A Deep Non-negative Matrix Factorization Model for Big Data Representation Learning.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Deep representations offer high performance but lack interpretability, hindering real-world applications.
• Interpretability of deep learning models is a significant challenge in big data analysis.

Purpose of the Study:

• To propose a novel deep matrix factorization method for learning interpretable, part-based deep representations.
• To address the challenge of interpretability in deep representations for big data applications.

Main Methods:

• A deep architecture with supervisor and student networks was designed for an end-to-end pattern mining framework.
• Non-negative constraints and a specialized interpretability loss (symmetric, apposition, non-negative constraint loss) were employed for training.
• The method facilitates knowledge transfer from the supervisor to the student network, improving representation robustness.

Main Results:

• The proposed deep matrix factorization method successfully learns interpretable, part-based deep representations.
• Experimental results on two benchmark datasets confirmed the method's superiority over existing approaches.
• The integrated interpretability loss enhanced the robustness of the learned deep representations.

Conclusions:

• The developed deep matrix factorization method effectively enhances the interpretability of deep representations.
• This approach offers a promising solution for pattern mining and big data analysis where interpretability is crucial.
• The findings highlight the potential of deep matrix factorization for robust and understandable AI models.