Multi-output parameter-insensitive kernel twin SVR model.

Area of Science:

• Machine Learning
• Computational Statistics

Background:

• Multi-output regression maps complex input spaces to multiple outputs.
• Existing methods like single-output SVR combinations neglect output correlations, while others suffer from high computational cost and parameter sensitivity due to noise.
• Addressing these limitations is crucial for advancing multivariate data analysis.

Purpose of the Study:

• To develop novel, efficient, and robust multi-output regression techniques.
• To introduce methods that overcome the drawbacks of existing multi-output support vector regression approaches.
• To enhance prediction accuracy and stability in multivariate regression tasks.

Main Methods:

• Developed multi-output twin support vector regression (M-TSVR) using two smaller quadratic programming problems for fast learning.
• Introduced multi-output parameter-insensitive twin support vector regression (M-PITSVR) for heteroscedastic noise, employing parameter-insensitive functions.
• Derived kernelized extensions for both M-TSVR and M-PITSVR to handle nonlinear relationships.

Main Results:

• Comparative experiments on twelve datasets demonstrated M-TSVR and M-PITSVR's fast learning speeds.
• The proposed methods achieved better and more stable prediction performance than existing multi-output regression techniques.
• Kernelized versions effectively addressed nonlinear multi-output regression challenges.

Conclusions:

• M-TSVR and M-PITSVR offer significant improvements in speed and performance for multi-output regression.
• These novel methods provide robust solutions for handling output correlations and noise in multivariate analysis.
• The developed techniques represent a valuable advancement in the field of machine learning for complex regression problems.