Successive overrelaxation for laplacian support vector machine.

Area of Science:

• Machine Learning
• Data Mining
• Computational Science

Background:

• Semisupervised learning (SSL) utilizes both labeled and unlabeled data.
• Laplacian Support Vector Machines (LapSVMs) are effective but computationally intensive.
• There is a need for more efficient SSL algorithms for large datasets.

Purpose of the Study:

• To propose a Fast Laplacian Support Vector Machine (FLapSVM) solver.
• To improve the computational efficiency of LapSVMs for classification tasks.
• To enhance scalability for large-scale machine learning problems.

Main Methods:

• Developed FLapSVM by optimizing the LapSVM formulation.
• Avoided complex matrix operations and variable switching inherent in standard LapSVM.
• Employed successive overrelaxation (SOR) technology for linear convergence.
• Integrated random subproblem scheduling and dual stopping conditions.

Main Results:

• FLapSVM demonstrates significantly faster computation speeds compared to standard LapSVM.
• The method is more suitable for large-scale problems due to reduced computational burden.
• The dual problem formulation allows direct application of the kernel trick.
• FLapSVM effectively processes large datasets that may not fit into memory.

Conclusions:

• FLapSVM offers a computationally efficient and scalable alternative for semisupervised classification.
• The proposed method maintains the advantages of LapSVM while overcoming its computational limitations.
• FLapSVM is a practical and effective solution for real-world large-scale SSL applications.