Decision-Focused Data Pooling for Contextual Stochastic Optimization
Résumé
Data scarcity poses a significant risk that hinders the deployment of advanced datadriven methods. In many cases of practical interest, decision-makers have access to data from similar, potentially unrelated, problem instances. Maximizing the benefits of data-driven methods thus necessitates novel methods to utilize all available data. In this work, we propose two methods to pool data when dealing with multiple contextuallydependent stochastic optimization problems. The first involves naively pooling data and training a global model to derive decisions across all problems, while the second leverages optimal transport for model aggregation. An essential contribution is the
Origine : Fichiers produits par l'(les) auteur(s)