Multiobjective Sparse Non-Negative Matrix Factorization.

Area of Science:

• Data Science
• Machine Learning
• Computational Science

Background:

• Non-negative Matrix Factorization (NMF) is widely used for its interpretability and part-based representation.
• Sparse constraints are often applied to NMF for enhanced feature extraction, typically via regularization models.
• Existing regularization methods face challenges with parameter tuning, bias effects, and the Restricted Isometry Property (RIP).

Purpose of the Study:

• To propose a novel multiobjective sparse NMF paradigm that avoids regularization parameter issues, bias effects, and RIP conditions.
• To develop a multiobjective memetic algorithm for generating diverse solutions with high factorization accuracy.
• To integrate prior knowledge for computational efficiency.

Main Methods:

• A multiobjective sparse Non-negative Matrix Factorization (NMF) paradigm is presented.
• A novel multiobjective memetic algorithm is developed, incorporating a masked projected gradient local search.
• A priori knowledge is integrated to optimize the search within the objective space.

Main Results:

• The proposed NMF paradigm effectively generates solutions with varying sparsity levels and high factorization accuracy.
• Experimental results demonstrate superior performance compared to traditional regularization algorithms.
• The method successfully addresses limitations associated with parameter tuning and bias effects in sparse NMF.

Conclusions:

• The novel multiobjective sparse NMF paradigm offers a robust alternative to existing methods.
• The developed algorithm provides a set of diverse, accurate solutions favorable for decision-making.
• This approach enhances the practical applicability of NMF in various research fields.