LDLE: Low Distortion Local Eigenmaps.

Area of Science:

• Computational mathematics
• Data science
• Machine learning

Background:

• Manifold learning aims to represent high-dimensional data in a lower-dimensional space.
• Existing methods often struggle with preserving local distances and embedding complex manifold structures.

Purpose of the Study:

• To introduce Low Distortion Local Eigenmaps (LDLE), a new manifold learning technique.
• To address limitations of current methods in embedding complex manifolds and preserving distances.

Main Methods:

• Constructing local data views using global eigenvectors of the graph Laplacian.
• Registering local views via Procrustes analysis to achieve a global embedding.
• Enabling embedding of closed and non-orientable manifolds by allowing controlled tearing.

Main Results:

• LDLE demonstrates significantly lower distortion compared to existing techniques.
• The method effectively preserves distances up to a constant scale.
• High-quality embeddings are achieved even with noisy or sparse datasets.

Conclusions:

• LDLE offers a robust approach to manifold learning with superior distance preservation.
• The technique can handle complex topological structures, including closed and non-orientable manifolds.
• LDLE provides a valuable tool for data visualization and analysis in challenging scenarios.