Neagging: an aggregation procedure based on normalized entropy

Abstract

The analysis of big data, namely in inhomogeneous large-scale data under the regression analysis context, is a research topic with growing interest in recent years, where bagging and magging are two well-known aggregation procedures. As this kind of data may be recorded in different time regimes or may be taken from multiple sources, inhomogeneities are expected to be present, compromising regression modelling. The classical framework of independent and identically distributed errors related to a single underlying model does not apply and the usual alternatives (such as time-varying coefficients models or mixture models, for instance) may represent prohibitive computational burden. This paper revises the methodology developed in a recent work where an aggregation procedure based on normalized entropy was proposed, with very promising results, and illustrates its performance with real data applications considering distinct scenarios.