Extended AMUSE algorithm and novel randomness approach for BSS model aggregation with methodology remarks

Abstract

In this paper we propose application of extended AMUSE blind signal separation method to improve a model prediction. In our approach we assume, that results generated by any regression model usually include both constructive and destructive components. In case of a few models, some of the components can be common to all of them. Our aim is to find the basis elements via AMUSE algorithm and distinguish the components with the constructive influence on the modelling quality from the destructive ones. We extend the standard AMUSE algorithm for cases with strong noises. The crucial question is to determine number of delays used in separation process and define criterion for destructive components identification. We propose novel method of randomness analysis to solve above problems. Due to complexity of the whole BSS aggregation method we include some methodological remarks as the framework for proposed approach.