Deep Recursive Embedding for High-Dimensional Data.

Area of Science:

• Machine Learning
• Data Science
• Dimensionality Reduction

Background:

• High-dimensional data embedding is crucial for data analysis and visualization.
• Existing methods like t-SNE and UMAP have limitations in scalability and out-of-sample mapping.
• Deep neural networks offer powerful tools for learning complex data representations.

Purpose of the Study:

• To develop a novel, scalable, and effective method for embedding high-dimensional data into a low-dimensional manifold.
• To leverage deep neural networks and mathematics-guided principles for improved embedding performance.
• To enable out-of-sample data mapping and handle extremely large datasets.

Main Methods:

• Proposing a generic Deep Embedding Network (DEN) framework.
• Introducing a Deep Recursive Embedding (DRE) strategy utilizing latent representations.
• Employing Kullback-Leibler (KL) divergence minimization as a guiding objective.
• Benchmarking DRE against t-distributed stochastic neighbor embedding (t-SNE) and Uniform Manifold Approximation and Projection (UMAP).

Main Results:

• DRE demonstrates superior performance in preserving both local and global data structures compared to state-of-the-art methods.
• The proposed method effectively maps out-of-sample data points.
• DRE exhibits excellent scalability for extremely large datasets.
• Experiments on public datasets validate the improved embedding quality.

Conclusions:

• Deep Recursive Embedding (DRE) offers a robust and efficient solution for high-dimensional data embedding.
• The DRE framework provides flexibility through various architectures and loss functions.
• This method advances the field of dimensionality reduction with improved performance and scalability.