Boosting Graph Similarity Search through Pre-Computation

Efficient Top-k Graph Similarity Search With GED Constraints

“…Similar to existing work for threshold-based similarity search [31], [37], in Definition 2, we introduce a maximum threshold τ max for top-k graph similarity search to prevent excessive GED computation for a large k.…”

“…The label set-based lower bound in Definition 5 is the most widely used lower bound function to estimate the edit cost of the unmapped part. DEFINITION 5 (Label set-based lower bound [30], [37]). The label set-based lower bound between two graphs g 1 and g 2 is defined as:…”

“…For example, Inves [29] captures edit errors in bridges, which are edges connecting mapped vertices to unmapped vertices. Nass [37] applies different lower bound functions to the unmapped part to obtain a tighter lower bound.…”

“…ParsK builds a hierarchical partition-based index motivated by a top-k string similarity search technique [41]. However, recent work has revealed that the filtering power of feature-based indexing techniques is inherently limited [29], [33], [37]. For example, Inves [29] and AStar + -LSa [33] have empirically showed that efficient verification techniques invalidate the filtering effect of existing feature-based indexing techniques.…”

“…Based on the algorithms, we explore two search frameworks that reduce GED computation overheads of unpromising graphs (Section III-B and III-C). To utilize the latest pre-computed GED-based index proposed in our previous work [37], we also propose a hybrid search framework that integrates threshold-based approach with our top-k search techniques (Section III-D). We conduct extensive experiments on widely used real datasets and show that the proposed frameworks significantly improve the performance of top-k graph similarity search (Section IV).…”