Boosting Graph Similarity Search through Pre-Computation

Abstract

Graph similarity search is to retrieve all graphs from a graph database whose graph edit distance (GED) to a query graph is within a given threshold. As GED computation is NP-hard, existing solutions adopt the filtering-and-verification framework, where the main focus is on the filtering phase to reduce the number of GED verifications. However, existing filtering techniques have inherently limited filtering capabilities, and suffer from a large number of GED verifications. To address the problem, in this paper, we propose a fundamentally different approach that utilizes pre-computed GEDs between data graphs in the filtering phase. Based on the approach, we develop a novel search framework Nass, which substantially reduces the verification workload. Because the efficiency of GED computation is essential in GED pre-computation, not to mention the verification of candidate graphs, we also propose an efficient GED computation algorithm as a part of Nass. We conduct extensive experiments on real datasets, and show Nass significantly outperforms the state-of-the art solutions.