Nearest clusters based partial least squares discriminant analysis for the classification of spectral data.

Area of Science:

• Multivariate data analysis
• Chemometrics
• Machine learning for spectral data

Background:

• Partial Least Squares Discriminant Analysis (PLS-DA) is effective for high-dimensional, collinear spectral data.
• Standard PLS-DA struggles with within-class data multimodality and nonlinearity.
• Existing methods like kernel PLS-DA and local PLS-DA have limitations in handling complex data distributions.

Purpose of the Study:

• To introduce Nearest Clusters based PLS-DA (NCPLS-DA) to explicitly address data multimodality and nonlinearity.
• To improve the classification performance of PLS-DA on spectral datasets.
• To provide a robust alternative for complex spectral data analysis.

Main Methods:

• Hierarchical clustering is applied to group samples into distinct clusters.
• Cluster centers are calculated to represent local data distributions.
• PLS-DA is applied locally using only the nearest clusters to a query point.

Main Results:

• NCPLS-DA effectively separates multimodal and nonlinear classes into locally linear and unimodal clusters.
• Experimental results on 17 datasets (12 UCI, 5 spectral) demonstrate superior performance.
• NCPLS-DA achieved the highest classification accuracy compared to PLS-DA, kernel PLS-DA, local PLS-DA, and k-NN.

Conclusions:

• NCPLS-DA offers a significant advancement in spectral data classification by handling multimodality and nonlinearity.
• The method provides a simple yet effective tool for improving classification accuracy in complex datasets.
• NCPLS-DA demonstrates strong potential for applications in chemometrics and other fields utilizing spectral analysis.