Robust Supervised Spline Embedding.

Area of Science:

• Machine Learning
• Data Science
• Computational Statistics

Background:

• High-dimensional data presents challenges like noise and the curse of dimensionality, limiting traditional classification algorithms.
• Existing methods often fail to preserve both manifold structure and discriminative information for valid inference.

Purpose of the Study:

• To develop a robust supervised spline embedding (RS2E) algorithm for accurate high-dimensional classification.
• To address limitations of current methods by preserving class-aware submanifolds and eliminating noise.

Main Methods:

• The RS2E algorithm preserves class-aware submanifold structure in a thin plate spline embedding space.
• It eliminates noise and outliers by exploiting intrinsic low complexity to recover clean manifolds.
• Class-aware submanifolds are separated by maximizing distances between data points and marginal points of other classes.
• The objective function is solved using the alternating direction method of multipliers with generalized power iteration.

Main Results:

• RS2E demonstrates superior performance in classification accuracy compared to other supervised dimensionality reduction algorithms.
• Experiments were conducted on real-world, GAN-generated, and artificially corrupted datasets.

Conclusions:

• The RS2E algorithm offers a robust solution for high-dimensional classification tasks.
• It effectively handles noise and preserves essential data structures, leading to enhanced classification accuracy.