Robust regression using support vector regressions

Noisy data and outliers has always been one of the main challenges in regression applications. The presence of these data among training data will produce several negative effects on the generalization ability of the built model in regression. Filtering noisy data and outliers as a preprocessing step in pattern recognition has many challenges in practice because it is really hard to distinguish noisy samples from the remaining ones. Hence, robustness of an algorithm is defined as its capability to train the model in such way that it suffers less from noisy data as well as outliers. In this study, we improve the performance of support vector regression (SVR) to be more robust. SVR is sensitive to noisy data because its trained model is built by support vectors which contain a small portion of training data. The constraints inequalities in the constraints of ε-insensitive SVR are changed to fuzzy inequalities without any changes in its loss function. This gives more flexibility to the SVR constraints for satisfaction. Then, we solve the quadratic programming (QP) problem and compare the new method with standard support vector regression model. Experimental results with different data sets show the superiority of the proposed method especially in presence of noisy data and outliers.