Robust and efficient learning with granular ball support vector regression.

Area of Science:

• Machine Learning
• Data Mining
• Computational Statistics

Background:

• Support Vector Regression (SVR) is a popular regression technique but suffers from high computational costs due to quadratic programming.
• The standard SVR epsilon-insensitive loss function is sensitive to noise and outliers, potentially degrading performance.
• Existing SVR methods face limitations in handling large datasets and noisy real-world applications.

Purpose of the Study:

• To introduce Granular Ball Support Vector Regression (GBSVR) as an efficient and robust regression method.
• To leverage the granular ball concept for simplifying complex data spaces in regression tasks.
• To address the computational expense and noise sensitivity issues inherent in traditional SVR.

Main Methods:

• Developed Granular Ball Support Vector Regression (GBSVR) using the granular ball concept.
• Implemented a discretization method for continuous attributes to enable granular ball construction.
• Grouped data points into granular balls based on feature proximity for a coarse, noise-tolerant representation.

Main Results:

• GBSVR significantly reduces computational cost by replacing numerous data points with fewer granular balls.
• The granular ball approach provides a noise-tolerant data representation, enhancing regression stability.
• Evaluations on benchmark datasets demonstrate that GBSVR outperforms existing state-of-the-art regression approaches.

Conclusions:

• GBSVR presents a computationally efficient and noise-tolerant solution for regression problems.
• The granular ball concept effectively simplifies data representation for improved machine learning performance.
• The proposed method shows superior performance compared to current state-of-the-art techniques, with open-source code available.