Fast Bayesian clustering and model selection for longitudinal data mixtures

Marco Corneli (1) , Elena Erosheva

1 Center of Modeling, Simulation & Interactions

The clustering of longitudinal data from a Bayesian perspective is considered , with particular attention to the selection of the number of components. Instead of using asymptotic criteria (e.g. BIC), we propose to directly maximize an exact quantity based on conjugated prior distributions of the model parameters. The prior parameters are estimated by gradient descent, via automatic differentiation. Using simulated data, we demonstrate that, in terms of accuracy of the obtained clustering, our approach is comparable to two frequentist approaches commonly used in this setting, and it outper-forms them in selecting the actual number of clusters.

Domaines: Statistiques [stat]; Machine Learning [stat.ML]

Marco Corneli, Elena Erosheva. Fast Bayesian clustering and model selection for longitudinal data mixtures. 2019. ⟨hal-02310069v1⟩