An Evolutionary Computation Approach for Optimizing Support Vector Regression in Predicting Pollen Levels

Tiberiu Alexandru MIHAI, Smaranda BELCIUG

Abstract

Support Vector Regression (SVR) is a machine learning technique widely used for modeling complex and non-linear data. Its performance is highly sensitive to the choice of the kernel and hyperparameters. This study explores the use of an evolutionary computation approach for optimizing the SVR parameters, namely a systematic framework that balances predictive accuracy, computational efficiency, and generalization. The proposed methodology optimizes the key SVR parameters - including the regularization constant (C), the kernel coefficient (γ), and the ε-insensitive loss parameter - while simultaneously exploring different kernel families. Unlike the conventional grid search optimization, the evolutionary algorithm dynamically refines the candidate solutions through selection, crossover, and mutation. In order to evaluate this method, it is applied to forecasting airborne pollen counts, a task which is complicated by variability, seasonality, and the non-linear dynamics of aeroallergen data. The experimental results demonstrate that kernel adaptation through evolutionary computation leads to a consistently improved accuracy and efficiency. Overall, this approach enables the construction of SVR models that are both more accurate and computationally efficient than those produced by using traditional optimization methods.

Keywords

Support vector regression, Evolutionary computation, Statistical analysis, Airborne pollen counts.

View full article