Ensemble methods in ordinal data classification

Abstract

Ordinal classification problems can be found in various areas, such as product recommendation systems, intelligent health systems and image recognition. This problems have the goal of learning how to classify certain instances (e.g a movie) in an ordinal scale (e.g. good, average, bad). The performance of supervised learned problems (such as ordinal classification) can be improved by using ensemble methods, where various models are combined to perform better decisions. While there are various ensemble methods for nominal classification, ranking and regression, ordinal classification has not received the same level of attention. The goal of this dissertation is, therefore, to introduce novel ensemble methods for the classification of ordinal data. To do this, first a new ordinal classification algorithm based on decision trees and the data replication method is presented, whose results show that this classifier might perform better than other non-ordinal classifiers. Then, the main ideas of this method are exploited to try and improve ensembles whose models share similarities with decision trees (i.e. AdaBoost.M1 with Decision Stumps and Random Forests).