Pair- ${v}$ -SVR: A Novel and Efficient Pairing nu-Support Vector Regression Algorithm.

Area of Science:

• Machine Learning
• Statistical Learning Theory
• Regression Analysis

Background:

• Classical nu-support vector regression (-SVR) involves solving a single large quadratic programming problem (QPP).
• Twin support vector regression (TSVR) improves learning speed by solving two smaller QPPs but can be further optimized.
• Accurate modeling of conditional mean and predictive variance is crucial, especially with heteroscedastic noise.

Purpose of the Study:

• To propose a novel and efficient pairing nu-support vector regression (pair-SVR) algorithm.
• To enhance prediction speed and generalization ability compared to existing TSVR and -SVR methods.
• To facilitate automatic estimation of conditional mean and predictive variance.

Main Methods:

• Developed a pair-SVR algorithm integrating advantages of TSVR and -SVR.
• Implemented two smaller QPPs for faster learning, inspired by TSVR.
• Introduced an insensitive zone and regularization term for improved prediction and generalization.
• Utilized a parameter to control fractions of support vectors (SVs) and errors.
• Estimated upper and lower bound functions to capture data distribution characteristics.

Main Results:

• The pair-SVR algorithm demonstrates significantly faster training and prediction speeds.
• Improved generalization ability is achieved compared to TSVR and classical -SVR.
• The method effectively captures data distribution characteristics for simultaneous mean and variance estimation.
• Experimental results validate the proposed algorithm's superiority.

Conclusions:

• The proposed pair-SVR algorithm offers a superior alternative for regression tasks.
• It efficiently combines the strengths of TSVR and -SVR, enhancing performance metrics.
• The algorithm's ability to handle heteroscedastic noise makes it valuable for complex datasets.