Numerically stable locality-preserving partial least squares discriminant analysis for efficient dimensionality...

Area of Science:

• Machine Learning
• Data Science
• Pattern Recognition

Background:

• Dimensionality reduction is crucial for high-dimensional data classification.
• Numerical stability of algorithms directly impacts classification accuracy.
• High-dimensional data often leads to scatter matrix singularity.

Purpose of the Study:

• To investigate numerical attributes of dimensionality reduction and discriminant subspace learning.
• To propose robust implementations of Locality-Preserving Partial Least Squares Discriminant Analysis (LPPLS-DA).
• To enhance class separability and classification accuracy.

Main Methods:

• Exploration of two robust LPPLS-DA implementations.
• Optimization of data projections for improved feature discrimination.
• Numerical experiments on synthetic and spectral datasets.

Main Results:

• Proposed LPPLS-DA methods demonstrated improved classification accuracy.
• Enhanced feature discrimination and optimized data projections were achieved.
• Superior performance compared to state-of-the-art dimensionality reduction techniques.

Conclusions:

• Robust LPPLS-DA implementations effectively address singularity issues in high-dimensional data.
• The proposed methods offer significant improvements in classification and dimension reduction.
• These findings highlight the importance of numerical stability in discriminant analysis.