A locally weighted learning method based on a data gravitation model for multi-target regression

Abstract

Locally weighted regression allows to adjust the regression models to nearby data of a query example. In this paper, a locally weighted regression method for the multi-target regression problem is proposed. A novel way of weighting data based on a data gravitation-based approach is presented. The process of weighting data does not need to decompose the multi-target data into several single-target problems. This weighted regression method can be used with any multi-target regressor as a local method to provide the target vector of a query example. The proposed method was assessed on the largest collection of multi-target regression datasets publicly available. The experimental stage showed that the performance of multi-target regressors can be significantly improved by means of fitting the models to local training data.