UDRN: Unified Dimensional Reduction Neural Network for feature selection and feature projection

Area of Science:

• Machine Learning
• Data Science
• Computational Biology

Background:

• Dimensional reduction (DR) techniques map high-dimensional data to a lower-dimensional latent space.
• Existing DR methods are categorized into feature selection (FS) and feature projection (FP), which are traditionally incompatible.
• FS risks data structure destruction, while FP lacks interpretability and sparsity.

Purpose of the Study:

• To propose a unified framework, the Unified Dimensional Reduction Network (UDRN), integrating FS and FP.
• To develop an end-to-end manifold learning framework that performs feature discovery and preserves data structure.
• To enhance the generalization ability of DR models using data augmentation priors.

Main Methods:

• Developed a novel network framework implementing FS and FP tasks separately via stacked networks.
• Introduced a stronger manifold assumption and a novel loss function.
• Leveraged data augmentation priors within the loss function to improve generalization.

Main Results:

• Demonstrated the advantages of UDRN over existing FS, FP, and FS&FP pipeline methods.
• Achieved superior performance in downstream tasks like classification and visualization.
• Validated UDRN's effectiveness on diverse datasets, including high-dimensional image and biological data.

Conclusions:

• The proposed UDRN effectively unifies feature selection and feature projection in an end-to-end manifold learning framework.
• UDRN offers improved performance in downstream tasks compared to traditional DR methods.
• The framework shows significant potential for analyzing complex, high-dimensional datasets.