Rule Discovery with Particle Swarm Optimization

This paper proposes Particle Swarm Optimization (PSO) algorithm to discover classification rules. The potential IF-THEN rules are encoded into real-valued particles that contain all types of attributes in data sets. Rule discovery task is formulized into an optimization problem with the objective to get the high accuracy, generalization performance, and comprehensibility, and then PSO algorithm is employed to resolve it. The advantage of the proposed approach is that it can be applied on both categorical data and continuous data. The experiments are conducted on two benchmark data sets: Zoo data set, in which all attributes are categorical, and Wine data set, in which all attributes except for the classification attribute are continuous. The results show that there is on average the small number of conditions per rule and a few rules per rule set, and also show that the rules have good performance of predictive accuracy and generalization ability.