Sliced Inverse Regression With Adaptive Spectral Sparsity for Dimension Reduction

Area of Science:

• Machine Learning
• Pattern Analysis
• Data Science

Background:

• Dimension reduction is crucial for feature representation and pattern classification.
• Sliced Inverse Regression (SIR) is effective but suffers from high computational complexity and underutilized sparsity for high-dimensional data.
• Existing SIR methods lack robust feature selection and model interpretability.

Purpose of the Study:

• To develop computationally efficient and interpretable dimension reduction techniques.
• To improve feature selection and model interpretability in high-dimensional data analysis.
• To enhance robust pattern classification under data corruption.

Main Methods:

• Computing SIR projection vectors in spectral space for faster approximated regression.
• Employing adaptive lasso for sparse and globally optimal solutions in variable selection.
• Designing a correntropy-based, class-wise regression model for robust pattern classification with smooth penalties.

Main Results:

• The proposed spectral SIR approach achieves faster computation speeds.
• Adaptive lasso effectively yields sparse solutions for improved feature selection.
• The correntropy-based class-wise regression demonstrates robustness in pattern classification tasks.
• Experimental results on facial images and gene microarray data show competitive performance.

Conclusions:

• The novel spectral SIR and adaptive lasso methods offer significant improvements in speed and interpretability.
• The correntropy-based regression model provides a flexible and robust approach for pattern classification.
• These advancements address key limitations of traditional SIR, enabling wider applications in machine learning and data analysis.