Alternate deflation and inflation of search space in reweighted sampling: An effective variable selection approach for PLS model

Based on assessment of randomized sub-model populations generated through reweighted binary matrix sampling (BMS), an innovative variable selection strategy for PLS regression model, called alternate deflation and inflation of search space (ADISS) is proposed. Normalized regression coefficients of best PLS sub-models population is used to formulate the weight vector for re-weighted BMS. Unlike the most existing algorithm, ADISS alternatively shifts between forward selection (inflation) and backward elimination (deflation) of variable space, minimizing the risk of accidental loss of informative variables. Compared with methods such as competitive adaptive reweighted sampling (CARS), variable iterative space shrinkage approach (VISSA), or Monte Carlo uninformative variable elimination (MC-UVE), proposed method showed lower cross-validation or prediction error for two different benchmark NIR data sets. ADISS frequently selects nearly the same sets of variables across multiple independent runs, that signifies stability of the output. The unsupervised execution, termination and projection of final variable set from the algorithm is important advantage while considering for large scale data.