On estimating model complexity and prediction errors in multivariate calibration: generalized resampling by random sample weighting (RSW)

摘要: Thepresentpaper focusesondeterminingthenumberofPLScomponentsbyusingresamplingmethodssuchascrossvalidation (CV), Monte Carlo cross validation (MCCV), bootstrapping (BS), etc. To resample the training data, randomnon-negative weights are assigned to original samples and a sample-weighted PLS model is developedwithout increasing computational burden much. Random weighting generalization of traditionalresampling methods expected have lower risk getting an insufficient set. For prediction,only with random less than threshold value selected ensure that predictionsamples influence on training. complicated because optimal number components isoften not unique or readily distinguished there might exist region complexity, thedistribution prediction errors can be more useful single root mean squared error prediction(RMSEP). Therefore, distribution estimated by repeated sample andused determine complexity. RSW compared its traditional counterparts like CV, MCCV, BS arecently proposed randomization test method demonstrate usefulness. Copyright 2010 John Wiley & Sons,Ltd.Keywords: complexity; weighting; validation; bootstrapping;