Efficient inference of rankings from multibody comparisons

Many of the existing approaches to assess and predict the performance of players, teams, or products in competitive contests rely on the assumption that comparisons occur between pairs of such entities. There are, however, several real contests where more than two entities are part of each comparison, e.g., sports tournaments, multiplayer board and card games, and preference surveys. The Plackett-Luce (PL) model provides a principled approach to infer the ranking of entities involved in such contests characterized by multibody comparisons. Unfortunately, traditional algorithms used to compute PL rankings suffer from slow convergence, which limits the application of the PL model to relatively small-scale systems. Here, we present an alternative implementation that allows for significant speed-ups and validate its efficiency in both synthetic and real-world sets of data. Further, we perform systematic cross-validation tests concerning the ability of the PL model to predict unobserved comparisons. We find that a PL model trained on a set composed of multibody comparisons is more predictive than a PL model trained on a set of projected pairwise comparisons derived from the very same training set, emphasizing the need of properly accounting for the true multibody nature of real-world systems whenever such information is available.