Application of the Fuzzy Classification for Linear Hybrid Prediction Methods

The paper discusses the problem of forecasting for samples with real-valued attributes. The goal is to estimate the effect of generated binary attributes on forecasting accuracy for the linear regression and the hybrid methods based on clustering. The initial set of attributes is expanded by binary attributes which are derived from the initial set by fuzzy classification. A comparative testing of the discussed forecasting methods on the initial samples and the resulting ones is performed. The test results on three different databases showed that the use of generated attributes for the classical linear regression resulted in the significant increase of the forecasting accuracy. In case of the linear regression with the clustering based on k-means the increase of forecasting accuracy was also observed. In case of the linear regression with the clustering based on the knn–method we registered a slight decrease, and an unstable result was obtained for the double linear regression.