Utilizing Diversity and Performance Measures for Ensemble Creation

Abstract

An ensemble is a composite model, aggregating multiple base models into one predictive model. An ensemble prediction, consequently, is a function of all included base models. Both theory and a wealth of empirical studies have established that ensembles are generally more accurate than single predictive models. The main motivation for using ensembles is the fact that combining several models will eliminate uncorrelated base classifier errors. This reasoning, however, requires the base classifiers to commit their errors on different instances – clearly there is no point in combining identical models. Informally, the key term diversity means that the base classifiers commit their errors independently of each other. The problem addressed in this thesis is how to maximize ensemble performance by analyzing how diversity can be utilized when creating ensembles. A series of studies, addressing different facets of the question, is presented. The results show that ensemble accuracy and the diversity measure difficulty are the two individually best measures to use as optimization criterion when selecting ensemble members. However, the results further suggest that combinations of several measures are most often better as optimization criteria than single measures. A novel method to find a useful combination of measures is proposed in the end. Furthermore, the results show that it is very difficult to estimate predictive performance on unseen data based on results achieved with available data. Finally, it is also shown that implicit diversity achieved by varied ANN architecture or by using resampling of features is beneficial for ensemble performance.