Modeling Pattern Set Mining Using Boolean Circuits

Abstract

Researchers in machine learning and data mining are increasingly getting used to modeling machine learning and data mining problems as parameter learning problems over network structures. However, this is not yet the case for several pattern set mining problems, such as concept learning, rule list learning, conceptual clustering, and Boolean matrix factorization. In this paper, we propose a new modeling language that allows modeling these problems. The key idea in this modeling language is that pattern set mining problems are modeled as discrete parameter learning problems over Boolean circuits. To solve the resulting optimisation problems, we show that standard optimization techniques from the constraint programming literature can be used, including mixed integer programming solvers and a local search algorithm. Our experiments on various standard machine learning datasets demonstrate that this approach, despite its genericity, permits learning high quality models.