Validity of training samples is important to regression model

however

outliers in training sample sets are hard to avoid

which will disturb the model accuracy.In order to overcome the influence of outliers to regression model

a new M-estimator based support vector machine (M-SVM) was introduced. It replaced the sum of square residuals in least squares support vector machine (LS-SVM) with objective function of M-estimator.An iterative modeling algorithm of the M-SVM was proposed.The M-SVM was used in a low-dimension data regression problem with outliers and in the NIR spectral analysis of gasoline research octane number.Experimental results showed that the M-SVM was more robust and accurate than other SVMs.