Model Selection and Predictive Inference

Gilbert Saporta (1)

1 CEDRIC - MSDMA - CEDRIC. Méthodes statistiques de data-mining et apprentissage

Methods based on penalized likelihood cannot be applied in many problems. Statistical learning theory provide the theoretical framework for predictive inference, but model choice based on VC dimension is often not feasible. In binary classification, ROC curve and AUC provide a reasonable criterium for model choice, combined with resampling.

Mots clés: model selection; ridge regression; statistical learning; discriminant analysis

Domaines: Statistiques [math.ST]

Gilbert Saporta. Model Selection and Predictive Inference. Trends and Challenges in Applied Mathematics, Technical University of Civil Engineering, Jun 2007, Bucarest, Romania. pp.92-97. ⟨hal-01125334⟩