Outlier resistent approach to linear support vector machines

Support vector machines are widely used for the classification of binary response variables. In this paper, a loss function is introduced that enables support vector machines to be more resistant against outliers. Although outliers may contain a lot of information, it is generally not desirable that these observations play a major role in determining the binary classification. The novel error function that is being introduced, the absolute outlier resistant hinge error, restricts the distorting effect of outliers by treating them differently. It attributes a decreasing marginal impact on the loss function to a support vector as it becomes more deviant. Support vector machines that make use of the novel hinge error have been applied to multiple data sets, including ones that are contaminated with artificial outliers. The experiments show signs of an increased resistance against outliers.