Heuristic extraction of co-occurrence patterns for effective construction of graph-based interpretable decision sets

Graph-based Interpretable Decision Set, gIDS in short, is an unordered set of graph classification rules composed with subgraph patterns. gIDS is recognized as an accurate and human interpretable graph classifier, and thus it can expect to be utilized in a wide variety of application areas including safety-critical domains. However, building gIDS often requires unfeasible computation time, since we need to solve three data mining problems, namely enumeration of frequent subgraph patterns, derivation of subgraph-based class association rules, and finding the optimal combination among class association rules. In this paper, in order to make gIDS more practical by reducing the building time, we consider the replacement of the complete enumeration of association rules in the second stage by heuristic rule extraction, and develop a heuristic rule extractor by extending the COM (co-occurrence rule miner) that efficiently derives a small set of rules with high classification ability. Our developed heuristic extractor can suppress the number of rules obtained, which also expects to reduce the computation time for combinatorial optimization in the third stage. The results of experiments using real world dataset show that our proposal can reduce computation time while keeping the classification performance and interpretability.