Collaborative Neurodynamic Approaches to Data Processing

Abstract

In the modern digital era, data has emerged as a pivotal asset for industries, driving innovation and improving operational efficiency. Data processing is the task of converting raw data into meaningful information. Among various data processing techniques, clustering and matrix factorization are critical tools that uncover hidden patterns and simplify complex datasets. Nevertheless, these methods encounter numerous challenges, including the selection of optimal parameters, ensuring algorithm scalability, and achieving accurate and reliable results on complex datasets. These challenges directly affect the effectiveness and efficiency of data processing, necessitating the development of advanced methods.

Collaborative neurodynamic optimization (CNO) emerges as a powerful hybrid intelligence framework for global optimization problems by employing multiple neurodynamic models computing in parallel for searching the global optima. It has been shown to be well-suited for various computational tasks. Given these advances, it is both beneficial and desirable to develop and apply CNO-driven methods in data processing to address the aforementioned challenges.

This thesis focuses on developing collaborative annealing methods for clustering and CNO-driven algorithms for binary matrix factorization, organized into four parts. The first part introduces collaborative annealing clustering algorithms based on power k-means and fuzzy c-means clustering, respectively. The second part presents a CNO algorithm for quadratic unconstrained binary optimization problems. The third part proposes an algorithm for capacitated clustering based on majorization-minimization and collaborative neurodynamic optimization. The final part develops a CNO-driven algorithm for binary matrix factorization, a neurodynamics-driven constrained matrix factorization approach to machine-cell and part-family formulation, and a neurodynamics-driven binary matrix factorization approach for bi-clustering binary data. The efficacy of these methods is demonstrated and substantiated through extensive experimental results, highlighting their superiority in handling complex data processing tasks compared to traditional approaches.

Date of Award	27 Dec 2024
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Jun WANG (Supervisor), Qingfu ZHANG (Co-supervisor) & Kay Chen TAN (Co-supervisor)