A Comprehensive Survey and Experimental Study of Subgraph Matching: Trends, Unbiasedness, and Interaction

Abstract

Subgraph matching is a fundamental problem in graph analysis. In recent years, many subgraph matching algorithms have been proposed, making it pressing and challenging to compare their performance and identify their strengths and weaknesses. We observe that (1) The embedding enumeration in the classic filtering-ordering-enumerating framework dominates the overall performance, and thus enhancing the backtracking paradigm is becoming a current research trend; (2) Simply changing the limitation of output size results in a substantial variation in the ranking of different methods, leading to biased performance evaluation; (3) The techniques employed at different stages of subgraph matching interact with each other, making it less feasible to replace and evaluate a single technique in isolation. Therefore, a comprehensive survey and experimental study of subgraph matching is necessary to identify the current trends, ensure unbiasedness, and investigate the potential interactions. In this paper, we comprehensively review the methods in the current trend and experimentally confirm their advantage over prior approaches. We unbiasedly evaluate the performance of these algorithms by using an effective metric, namely embeddings per second. To fully investigate the interactions between various techniques, we select 10 representative techniques for each stage and evaluate all the feasible combinations.