Support vector weighted fuzzy regression

Abstract

Based on the idea of Support Vector Machine (SVM) methodology, a new robust support vector linear regression modelling known as Support Vector Weighted Fuzzy Regression (SVWFR) is introduced, for the case when the values of response variable are fuzzy rather than crisp. The extension of the proposed method to the nonlinear case is also investigated. In the proposed approach, a weighted operation is employed to improve the robustness of usual support vector fuzzy regression models by assigning weights to the support hyperplanes constraints. While the fuzzy machine learning-based models are typically sensitive to outliers, the advantages of the proposed models are their robustness with respect to outlier data. The efficiency and applicability of the proposed models are investigated by using three data sets: a synthetic dataset including outliers, a textile engineering data set, and a stress-test simulation with artificially introduced anomalies. Across all cases, the introduced models consistently outperformed current fuzzy regression also approaches, based on three well-known goodness of fit indices. Sensitivity analysis of nonlinear SVWFR parameters is examined.