Learning from pairwise comparisons : an empirical analysis

Details

Supervisors

Hendrickx, Julien

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées, à finalité spécialisée

Abstract

We are interested in inferring the inherent ratings of a set of items based on pairwise comparisons among them. More specifically, we expand on the results of Hendrickxet al., 2019 [10] and 2020 [12] and empirically handle three problematics about pairwise comparisons. We analyze the quality of the average graph resistance as an upper bound to the error rate of the empirical method and show that even in the worst case scenario this bound holds. We improve the accuracy of the estimation by considering the minimum number of victories of an item and by designing selection tools allowing to select the best performing method for a given problem. Two selection tools are implemented: one based on a preliminary estimation of the weights and the other on a decision tree. Finally, the surprising behaviour of the empirical method performing better than the artificial method is explained by considering realisations that differ from their expectation.