Classifiers selection for ensemble learning based on accuracy and diversity

Abstract Ensemble learning is a learning method where a collection of a finite number of classifiers is trained for the same classification task and thus it can gain better performance at the cost of computation. Previous research has proved that it may be better to ensemble many instead of all of classifiers at hand. Thus classifiers selection became a crucial problem for ensemble learning. To select the best classifier set from a pool of classifiers, the classifier diversity is the most important property to be considered. In this paper, a kind of selection method based on accuracy and diversity is proposed in order to achieve better classification performance. Classifiers correlation in our method is calculated using Q statistics diversity measures based on correlation between errors. Experiments were carried out on five data sets from UCI Machine Learning Repository. Twenty classifiers and six combination rules were included in our experiments. The experimental results are encouraging and validate the effectiveness of the proposed classifiers selection method.