Discrete Particle Swarm Optimization with local search strategy for Rule Classification

Abstract

Rule discovery is an important classification method that has been attracting a significant amount of researchers in recent years. Rule discovery or rule mining uses a set of IF-THEN rules to classify a class or category. Besides the classical approaches, many rule mining approaches use biologically-inspired algorithms such as evolutionary algorithms and swarm intelligence approaches. In this paper, a Particle Swarm Optimization based discrete implementation with a local search strategy (DPSO-LS) was devised. The local search strategy helps to overcome local optima in order to improve the solution quality. Our DPSO-LS uses the Pittsburgh approach whereby a rule base is used to represent a `particle'. This rule base is evolved over time as to find the best possible classification model. Experimental results reveal that DPSO-LS outperforms other classification methods in most cases based on rule size, TP rates, FP rates, and precision.