Manifoldron: Direct Space Partition via Manifold Discovery.

Area of Science:

• Machine Learning
• Computational Mathematics
• Data Science

Background:

• Neural networks (NNs) partition sample spaces using convex polytopes, but face challenges in interpretability and decision boundary flexibility.
• Parametric models risk shortcut solutions, while non-parameterized models often lack sufficient power or fail to capture data manifold structures.

Purpose of the Study:

• Introduce Manifoldron, a new machine learning model that derives decision boundaries directly from data via manifold structure discovery.
• Analyze Manifoldron's characteristics, including its manifold characterization capability and relationship to NNs.

Main Methods:

• Developed a novel machine learning model, Manifoldron, for data partitioning based on manifold structure discovery.
• Systematically analyzed Manifoldron's performance and characteristics against existing machine learning models.

Main Results:

• Manifoldron demonstrates competitive performance compared to mainstream machine learning models.
• Experimental results validated on synthetic, benchmark, and real-world datasets.

Conclusions:

• Manifoldron offers a promising alternative for data partitioning by leveraging manifold structures.
• The model addresses limitations of traditional parametric and non-parameterized machine learning approaches.