Section: New Results

Dealing with Missing Not at Random Values in Model-based Clustering

Participant : Christophe Biernacki.

Missing values are current in modern data sets. In many situations, making the simplifying hypothesis that they are missing at random is not realistic. However, it is very challenging to propose sensible models which address the underlying missing process. We make such proposals specific to the clustering context, namely making the assumption that missing values are missing at random conditionally to clusters, thus leading to a quite natural not missing at random marginal model. A working paper is in progress.

It is a joint work with Julie Josse of Ecole Polytechnique and Gilles Celeux of Inria Saclay - Île de France.