Discriminant locality preserving projections based on L1-norm maximization.

Area of Science:

• Computer Science
• Machine Learning
• Pattern Recognition

Background:

• Conventional discriminant locality preserving projection (DLPP) is a manifold learning technique for dimensionality reduction.
• The L2-norm basis of conventional DLPP makes it sensitive to outliers, limiting its real-world applicability.
• Outliers and small sample sizes are common challenges in pattern recognition tasks.

Purpose of the Study:

• To develop an effective and robust version of DLPP that overcomes the limitations of conventional methods.
• To enhance the performance of dimensionality reduction in the presence of outliers and small sample sizes.
• To improve the accuracy and reliability of pattern recognition systems.

Main Methods:

• Proposes a novel DLPP method utilizing L1-norm maximization.
• Learns local optimal projection vectors by maximizing the ratio of L1-norm-based between-class dispersion to within-class dispersion.
• Employs manifold learning principles adapted for robust feature extraction.

Main Results:

• The proposed L1-norm DLPP method demonstrates robustness against outliers.
• The method effectively addresses the small sample size problem in dimensionality reduction.
• Experimental validation on artificial, image, and biometric datasets confirms the method's effectiveness.

Conclusions:

• The L1-norm DLPP offers a significant improvement over conventional DLPP for pattern recognition.
• This robust approach enhances the applicability of DLPP in real-world scenarios with noisy data.
• The method provides a reliable dimensionality reduction technique for various machine learning applications.