Unsupervised Temporal Rule Mining with Genetic Programming and Specialized Hardware

Rule mining is the practice of discovering interesting and unexpected rules from large data sets. Depending on the exact problem formulation, this may be a very complicated problem. Existing methods typically make strong simplifying assumptions about the form of the rules, and limit the measure of rule quality to simple properties, such as confidence. Because confidence in itself is not a good indicator of how interesting a rule is to the user, the mined rules are typically sorted according to some secondary interestingness measure. In this paper we present a rule mining method that is based on genetic programming. Because we use specialized pattern matching hardware to evaluate each rule, our method supports a very wide range of rule formats, and can use any reasonable fitness measure. We develop a fitness measure that is well-suited for our method, and give empirical results of applying the method to synthetic and real-world data sets.