Graph Reachability Pruning: Adaptive Data Reduction for Inexact Subgraph Matching

Abstract

Subgraph matching is a fundamental graph analysis capability that many fields in science and technology depend on to retrieve information from data that are structured as graphs. These fields span network and cyber security, social and digital media, and the physical and life sciences. Because subgraph isomorphism is an NP-Complete problem, with depth-first search approaches having factorial time complexity and breadth-first approaches having factorial space complexity, efficient search space reduction is fundamental to the timely execution of queries against large, real world graphs having hundreds of millions or more nodes and edges. In this paper, we introduce Graph Reachability Pruning (GRP), a neighborhood-based data reduction algorithm that discards all infeasible and many suboptimal regions of a knowledge graph prior to performing subgraph matching. The algorithm automatically adapts a missing neighbor tolerance parameter to enable structurally inexact matching without requiring cumbersome parameter tuning by an analyst and without relying on a space consuming pre-computed index. It outperforms the only known alternative algorithm, Inexact Local Constraint Checking (ILCC), which applies a single tolerance parameter uniformly across an entire query graph. In our real world case study, ILCC often fails to reduce knowledge graph data by more than a single digit percentage, whereas GRP often discards more than ninety-nine percent of the data. GRP operates independently from any search algorithm it might be paired with. Other data reduction methods are often integrated into, and operate concurrently with, a coupled search algorithm. This property of GRP gives it a versatility that is exceptionally useful.