applications in reliable graph learning and low-resource graph learning.

  • Takeaways & Limitations

    GraphDA offers a structured design space for augmenting graph information and supports research on data-centric deep graph learning.

  • Takeaways & Limitations

    Existing GraphDA methods remain under-explored for heterophilic graphs, automated augmentation selection, and semantic-preserving transformations.

Abstract

from arXiv · show

Graph neural networks, a powerful deep learning tool to model graph-structured data, have demonstrated remarkable performance on numerous graph learning tasks. To address the data noise and data scarcity issues in deep graph learning, the research on graph data augmentation has intensified lately. However, conventional data augmentation methods can hardly handle graph-structured data which is defined in non-Euclidean space with multi-modality. In this survey, we formally formulate the problem of graph data augmentation and further review the representative techniques and their applications in different deep graph learning problems. Specifically, we first propose a taxonomy for graph data augmentation techniques and then provide a structured review by categorizing the related work based on the augmented information modalities. Moreover, we summarize the applications of graph data augmentation in two representative problems in data-centric deep graph learning: (1) reliable graph learning which focuses on enhancing the utility of input graph as well as the model capacity via graph data augmentation; and (2) low-resource graph learning which targets on enlarging the labeled training data scale through graph data augmentation. For each problem, we also provide a hierarchical problem taxonomy and review the existing literature related to graph data augmentation. Finally, we point out promising research directions and the challenges in future research.

1. INTRODUCTION

Deep graph learning performs well on graph-structured tasks but depends heavily on high-quality graphs and labels, which are often noisy, scarce, or incomplete. This survey formalizes GraphDA, organizes its techniques and applications, and identifies future challenges.

  • GNN effectiveness in deep graph learning depends substantially on high-quality input graphs and ground-truth labels.
  • Real-world graph learning is challenged by the time-consuming, labor-intensive, and incomplete nature of graph labeling.
  • The survey provides a formal formulation and comprehensive taxonomy of GraphDA based on feature-oriented, structure-oriented, and label-oriented augmentation modalities.
  • It systematically reviews GraphDA applications in low-resource graph learning and reliable graph learning, alongside related learning paradigms.
  • Compared with earlier surveys focused mainly on individual deep graph learning problems, this work reviews GraphDA techniques and applications across two representative data-centric problems.

2. PRELIMINARIES

The preliminaries define attributed graphs and their matrix notation, describe GNNs through message passing, and classify deep graph learning tasks by their target graph components. The task categories cover node-level, edge-level, and graph-level prediction.

  • 2.1 Notations and Definitions: An attributed graph is represented as G = (A, X), where A is the adjacency matrix and X is the node feature matrix.The notation uses n for nodes, m for edges, and d for feature dimension.
  • 2.2 Graph Neural Networks: GNNs extend neural networks to graph data and commonly extract component representations through recursive message passing.Node neighborhoods are represented through the adjacency matrix, and initial node representations are node features.
  • 2.3 Deep Graph Learning Tasks: Node-level tasks map graphs to node properties, with semi-supervised node classification predicting labels for partially unlabeled nodes.A typical implementation combines a GNN node encoder with a multi-class classifier.
  • 2.3 Deep Graph Learning Tasks: Edge-level tasks predict edge properties, such as whether a specified pair of nodes is connected in link prediction.A common implementation classifies edge embeddings produced from node representations.
  • 2.3 Deep Graph Learning Tasks: Graph-level tasks treat each graph as a sample and infer graph properties from graph representations formed by aggregating node embeddings.Graph classification is presented as a representative example.

3. TECHNIQUES OF GRAPH DATA AUGMENTATION

GraphDA transforms graphs to enrich preserved information, with learnable methods trained through decoupled, joint, or bi-level optimization. Its techniques augment features, structure, or labels using operations such as perturbation, sampling, diffusion, generation, and node manipulation.

  • GraphDA formulation: GraphDA learns or applies a transformation fθ to generate augmented graphs that enrich or enhance information preserved from input graphs.The framework jointly considers augmentation loss Laug and downstream utility loss Lutility.
  • Training paradigms: Learnable GraphDA uses decoupled training, joint training, or bi-level optimization to coordinate augmentation and prediction objectives.Decoupled training uses two stages; joint training optimizes both models together; bi-level optimization alternates updates.
  • Taxonomy: GraphDA techniques are categorized by augmented modality into feature-oriented, structure-oriented, and label-oriented methods.The taxonomy also organizes applications across data-centric deep graph learning problems.
  • Structure-oriented augmentations: Structure-oriented augmentation modifies graph connectivity through edge perturbation, rewiring, diffusion, sampling, node dropping, node insertion, or graph generation.These operations can preserve task-relevant information, improve graph utility, alter global topology, or enlarge graph-level training scales.
  • Structure-oriented augmentations: Graph sampling selects subgraphs while retaining portions of nodes and their underlying linkages, whereas node dropping removes nodes together with associated edges and, for attributed graphs, features.Node insertion adds virtual nodes and connecting edges, requiring feature initialization for attributed graphs.
  • Structure-oriented augmentations: Graph generation learns distributions from observed graphs to create additional training graphs, including graph coarsening, condensation, and edge mixing.It is commonly used to increase training-graph scale for graph-level tasks such as graph classification.

