[论文解读] Out-Of-Distribution Generalization on Graphs: A Survey

Haoyang Li, Xin Wang+2 more|arXiv (Cornell University)|Feb 16, 2022

Advanced Graph Neural Networks被引用 26

一句话总结

本综述将 Graph OOD 泛化形式化，将现有方法分为数据、模型和学习策略三个分支，回顾理论与数据集，并讨论未来方向。

ABSTRACT

Graph machine learning has been extensively studied in both academia and industry. Although booming with a vast number of emerging methods and techniques, most of the literature is built on the in-distribution hypothesis, i.e., testing and training graph data are identically distributed. However, this in-distribution hypothesis can hardly be satisfied in many real-world graph scenarios where the model performance substantially degrades when there exist distribution shifts between testing and training graph data. To solve this critical problem, out-of-distribution (OOD) generalization on graphs, which goes beyond the in-distribution hypothesis, has made great progress and attracted ever-increasing attention from the research community. In this paper, we comprehensively survey OOD generalization on graphs and present a detailed review of recent advances in this area. First, we provide a formal problem definition of OOD generalization on graphs. Second, we categorize existing methods into three classes from conceptually different perspectives, i.e., data, model, and learning strategy, based on their positions in the graph machine learning pipeline, followed by detailed discussions for each category. We also review the theories related to OOD generalization on graphs and introduce the commonly used graph datasets for thorough evaluations. Finally, we share our insights on future research directions. This paper is the first systematic and comprehensive review of OOD generalization on graphs, to the best of our knowledge.

研究动机与目标

• Formalize the Graph OOD generalization problem and its motivation.
• Categorize existing Graph OOD methods into data, model, and learning-strategy classes.
• Review theoretical foundations related to Graph OOD generalization.
• Summarize commonly used datasets and evaluation practices for Graph OOD.
• Provide insights and directions for future research in Graph OOD generalization.

提出的方法

• Define Graph OOD generalization and highlight distribution-shift challenges in graphs.
• Propose a three-branch taxonomy (Data, Model, Learning Strategy) aligned with the graph ML pipeline.
• Detail data augmentation strategies (structure-wise, feature-wise, mixed-type) to improve OOD generalization.
• Describe model-based approaches including disentanglement- and causality-based graph models.
• Outline learning-strategy methods such as graph invariant learning, graph adversarial training, and graph self-supervised learning.

实验结果

研究问题

• RQ1 What is the formal problem of graph OOD generalization and why does it differ from in-distribution generalization?
• RQ2 How can graph OOD generalization methods be categorized and what are representative techniques in each category?
• RQ3 What theories underlie Graph OOD generalization and what datasets exist to evaluate it?
• RQ4 What are the key challenges and future directions for graph OOD generalization?
• RQ5 How do data, model, and learning-strategy approaches compare in addressing topology-level and feature-level distribution shifts?

主要发现

• A formal problem definition is provided for Graph OOD generalization, emphasizing shifts between training and testing distributions.
• A three-category taxonomy is proposed: Data (augmentation), Model (disentanglement and causality), and Learning Strategy (invariant learning, adversarial training, self-supervised learning).
• Structured, feature-wise, and mixed-type graph augmentations are surveyed with representative methods and their aims toward improving OOD generalization.
• Disentanglement-based and causality-based graph models are identified as core model approaches for robust representations under distribution shifts.
• The survey summarizes theories and datasets used to evaluate Graph OOD generalization and offers directions for future research.