Binary matrix factorization via collaborative neurodynamic optimization.

Area of Science:

• Computational science
• Data science
• Machine learning

Background:

• Binary matrix factorization (BMF) is crucial for reducing dimensions in high-dimensional binary datasets.
• Existing BMF methods face challenges with local optima and factorization error.

Purpose of the Study:

• To present a collaborative neurodynamic optimization approach for BMF.
• To improve factorization accuracy and enable effective pattern discovery.

Main Methods:

• Utilizes multiple discrete Hopfield networks for concurrent local optima search.
• Incorporates particle swarm optimization to escape local minima and refine solutions.
• Reformulates BMF using combinatorial and quadratic unconstrained binary optimization.

Main Results:

• Demonstrates superior performance over six baseline algorithms on eight benchmark datasets.
• Achieves lower factorization error compared to existing methods.
• Successfully applied to pattern discovery on three distinct datasets.

Conclusions:

• The proposed neurodynamic approach offers a powerful and effective method for BMF.
• This technique enhances dimension reduction and pattern discovery capabilities for binary data.