Out-of-Distribution Generalization on Graphs: A Survey

Area of Science:

• Graph machine learning
• Artificial intelligence
• Data science

Background:

• Most graph machine learning relies on the in-distribution hypothesis, assuming training and testing data are identically distributed.
• Real-world graph data often exhibits distribution shifts, leading to significant performance degradation in existing models.
• Out-of-distribution (OOD) generalization for graphs is crucial for robust performance in diverse, dynamic environments.

Purpose of the Study:

• To provide a comprehensive survey of out-of-distribution (OOD) generalization techniques for graph machine learning.
• To formally define the OOD generalization problem in the context of graph data.
• To review and categorize recent advances in OOD generalization for graphs.

Main Methods:

• Categorization of existing OOD generalization methods into three classes: data-centric, model-centric, and learning strategy-centric.
• Detailed discussion of methods within each category based on their integration into the graph machine learning pipeline.
• Review of theoretical foundations and commonly used benchmark datasets for evaluating OOD graph generalization.

Main Results:

• A structured overview of the current landscape of OOD generalization in graph machine learning.
• Identification of key challenges and limitations in existing approaches.
• Synthesis of theoretical underpinnings and empirical evaluation strategies.

Conclusions:

• OOD generalization is a critical area for advancing graph machine learning beyond the in-distribution assumption.
• A systematic categorization provides a framework for understanding and developing new OOD graph methods.
• Future research should focus on addressing the identified challenges to improve the robustness and applicability of graph models in real-world scenarios.