Melvin's digital garden

Leardi2000

CREATED: 200907010656 LINK: url:~/Modules/Literature/Leardi2000.pdf Efficient feature selection can improve predictive ability of models and reduce complexity (Ref 3)

Methods of feature selection in PLS models for spectral data include

  • iterative variable selection (Ref 4)
  • uninformative variable elimination (Ref 5)
  • iterative predictor weighting (Ref 6)

Drawback of techniques is that selected features (wavelengths) are scattered throughout the spectrum.

GA produced more interpretable results as selected wavelengths are less dispersed.

Techniques of feature selection usually assumes that is no autocorrelation among the variables, this does not hold for spectral data. If wavelength n is relevant, wavelengths n-1 and n+1 should also be relevant.

Links to this note