Signature Pattern Covering via Local Greedy Algorithm and Pattern Shrink

Pattern mining is a fundamental problem that has a wide range of applications. In this paper, we study the problem of finding a minimum set of signature patterns that explain all data. In the problem, we are given objects where each object has an item set and a label. A pattern is called a signature pattern if all objects with the pattern have the same label. This problem has many interesting applications such as assertion mining in hardware design and identifying failure causes from various log data. We show that the previous pattern mining methods are not suitable for mining signature patterns and identify the problems. Then we propose a novel pattern enumeration method which we call Pattern Shrink. Our method is strongly coupled with another novel method that is very similar to finding a local optimum with a negligible loss in performance. Our proposed methods show a speedup of more than ten times over the previous methods. Our methods are flexible enough to be extended to mining high confidence patterns, instead of signature patterns.