Distributed support vector machine in master-slave mode.

Area of Science:

• Machine Learning
• Optimization Algorithms
• Distributed Computing

Background:

• Support Vector Machines (SVM) are effective learning algorithms.
• Alternating Direction Method of Multipliers (ADMM) is a powerful technique for distributed optimization.
• Existing distributed SVM methods can be improved for efficiency and convergence.

Purpose of the Study:

• To propose a novel distributed SVM algorithm using ADMM in a master-slave configuration.
• To enhance the convergence rate of distributed SVM training.
• To demonstrate the superiority of the proposed method in terms of convergence and accuracy.

Main Methods:

• Formulating distributed SVM as a regularized optimization problem.
• Solving convex optimization sub-problems using ADMM in a master-slave architecture.
• Employing an over-relaxation technique to accelerate convergence.

Main Results:

• The proposed Master-Slave Distributed SVM (MS-DSVM) algorithm achieves linear convergence.
• MS-DSVM exhibits the fastest convergence rate among standard distributed ADMM algorithms.
• Numerical examples confirm superior convergence and accuracy compared to existing ADMM-based methods.

Conclusions:

• The MS-DSVM algorithm effectively integrates distributed SVM and ADMM.
• The proposed method offers significant improvements in convergence speed and accuracy for distributed SVM.
• MS-DSVM represents an advancement in distributed machine learning optimization.