MDSR-NMF: Multiple deconstruction single reconstruction deep neural network model for non-negative matrix...

Area of Science:

• Machine Learning
• Data Science
• Deep Learning

Background:

• High-dimensional datasets pose challenges for analysis.
• Dimension reduction is crucial for managing large data.
• Existing methods may lack robustness or efficiency.

Purpose of the Study:

• To introduce a novel deep learning architecture for non-negative matrix factorization (NMF).
• To achieve robust low-rank approximation of high-dimensional data.
• To enhance classification and clustering performance.

Main Methods:

• Developed a deep learning architecture with multiple deconstruction and single reconstruction layers.
• Employed a two-stage approach: pretraining and stacking.
• Modified the sigmoid function for non-negativity and reduced data loss.
• Utilized Xavier initialization to address gradient issues.
• Incorporated a regularizer in the objective function for optimal matrix approximation.

Main Results:

• The proposed MDSR-NMF model demonstrated superior performance compared to six established dimension reduction methods.
• Effectiveness validated across five diverse datasets for classification and clustering tasks.
• Analysis confirmed the model's computational efficiency and convergence properties.

Conclusions:

• MDSR-NMF provides a robust and effective solution for dimension reduction in high-dimensional datasets.
• The novel deep learning architecture offers advantages in low-rank approximation and data analysis tasks.
• The model's performance and efficiency are well-established through empirical evaluation.