Local classification: Locally weighted-partial least squares-discriminant analysis (LW-PLS-DA)

Area of Science:

• Chemometrics
• Machine Learning
• Data Mining

Background:

• Traditional methods struggle with non-linear classification tasks.
• Partial Least Squares-Discriminant Analysis (PLS-DA) is effective but often limited to linear relationships.
• Developing flexible and accurate non-linear classification algorithms is crucial for complex datasets.

Purpose of the Study:

• To extend the locally weighted partial least squares (LW-PLS) approach for non-linear classification.
• To introduce a discriminant version of LW-PLS, termed locally weighted-partial least squares-discriminant analysis (LW-PLS-DA).
• To evaluate the performance of LW-PLS-DA against existing non-linear classification methods.

Main Methods:

• The proposed LW-PLS-DA algorithm identifies nearest training samples to predict unknown samples.
• A PLS-DA model is built using only these selected calibration samples.
• A non-uniform, distance-based weighting scheme modulates the influence of training samples.

Main Results:

• LW-PLS-DA achieved over 99% classification accuracy on three simulated non-linear datasets.
• On a real-world dataset (rice variety classification), LW-PLS-DA yielded an average correct classification rate of 93%.
• Performance was comparable or superior to k-nearest neighbors, kernel-PLS-DA, and counterpropagation neural networks.

Conclusions:

• LW-PLS-DA is a simple, flexible, and accurate method for non-linear classification.
• The algorithm effectively handles datasets with high degrees of non-linearity.
• LW-PLS-DA demonstrates strong potential as an alternative to existing non-linear classification techniques.