Feature Variable Extraction Using Competitive Adaptive Reweighting Sampling (CARS) Method

This article introduces Competitive Adaptive Reweighting Sampling (CARS), a feature variable extraction method for MATLAB-based pattern recognition (classification and regression). The algorithm utilizes Adaptive Reweighted Sampling (ARS) technology to select wavelength points exhibiting large absolute regression coefficients in Partial Least Squares (PLS) models while eliminating low-weight wavelength points. Through cross-validation, the method identifies subsets with minimal Root Mean Square Error of Cross-Validation (RMSECV) values, enabling efficient discovery of optimal variable combinations. Key implementation aspects include: 1) Weight calculation based on regression coefficient magnitudes 2) Exponential decay function for feature elimination 3) Cross-validation loops for subset evaluation. These techniques significantly enhance model accuracy and stability, particularly when handling large-scale datasets. Thus, CARS represents a highly effective feature variable extraction methodology for spectral data analysis and multivariate calibration.