Deep Ranking with Heterogeneous Effect

Classical parametric ranking models, such as the Placket–Luce model, often fail to distinguish an object’s intrinsic utility from context-driven advantages. To address this, we propose a semiparametric framework that models the log-score as an additive combination of a latent parameter and a non-linear covariate effect approximated by a deep neural network. We establish model identifiability under mild hypergraph connectivity assumptions and existence of the maximum likelihood estimator. Our analysis characterizes the non-asymptotic error from both the high-dimensional intrinsic score estimator and the neural component. Empirically, the method outperforms baselines on synthetic and professional tennis data, successfully capturing complex, intransitive interactions.