Ensemble learning for the partial label ranking problem

Abstract

The partial label ranking problem is a general interpretation of the preference learning scenario known as the label ranking problem, the goal of which is to learn preference classifiers able to predict a complete ranking with ties over the finite set of labels of the class variable. In this paper, we use unsupervised discretization techniques (equal-frequency and equal-width binning) to heuristically select the threshold for the numerical features in the algorithms based on induction of decision trees (partial label ranking trees algorithm). Moreover, we adapt the most well-known averaging (bootstrap aggregating and random forests) and boosting (adaptive boosting) ensemble methods to the partial label ranking problem, in order to improve the robustness of the built classifiers. We compare the proposed methods with the nearest neighbors-based algorithm (instance based partial label ranking) over the standard benchmark datasets, showing that our versions of the ensemble methods are superior in terms of accuracy. Furthermore, they are affordable in terms of computational efficiency.