New Ensemble Method for Prediction Improvement with Application in Telecommunication

In this paper we propose a system for invoice amount prediction. The approach we applied is based on ensemble methods integrating information from many models predicting customer´s invoice amount. Results generated by different models can be decomposed into constructive and destructive components. The elimination of the destructive ones should improve final results. After basis components are estimated by Blind Signal Separation methods, the main problem is their proper classification as destructive or constructive. For this task we propose a multicriteria system. The example from the telecom market confirms validity of our approach.