Résumé : Title : Variable selection and outlier detection as a MIP
Abstract : Dimension reduction or feature selection is an effective strategy to handle contaminated data and to deal with high dimensionality while providing better prediction. To deal with outlier proneness and spurious variables, we propose a method performing the outright rejection of discordant observations together with the selection of relevant variables. To solve this problem, it is recasted as a mixed integer program which allows the use of efficient commercial solver. Also we propose an alternate projected gradient algorithm (proximal) so get a nice appoximated solution.
Lieu : A304
Notes de dernières minutes : Séminaire commun Pole 2 / Pole 3