Beyond majority: Label ranking ensembles based on voting rules

Label ranking is a machine learning task that deals with mapping an instance to a ranking of labels, representing the labels’ ordered relevance to the instance. Three recent studies have suggested the use of ensembles to improve the performance of simple label ranking models. However, none of them has explicitly inspected the question of how should the results obtained by simple models be aggregated into a single combined output. While classification tasks and regression tasks typically employ trivial aggregation techniques (i.e., majority voting and averaging, respectively), the case of label ranking tasks is not straightforward. The contribution of this paper is twofold. First, we propose to apply voting rules, typically used in the field of social choice, as the aggregation technique for label ranking ensembles. Our evaluation reveals that there is no single rule that consistently outperforms all other voting rules, and that under different settings different voting rules perform the best. Second, we propose a novel aggregation method for label ranking ensembles that learns the best voting rule to be used in a given setting. An extensive evaluation of the proposed method on semi-synthetic as well as real-world datasets shows that it obtains prediction performance that is significantly higher than that of the aggregation techniques currently used by state-of-the-art label ranking ensembles.