Fast-Solving Quasi-Optimal LS-S³VM Based on an Extended Candidate Set.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Semisupervised learning is crucial as most real-world data lacks labels.
• Existing methods for semisupervised least squares support vector machines (LS-S3VM) struggle with efficient and effective decision hyperplane derivation.
• The need for improved training algorithms for LS-S3VM is evident.

Purpose of the Study:

• To propose a fully weighted model for LS-S3VM and an efficient integer programming (IP) model for solving it.
• To develop a novel algorithm that addresses the challenge of deriving optimal decision hyperplanes in LS-S3VM.
• To enhance the training process by effectively integrating information from unlabeled data.

Main Methods:

• A fully weighted LS-S3VM model is proposed and transformed into a simple integer programming (IP) model.
• A new indicator is designed based on data distances to the hyperplane to guide label reversal during training.
• An extended candidate set of unlabeled data is constructed, and two strategies determine descent directions.
• A novel method for finding a good starting point is developed based on the IP model properties.

Main Results:

• A fast algorithm for quasi-optimal LS-S3VM solutions is presented, balancing computational cost and overfitting avoidance.
• Experimental results demonstrate the proposed algorithm's effectiveness compared to existing methods.
• The algorithm excels in computational complexity, generalization ability, and flexibility, with comparable performance in other aspects.

Conclusions:

• The proposed algorithm offers a quasi-optimal solution for LS-S3VM, achieving low computational cost and preventing overfitting.
• The integration of an extended candidate set and a carefully computed starting point leads to a fast and effective training method.
• The developed strategies enhance LS-S3VM performance across multiple key metrics, offering a significant advancement in semisupervised learning.