[Paper Review] Data-centric Graph Learning: A Survey

Yuxin Guo, Deyu Bo+6 more|arXiv (Cornell University)|Oct 8, 2023

Advanced Graph Neural Networks4 citations

TL;DR

This survey introduces a data-centric framework for graph learning, categorizing methods by learning stage (pre-processing, training, inference) and data structure (topology, features, labels). It proposes a novel taxonomy to guide when and how to modify graph data to enhance model performance, while addressing issues like bias, noise, and heterophily through data-centric solutions.

ABSTRACT

The history of artificial intelligence (AI) has witnessed the significant impact of high-quality data on various deep learning models, such as ImageNet for AlexNet and ResNet. Recently, instead of designing more complex neural architectures as model-centric approaches, the attention of AI community has shifted to data-centric ones, which focuses on better processing data to strengthen the ability of neural models. Graph learning, which operates on ubiquitous topological data, also plays an important role in the era of deep learning. In this survey, we comprehensively review graph learning approaches from the data-centric perspective, and aim to answer three crucial questions: (1) when to modify graph data, (2) what part of the graph data needs modification to unlock the potential of various graph models, and (3) how to safeguard graph models from problematic data influence. Accordingly, we propose a novel taxonomy based on the stages in the graph learning pipeline, and highlight the processing methods for different data structures in the graph data, i.e., topology, feature and label. Furthermore, we analyze some potential problems embedded in graph data and discuss how to solve them in a data-centric manner. Finally, we provide some promising future directions for data-centric graph learning.

Motivation & Objective

• To address the growing need for data-centric approaches in graph learning, shifting focus from model architecture to data quality and processing.
• To identify when in the graph learning pipeline (pre-processing, training, inference) data modifications should occur for optimal model performance.
• To determine which components of graph data—topology, features, or labels—should be modified to unlock the full potential of graph neural networks.
• To mitigate the impact of problematic data, such as noisy, biased, or heterophilic graphs, through data-centric strategies.
• To propose future directions for co-development of graph data and models, emphasizing standardization, continuous learning, and collaborative design.

Proposed method

• Proposes a novel taxonomy based on the stages of the graph learning pipeline: pre-processing, training, and inference, to guide data modification decisions.
• Categorizes data-centric methods by data structure—topology, features, and labels—offering targeted processing strategies for each.
• Introduces data augmentation, graph structure learning, and feature/label refinement as core techniques to improve data quality and model robustness.
• Applies unsupervised and self-supervised methods to detect and mitigate selection bias, unfairness, and spurious correlations in graph data.
• Leverages model predictions to iteratively refine graph data, enabling continuous learning and data evolution.
• Proposes co-development of graph data and models through integrated design, exemplified by frameworks like GraphStorm.

Experimental results

Research questions

• RQ1 When in the graph learning pipeline should graph data be modified to maximize model performance?
• RQ2 Which components of graph data—topology, features, or labels—most significantly influence the representational capacity of graph models?
• RQ3 How can data-centric methods safeguard graph models against noise, bias, and heterophily in real-world graphs?
• RQ4 What role can continuous learning and model feedback play in refining graph data over time?
• RQ5 How can graph data and models be co-developed to achieve mutual optimization in a data-centric paradigm?

Key findings

• Data-centric approaches significantly improve graph neural network performance by focusing on data quality rather than model complexity.
• Modifying graph topology through structure learning or augmentation enhances model generalization, especially in heterophilic graphs.
• Feature and label refinement techniques reduce the impact of noisy or biased data, improving model fairness and robustness.
• Unsupervised and self-supervised methods can detect and mitigate selection bias and spurious correlations in node representations.
• Continuous learning frameworks that use model predictions to refine graph data enable adaptive and evolving data pipelines.
• Co-development of data and models, as demonstrated by GraphStorm, leads to more efficient and effective graph learning systems.