Ensemble Learning With An Archive-based Genetic Algorithm For Feature Subspace

Nowadays,data generated in business,engineering,medicine and other fields has risen excessively.The data usually has a large number of attributes.Redundant or irrelevant features may increase the complexity of learning models,and even damage the performances.Therefore,how to select an effective feature subset,and improve the performance of learning model is one of big challenges in data mining and other application fields.Firstly,the thesis introduces some common theories and methods of feature selection which affords an effective way to improve the classification accuracies,reduce the complexities of learning models,and enhance the interpretability of the models.Secondly,the thesis does a brief literature review on subspace ensemble learning which combines subspace learning with ensemble models.Thirdly,we propose a feature subspace ensemble learning model based on the archive genetic algorithm.This model contains two populations.One is the evolutionary population.The individuals are the various subspace selections of the original feature space.The multi-objective optimization strategy is used to optimize the individuals in a niching genetic algorithm.The other is the archived population.The individuals in the evolutionary population provides candidates for updating the archived population.The elite individuals of the evolutionary population in each generation are used to update the archive population.These final elites in the archive are the subspaces for the base classifiers in the ensemble.The proposed model leverages the diversity and the accuracy based on the two populations.Experimental results show that the proposed model can outperform other conventional ensemble learning algorithms.The proposed model can achieve good performances with small feature subset in both traditional datasets and customer credit dataset.In this paper,the archive genetic algorithm is used to produce diverse and accurate subspaces to achieve small feature subsets and further improve the classification performance.