A weighting initialization strategy for weighted support vector machines

Abstract

This paper presents a problem independent weighting strategy for weighted support vector machines (SVMs). SVMs can be applied with a weighting to each training vector to reflect the importance of different classes or training samples. Weightings are often assigned to the two classes inversely proportional to the sample count of each class, or according to a priori knowledge. Such a strategy can be applied to skewed data sets to balance the importance, error contribution and cost between the two classes. In this paper we propose a strategy to give each training pattern a weighting according to their distances to the classifier. The strategy regards the importance of the training patterns to the training process but not the importance of the data to the problem, thus it is suitable for general SVM applications. Experiments show that the performance of the proposed method is competitive to standard SVM while the training processes are even sped up. © Springer-Verlag Berlin Heidelberg 2005.