Mine - a framework for dynamic regressor selection

Abstract

Dynamic Regressor Selection (DRS) systems work by selecting the most competent regressors from an ensemble to estimate the target value of a given test pattern. Hence, the central issue in dynamic selection techniques is how to define the competence of the regressors to select the most competent ones. This competence is usually quantified using a single measure, such as the performance of the regressors in local regions of the feature space around the test pattern, called the region of competence. However, to decide what is the best measure to correctly calculate the level of competence is a hard task, because no one is the best for any task. Works using ensemble of classifiers present a wide variety of measures that are used to calculate the competence. Using ensemble of regressors, many of these measures can not be used or adapted. Thus, in this work, we present a framework for DRS, called Meta INtEgration (MINE), that aims at selecting and combining the most competent regressors from a homogeneous ensemble during the evaluation of a given test pattern. The proposed framework uses the combination of different measures extracted from the region of competence, as a criterion for the selection and combination of the regressors. Also, we have done a survey in the literature on some measures used with regression problems to test the performance of the dynamic regression selection algorithms found in the literature. The measures are extracted from region of competence and they are aimed at capturing different behaviors of the regressors. Thus, for each test pattern, only the most competent regressors are selected and combined. Using the MINE framework, comprehensive experiments on 20 regression datasets show that MINE improves the final estimate performance when compared to state-of-the-art techniques. Also, experiments are performed on 15 real regression problems datasets using the state-of-the-art dynamic regressor selection techniques by changing only the measure that computes the competence. The results show that the measures have different performance throughout the datasets and none of them are better in all situations.