Section: New Results

An AIC-like criterion for semi-supervised classification

Participants : Christophe Biernacki, Vincent Vandewalle.

In semi-supervised classification, generative models take naturally into account unlabeled data and parameter estimation can be easily performed through the EM algorithm. However, traditional model selection criteria either does not take into consideration the predictive purpose (AIC or BIC criteria) or involve a high computational cost because of the EM mechanism (cross validation criteria). Alternatively, we propose the penalized model selection criterion AICcond which aims to estimate the predictive power of a generative model by approximating its predictive deviance. AICcond has similar computational cost to AIC, owns good consistency theoretical properties and highlights encouraging behaviour for variable and model selection in comparison to other standard criteria.

This joint work with Gilles Celeux and Gérard Govaert is now published in[16] .