A framework for dynamic classifier selection oriented by the classification problem difficulty

Abstract

This paper describes a framework for Dynamic Classifier Selection (DCS) whose novelty resides in its use of features that address the difficulty posed by the classification problem in terms of orienting both pool generation and classifier selection. The classification difficulty is described by meta-features estimated from problem data using complexity measures. Firstly, these features are used to drive the classifier pool generation expecting a better coverage of the problem space, and then, a dynamic classifier selection based on similar features estimates the ability of the classifiers to deal with the test instance. The rationale here is to dynamically select a classifier trained on a subproblem (training subset) having a similar level of difficulty as that observed in the neighborhood of the test instance defined in a validation set. A robust experimental protocol based on 30 datasets, and considering 20 replications, has confirmed that a better understanding of the classification problem difficulty may positively impact the performance of a DCS. For the pool generation method, it was observed that in 126 of 180 experiments (70.0%) adopting the proposed pool generator allowed an improvement of the accuracy of the evaluated DCS methods. In addition, the main results from the proposed framework, in which pool generation and classifier selection are both based on problem difficulty features, are very promising. In 165 of 180 experiments (91.6%), it was also observed that the proposed DCS framework based on the problem difficulty achieved a better classification accuracy when compared to 6 well known DCS methods in the literature.