A New Micro-Batch Approach for Partial Least Square Clusterwise Regression

Gaël Beck (1, 2) , Hanane Azzag (1) , Stéphanie Bougeard (3) , Mustapha Lebbah (1) , Ndèye Niang (4)

1 UP13 - Université Paris 13
2 LIPN - Laboratoire d'Informatique de Paris-Nord
3 ANSES - Agence nationale de sécurité sanitaire de l'alimentation, de l'environnement et du travail
4 CEDRIC - MSDMA - CEDRIC. Méthodes statistiques de data-mining et apprentissage

Current implementations of Clusterwise methods for regression when applied to massive data either have prohibitive computational costs or produce models that are difficult to interpret. We introduce a new implementation Micro-Batch Clusterwise Partial Least Squares (mb-CW-PLS), which is consists of two main improvements: (a) a scalable and distributed computational framework and (b) a micro-batch Clusterwise regression using buckets (micro-clusters). With these improvements, we are able to produce interpretable regression models with multicollinearity within a reasonable time frame.

Keywords: PLS; Clusterwise; Spark

Domains: Statistics [stat]

Gaël Beck, Hanane Azzag, Stéphanie Bougeard, Mustapha Lebbah, Ndèye Niang. A New Micro-Batch Approach for Partial Least Square Clusterwise Regression. Procedia Computer Science, 2018, 144, pp.239-250. ⟨10.1016/j.procs.2018.10.525⟩. ⟨hal-02471601⟩