On Hybrid SPLS-Elastic Net Feature Selection and Prediction: Application to Quality Assurance in Industrial Manufacturing Processes

Monday, May 5, 2014 - 1:30pm - 2:30pm
Lind 305
Guy-vanie Miakonkana (University of Minnesota, Twin Cities)
In this work we investigate on hybrid statistical methods based on a combination of Sparse Partial Least Squares methodology and Elastic Net regression. Through simulations, we first observe that while the Elastic Net methodology struggles much more than the Sparse Partial Least Squares in identifying noise variables, the Sparse Partial Least Square retains less relevant predictors in the models than the Elastic Net. In addition, the Sparse Partial Least Squares regression coefficients estimates of relevant predictors are inflated proportionally to the variance of the corresponding predictor. In order to mitigate the shortfalls of each methodology, we explore combinations of both techniques to form a hybrid Sparse Partial Least Squares-Elastic Net statistical method. A two-step approach that combines the strengths of these techniques is proposed so as to reduce the prediction error.