Instance Hardness as a Decision Criterion on Dynamic Ensemble Structure

Abstract

Dynamic Feature Selection selects the best feature subset for each individual instance. In other words, each instance will be classified using its own feature subset. Recently, this idea has been combined with other dynamic approaches to create dynamic ensembles, in which an ensemble structure is selected for each testing instance. Nevertheless, dynamic selection approaches (feature and classifiers) are time consuming, once they aim to select a different setting for each testing instance. The integration of these two dynamic approaches in ensembles have obtained good accuracy results, but at the same time, the computational complexity increases when these approaches are used together. Due to this fact, this paper introduces the idea of using a decision criterion to reduce the complexity of the dynamic approaches in ensembles. The main aim is to use a complex dynamic ensemble for difficult instances, while the other instances will be classified by single classifiers. In order to validate the proposed approach, an empirical analysis will be conducted. The results of the empirical analysis indicate that is possible to reduce the processing time using a simple decision criteria, Instance Hardness, without deteriorating the accuracy of these systems.