4. GRAPH DATA AUGMENTATION FOR LOW-RESOURCE GRAPH LEARNING

GraphDA addresses low-resource graph learning by enlarging useful training signals when labels are scarce. The survey covers generative and contrastive self-supervision alongside consistency training, self-training, interpolation, and imbalance-oriented augmentation.

  • Overview: GraphDA is investigated for low-resource graph learning because ground-truth labels are scarce, costly to obtain, and can leave models vulnerable to overfitting.The section covers graph self-supervised and semi-supervised learning.
  • Graph Self-Supervised Learning: Graph generative modeling augments input graphs through edge perturbation, feature masking, or node dropping, then reconstructs masked features or structure.These methods follow a graph autoencoder-style workflow for learning node representations.
  • Graph Self-Supervised Learning: Graph contrastive learning creates augmented views, treating views from the same original sample as positive pairs and views from different samples as negative pairs.Contrastive loss pulls positive representations together and pushes negative representations apart in latent space.
  • Graph Self-Supervised Learning: Learnable augmentation schemes adapt edge dropping and feature masking to provide diverse node contexts, including adversarial self-supervised learning without handcrafted augmentation.Some contrastive methods also combine multiple predefined augmentation strategies.
  • Graph Semi-Supervised Learning: Consistency training enforces agreement between predictions or representations obtained from different stochastic augmentations of the same input.NodeAug uses feature corruption and edge perturbation while minimizing KL-divergence between original and augmented representations.
  • Graph Semi-Supervised Learning: Semi-supervised GraphDA expands limited supervision through pseudo-labeling, feature-and-label interpolation, and minority-class augmentation.GraphMix combines Manifold Mixup and pseudo-labeling, while imbalance-oriented methods address power-law class distributions such as Pubmed’s 5.25% minority class.

5. GRAPH DATA AUGMENTATION FOR RELIABLE GRAPH LEARNING

GraphDA supports reliable graph learning by improving graph quality, model expressiveness, robustness, and scalability through augmentation of graph structure, features, and connectivity.

  • Overview: Reliable graph learning is framed as improving robustness, expressiveness, and scalability by augmenting input graphs.Table 3 summarizes representative GraphDA works and distinguishes decoupled training, joint training, and bi-level optimization.
  • Graph Structure Learning: Graph structure learning rewires suboptimal graphs caused by fake connections, over-personalized users, and construction heuristics.
  • Graph Feature Denoising: Feature denoising rewrites node features to tolerate abnormal inputs and propagate information when features are missing.
  • Graph Adversarial Defense: GraphDA defends against adversarial attacks by recovering or enhancing poisoned graph structure and features.Examples use low-rank reconstruction, similarity-based link pruning, and topology or feature regularization.
  • Boosting GNN Expressive Power: Distance, random, and structural-count features augment node representations to exceed the expressive limitations of 1-WL-based GNNs.Virtual-node insertion provides another structure-oriented route for improving graph representations.
  • Scalable and Robust Training: Graph rewiring methods address over-smoothing, while sampling, diffusion, generation, and partitioning improve GNN scalability.GraphSAGE, FastGCN, LADIES, GraphSAINT, and Cluster-GCN reduce neighborhood-search, sampling, or memory costs.

6. FUTURE DIRECTIONS

Future GraphDA research must extend beyond simple graph types, automate augmentation selection, preserve graph semantics, and establish stronger foundations for trustworthy DGL.

  • Data Augmentation beyond Simple Graphs: Augmentation for heterophilic, heterogeneous, hyper-, multiplex, and dynamic graphs remains largely unexplored.Existing work often targets homophilic, plain, or attributed graphs and may increase assortativity or remove disassortative edges.
  • Automated and Generalizable Augmentation: Automated GraphDA is needed because augmentation effects vary across datasets and labels may be unavailable for representation evaluation.The survey also identifies generalizable methods that adapt across domains without learning from scratch for each graph.
  • Semantic-Preserving Graph Data Augmentation: Arbitrary graph perturbations can change structural and semantic patterns, motivating label-consistent or semantic-preserving augmentation.Latent-space augmentation is cited as one way to avoid semantic perturbation.
  • Graph Data Augmentation for Trustworthy DGL: Trustworthy DGL requires GraphDA research on fairness, causality, explainability, regulatory compliance, and the theoretical reasons augmentation works.The survey characterizes existing GraphDA theory as limited and largely surface-level.

7. CONCLUSION

The survey formalizes GraphDA, organizes it by augmentation modality, reviews its applications to low-resource and reliable graph learning, and identifies future challenges and opportunities.

  • The survey presents a formal formulation and taxonomy of GraphDA based on feature-wise, structure-wise, and label-wise augmentation.
  • It reviews GraphDA applications to low-resource graph learning and reliable graph learning, including prevailing GraphDA-based algorithms.
  • It outlines current challenges and opportunities for future GraphDA research.
Loading 2202.08235v3…