Ranking problems in modern applications often come with covariate vectors whose dimension grows with the number of items being compared. Ignoring these covariates can leave predictive performance on the table, but including all of them naively is statistically dangerous. The authors study a Plackett–Luce model that combines item-specific utilities with high-dimensional covariate effects, assuming the covariate coefficients are sparse.
They derive a necessary and sufficient condition for identifiability, expressed in terms of the Kruskal rank of the restricted covariate design. On the estimation side, they propose an ℓ1-regularized maximum likelihood estimator and show that, under suitable design and connectivity conditions, the estimator exists and is unique with probability tending to one. They also prove uniform consistency and, with an additional minimum-signal condition, exact support recovery.
The theoretical analysis combines a hypergraph-based chaining technique with a profile-likelihood perturbation argument, yielding rates that explicitly depend on the comparison hypergraph's connectivity and the sparsity level. Numerical experiments on synthetic data and a horse racing dataset corroborate the theoretical findings, suggesting the method is both practically usable and theoretically sound.