Keywords: graph neural network, graph edit distance
TL;DR: We propose GRAPHEDX, a novel neural model for approximate GED computation that flexibly incorporates variable edit costs.
Abstract: Graph Edit Distance (GED) measures the minimal cost to transform one graph into another through node and edge insertions, deletions, and substitutions. Despite its flexibility in adapting to different cost settings, existing neural models for GED have focused on fixed-cost scenarios, limiting their versatility. To address this, we propose GRAPHEDX, a novel neural model that handles GED computation with variable costs. By reformulating GED as a Quadratic Assignment Problem (QAP) and introducing a neural set divergence framework, GRAPHEDX approximates the QAP objective using differentiable surrogates based on node and node-pair embeddings. Extensive experiments demonstrate that GRAPHEDX not only adapts to customizable cost settings but also outperforms state-of-the-art methods in prediction accuracy, offering a flexible solution for diverse graph comparison tasks.
Submission Type: Extended abstract (max 4 main pages).
Poster: jpg
Poster Preview: jpg
Submission Number: 181
Loading