Apprentissage de données fonctionnelles par modèles multi-tâches : Application à la prédiction de performances sportives

Abstract

Gaussian process regression is a common tool of supervised learning that provides a convenient probabilistic framework, leading to predictions with proper uncertainty quantification. The learning procedure in such model generally focuses on hyper-parameters estimation of the covariance structure, rather than the prior mean of the process. Therefore, the quality of prediction might severely decrease, with an inappropriate prior mean, as we move away from observation points. This presentation introduces a multi-task extension of the GP framework, where data are supposed to come from several individuals sharing some structure altogether. This approach offers more reliable predictions even when a new individual is observed on few or sparse input locations. Then, the model is enhanced with a clustering component to provide cluster-specific GP predictions. This work has initially been tailored to handle the problem of talent identification in sports, and the talk is thus illustrated with this application involving function data.