Given that real-world classification tasks always have irrelevant or noisy features which degrade both prediction accuracy and computational efficiency, feature selection is an effective data reduction technique showing promising performance. This paper presents a cooperative coevolution framework to make the feature selection process embedded into the classification model construction within the genetic-based machine learning paradigm. The proposed approach utilizes the divide-and-conquer strategy to manage two populations in parallel, corresponding to the selected feature subsets and the rule sets of classifier respectively, in which a memetic feature selection algorithm is adopted to evolve the feature subset population while a Pittsburgh-style learning classifier system is used to carry out the classifier evolution. These two coevolving populations cooperate with each other regarding the fitness evaluation and the final solution is obtained via collaborations between the best individuals from each population. Empirical results on several benchmark data sets chosen from the UCI repository, together with a non-parametric statistical test, validate that the proposed approach is able to deliver classifiers of better prediction accuracy and higher stability with fewer selected features, compared with the original learning classifier system. In addition, the incorporated feature selection process is shown to help improve the computational efficiency as well.
展开▼
