Constructing Compact Dual Ensembles for Efficient Active Learning

A good ensemble is one whose members are both accurate and diverse. Active learning requires a small number of highly accurate classifiers so that they will not disagree with each other too often. Ensemble method, however, are not good candidates for active learning because of their different design purposes. In this paper, we propose to use dual ensembles for active learning in binary-class domains, and investigate how to use the diversity of the member classifiers of an ensemble for efficient active learning. As active learning requires iterative training of the member classifiers in an ensemble, it is imperative to maintain a small number of classifiers in an ensemble for learning efficiency. We empirically show using benchmark data that (1) number of classifiers varies fro different data sets to achieve a good (stable) ensemble; (2) feature selection can be applied to classifier selection to construct compact ensembles with high performance. A real-world application is used to demonstrate the effectiveness of the proposed approach.