Graph Data Augmentation for Graph Machine Learning: A Survey (2202.08871v2)

Abstract: 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.

An Analysis of Graph Data Augmentation for Graph Machine Learning

The paper "Graph Data Augmentation for Graph Machine Learning: A Survey" authored by T. Zhao et al., offers a comprehensive survey on graph data augmentation (GDA) techniques. This document serves as a systematic assessment, exploring various methods and taxonomies relevant to improving model performance in graph machine learning (GML) through data augmentation strategies.

Graph machine learning poses unique challenges due to the intrinsic non-Euclidean structure of graph data. Traditional data augmentation (DA) techniques common in other domains like computer vision and natural language processing fail to directly translate due to these complexities. In response to this, the paper presents an organized exploration of GDA methodologies, seeking to academically enrich and technically facilitate advancements in GML.

Overview

The authors categorize existing GDA methods via three taxonomies: the data manipulated, the focus of downstream tasks, and the learning strategy involved. This systematic categorization not only aids in understanding existing GDA approaches but also serves as a framework to evaluate future developments in this domain.

Key Insights and Techniques

1. Data Perspective:

• Structure Augmentations: These modifications involve altering the connectivity patterns within graphs, such as through edge manipulation. Techniques such as DropEdge revel in dropping specific edge sets to preserve structural diversity.
• Feature Augmentations: Altering the feature vectors of nodes enhances the quality of graph datasets. Methods like Attribute Masking introduce noise into graph node features deliberately to improve model generalization.
• Label Augmentations: They consist of modifications to node or graph labels and focus on techniques like Mixup, where interpolation is used over node features and labels to introduce new training samples.

2. Learning Perspective:

• Rule-based Augmentations: Including techniques such as random edge dropping and node removal, these are computationally beneficial and aid in data diversity by removing specific structural elements within the graph.
• Learned Augmentations: Learning-driven methods like graph structure learning intelligently modify graph structure based on downstream requirements, enhancing robustness against adversarial attacks by learning task-specific graph features.

3. Self-supervised Learning Applications:

• Contrastive Learning: GDA enables these models to create multiple views of graph data, fostering enhancement in feature extraction processes. Techniques like GRACE leverage graph data augmentation to perform contrastive learning at node levels.
• Non-contrastive Learning: Uses GDA strategies to improve representation learning without relying on contrastive methods, often employing simple data corruptions.
• Consistency Training: Ensures the stability of GNN predictions against stochastic augmentations of graph data, reinforcing node classification tasks.

Implications and Future Directions

Graph data augmentation has strong implications for the robustness, scalability, and generalization of GNNs. Future developments in GDA may explore more adaptive automated strategies, enhancing domain adaptation and supporting scalability for larger datasets. Furthermore, a comprehensive evaluation framework incorporating both quantitative and practical evaluations would offer a more holistic view of GDA's efficacy in GML.

The theoretical underpinnings of GDA require further exploration, particularly with rigorous analysis into effects on model robustness and generalization. Understanding the theoretical implications will allow for the design of augmentation strategies more aligned with improving robust GML outcomes.

Conclusion

This survey by Zhao et al. highlights the significant contributions and potential of graph data augmentation in enhancing graph neural network methodologies. As GML applications continue to expand across domains, the exploration and development of sophisticated, theoretically grounded GDA techniques will be pivotal in addressing both current and emergent challenges within graph machine learning landscapes. This work, therefore, serves as an essential reference point for further research and application development in graph data augmentation.