Efficient Search in Graph Edit Distance: Metric Search Trees vs. Brute Force Verification

TL;DR

This work tackles the efficiency of graph similarity search by evaluating whether Cascading Metric Trees (CMT) with Upper-Lower Bounds (UBLB) can outperform brute-force verification in Graph Edit Distance (GED) based search. The methodology builds a precomputed CMT, prunes queries using UBLB, and uses brute-force verification on a suspected set to filter false positives, with evaluation on PubChem graph data. Contrary to expectations, CMT did not consistently outperform brute-force verification and often lagged in speed, suggesting that computing GED upper and lower bounds remains a bottleneck. The findings highlight the need for new approaches to accelerate GED computations and call for broader testing across datasets and optimized implementations to realize any potential gains from tree-based pruning.

Abstract

This report evaluates the efficiency of Graph Edit Distance (GED) computation for graph similarity search, comparing Cascading Metric Trees (CMT) with brute-force verification. Despite the anticipated advantages of CMT, our findings indicate it does not consistently outperform brute-force methods in speed. The study, based on graph data from PubChem, suggests that the computational complexity of GED-based GSS remains a challenge.