Spectral data classification using locally weighted partial least squares classifier

Partial least squares discriminant analysis (PLS-DA) is an effective chemometric method for handling ill-conditioned problems in data matrices, such as small-sample-size, high dimensionality and high collinearity. Although PLS-DA has been widely used in the classification of spectral data, it is often confronted with performance degradation when physical and chemical properties of a testing object have complex effects on spectra, such as detector-based and chemical-based nonlinearity. Locally weighted partial least squares (LW-PLS) is a variant of PLS for regression to address nonlinearity in data. It utilizes the Euclidean distance based similarity to weight training samples and then constructs local PLS models for prediction. However, using LW-PLS for classification is still blank and its classification performance has yet to be reported. In this paper, we extend LW-PLS for the classification of spectral data, resulting LW-PLSC. Experimental results on ten UCI benchmark and two spectral datasets show that LW-PLSC can outperform five baseline methods, achieving the highest classification accuracies most of the time.