Adaptive Splitting and Selection Algorithm for Regression

Developing system for regression tasks like predicting prices, temperature is not a trivial task. There are many of issues which must be addressed such as: selecting appropriate model, eliminating irrelevant inputs, removing noise, etc. Most of them can be solved by application of machine learning methods. Although most of them were developed for classification tasks, they can be successfully applied for regression too. Therefore, in this paper we present Adaptive Splitting and Selection for Regression algorithm, whose predecessor was successfully applied in many classification tasks. The algorithm uses ensemble techniques whose strength comes from exploring local competences of several predictors. This is achieved by decomposing input space into disjointed competence areas and establishing local ensembles for each area respectively. Learning procedure is implemented as a compound optimisation process solved by means of evolutionary algorithm. The performance of the system is evaluated in series of experiments carried on several benchmark datasets. Obtained results show that proposed algorithm is valuable option for those who look for regression method.