Global-local least-squares support vector machine (GLocal-LS-SVM).

Area of Science:

• Machine Learning
• Computational Science
• Data Science

Background:

• Decentralized data sources and large datasets present significant computational challenges for traditional machine learning algorithms.
• Input-space complexities can hinder the performance and scalability of existing models.
• Least-squares support vector machine (LS-SVM) models struggle with efficiency on extensive and distributed datasets.

Purpose of the Study:

• To introduce a novel machine learning algorithm, the global-local least-squares support vector machine (GLocal-LS-SVM).
• To address challenges posed by decentralized data, large datasets, and input-space issues.
• To enhance computational efficiency while maintaining high classification performance.

Main Methods:

• A double-layer learning approach combining multiple local LS-SVM models and a global LS-SVM model.
• Identification and extraction of informative data points (support vectors) from local input space regions.
• Merging local support vectors to create a reduced training set for the global model.

Main Results:

• GLocal-LS-SVM demonstrated comparable or superior classification performance against standard LS-SVM and other state-of-the-art models.
• Significant improvements in computational efficiency were observed; training time was reduced to 2% of standard LS-SVM for a 9,000-instance dataset.
• The algorithm effectively handles large datasets and decentralized data sources.

Conclusions:

• GLocal-LS-SVM offers a robust solution for efficiently processing large and decentralized datasets.
• The algorithm maintains high classification accuracy, making it suitable for practical applications.
• Its computational efficiency positions it as a valuable tool across various domains requiring machine learning analysis.