DMT-EV: An Explainable Deep Network for Dimension Reduction.

Area of Science:

• Data Science
• Machine Learning
• Computational Biology

Background:

• Dimension reduction (DR) is vital for analyzing high-dimensional data in fields like image recognition and single-cell sequencing.
• Current DR methods struggle with preserving global/local features and generalizability, lacking explainability.
• Understanding feature contributions in DR is crucial for diagnosing embedding processes and identifying key components.

Purpose of the Study:

• To introduce DMT-EV, a deep neural network method for dimension reduction.
• To enhance both the performance (structural maintainability) and explainability of DR.
• To provide an interactive visual interface for user-guided DR optimization.

Main Methods:

• DMT-EV utilizes data augmentation and a manifold-based loss function for improved embedding performance.
• Explainability is achieved through saliency maps to analyze parameter contributions during embedding.
• An integrated visual interface allows users to adjust parameters for better DR and explainability.

Main Results:

• DMT-EV demonstrates superior performance in preserving data structure compared to state-of-the-art methods.
• The saliency map-based approach provides clear insights into the DR embedding process.
• Experimental comparisons confirm DMT-EV's effectiveness in both performance and explainability.

Conclusions:

• DMT-EV offers a significant advancement in dimension reduction by combining high performance with crucial explainability.
• The method addresses limitations of existing DR techniques, particularly in feature preservation and interpretability.
• The interactive visual interface facilitates user engagement and optimization of DR outcomes.