Experiments with Feature-Prior Hybrid Ensemble Method for Classification
Junyang Zhao, Zhili Zhang, Chongzhao Han, Lijiang Sun · 2014
In multiple classifier systems, base classifiers are trained to be accurate and diverse by a set of training data. The generation of training data is necessary and important in classifier ensemble, which can be achieved by instance selection (IS) or feature selection (FS) on initial data. In this paper, a feature-prior FS-IS hybrid ensemble method is proposed by integrating feature selection with instance selection. The influence of instance selection to feature selection and the effect of feature-prior and instance-prior methods to ensemble accuracy are discovered. Dataset experiments indicate that feature-prior model generally performs better in comparison with previous instance-prior model, and feature selection is convincible to be used in classifier ensemble prior to instance selection.