Dataset analysis for classifier ensemble enhancement

Faculty of Engineering and Architecture Department of Electrical and Electronic Engineering Doctor of Philosophy Dataset Analysis for Classifier Ensemble Enhancement by Emanuele Tamponi We developed three different methods for dataset analysis and ensemble enhancement. They share the underlying idea that an accurate preprocessing and adaptation of the data can improve the system performance, without changing the classification model. Correlation Score is a generic framework for assessing encoding techniques by measuring the correlation between the encoded feature vectors and the corresponding class labels; experiments show its effectiveness in discovering the best encoding configurations between those tested, on a wide range of classification domains. Multi-Resolution Complexity Analysis is a method for assessing the local complexity inside a given domain. It is able to split a domain into regions of different classification complexity, giving insights on the inner structure of the populations inside the domain. Finally, Forests of Local Trees are a novel training algorithm for ensemble classifiers. They are based on the concept of local trees: classifiers trained with a bias toward a certain region of the domain. This bias enhances the diversity inside the ensemble, leading to improved performance. These three topics are meant as a foundation for a more complex framework, that will eventually utilize them organically. Emanuele Tamponi gratefully acknowledges Sardinia Regional Government for the financial support of her PhD scholarship (P.O.R. Sardegna F.S.E. Operational Programme of the Autonomous Region of Sardinia, European Social Fund 2007-2013 – Axis IV Human Resources, Objective l.3, Line of Activity l.3.1.).