Graph Data Selection for Domain Adaptation: A Model-Free Approach

Graph domain adaptation (GDA) is a fundamental task in graph machine learning, with techniques like shift-robust graph neural networks (GNNs) and specialized training procedures to tackle the distribution shift problem. Although these modelcentric approaches show promising results, they often struggle with severe shifts and constrained computational resources. To address these challenges, we propose a novel model-free framework, GRADATE (GRAph DATa sElector), that selects the best training data from the source domain for the classification task on the target domain. GRADATE picks training samples without relying on any GNN model's predictions or training recipes, leveraging optimal transport theory to capture and adapt to distribution changes. GRADATE is data-efficient, scalable and meanwhile complements existing model-centric GDA approaches. Through comprehensive empirical studies on several real-world graph-level datasets and multiple covariate shift types, we demonstrate that GRADATE outperforms existing selection methods and enhances off-the-shelf GDA methods with much fewer training data.

How to select the most relevant source domain data, based on available validation data, for better graph-level classification accuracy evaluated on the target domain ?

In this paper, we propose a model-free method, GRADATE (GRAph DATa sElector), that selects a subset of important training data in the source domain independently of any specific GNN model design, making it both data-efficient and versatile. GRADATE reduces computational overhead and enables quick adaptation to unseen graph domains based on available validation data. Conceptually, GRADATE first leverages Fused Gromov-Wasserstein (FGW) distance [57] to compare graph samples. We provide a theoretical justification to demonstrate FGW's unique advantage over multi-layer GNNs for graph comparison. Then, FGW is used as a building block to measure the dataset-level distance between training and validation sets, which is termed as Graph Dataset Distance (GDD). Through 39th Conference on Neural Information Processing Systems (NeurIPS 2025).