Data Distillation for Enhancing Neural Frameworks in Learning Graph Distance Functions

Graph Edit Distance (GED) is a widely used method for measuring the similarity between two graphs, with numerous applications in domains such as molecular comparison and social network analysis. However, computing GED is NP-hard, making exact calculations computationally expensive for large datasets. While recent neural network-based approaches have emerged to approximate GED, these methods are typically supervised and require large amounts of training data with exact GED computations, which has a significant computational cost. To address the challenge of requiring large amounts of computational resources for training neural networks, data distillation techniques have been proposed in the context of graph and node classification tasks, but there is a lack of similar methods specifically for GED prediction. In this work, we explore approaches for designing data distillation techniques specifically for the GED task. One such approach leverages computational trees to reduce the size of training datasets required for GED prediction models. These approaches can improve efficiency and scalability of GED prediction models, with potential applications across various graph-based domains.