Bounded influence support vector regression for robust single-model estimation.

Area of Science:

• Machine Learning
• Statistical Modeling
• Data Science

Background:

• Support Vector Regression (SVR) is a standard technique for real-valued function estimation.
• Traditional SVR methods struggle with datasets containing significant outliers in predictor or response variables.
• Outlier contamination is a common challenge in real-world data analysis across various scientific domains.

Purpose of the Study:

• To develop an improved SVR algorithm capable of robustly handling severe outlier contamination.
• To enhance regression accuracy by downweighting the influence of outliers in both predictor and response variables.
• To create a method that can effectively identify and utilize the dominant subset of data in corrupted datasets.

Main Methods:

• Proposed a bounded influence Support Vector Regression (SVR) approach.
• Implemented an adaptive weighting strategy incorporating a robust scale estimator for residuals.
• Utilized a kernelized hat matrix statistic for identifying and mitigating leverage points.

Main Results:

• Simulated datasets demonstrated the algorithm's robustness against outliers in both linear and nonlinear scenarios.
• The bounded influence SVR successfully downweighted outliers, improving model performance.
• Real-world chemical and astronomical datasets with severe outlier contamination showed the practical effectiveness of the proposed method.

Conclusions:

• The developed bounded influence SVR is a robust alternative to standard SVR for contaminated datasets.
• The adaptive weighting strategy effectively addresses outliers, leading to more reliable regression estimates.
• This approach offers improved performance in practical applications with significant data corruption.