A Hybrid Genetic Algorithm for Simultaneous Feature Selection and Rule Learning

This paper proposes a hybrid genetic rule learning algorithm which incorporating feature selection technique. The chromosome of rule individual composed of two vectors: a rule condition vector representing the conjunction of rule conditions and a feature selection vector representing the selected features. In order to improve the performance of the algorithm, a local search method embedded in the evolution process is proposed. In the local search procedure, the minimum information entropy heuristic is used to specify the importance of features. Irrelevant features are removed and useful features are added. When adding a relevant feature, the corresponding rule condition is also adjusted to improve the rule quality. Experiments show that this hybrid model works well in practice.