A novel relative density based support vector machine

Relative density based SVM does not use any kernel to obtain the points near the optimal decision plane. It can be used to detect and eliminate classification noise so that cross validation is not necessary to be used. However, it relies on a search tree to find nearest neighbors to maintain a low time complexity. High dimensionality will lead to increase of complication of structure of the tree and the time complexity. Thus, the performance of relative density based SVM deteriorates greatly in high dimensional data. In this paper, the concept of “location difference of multiple distances” is introduced to improve the performance of relative density based SVM. The proposed algorithm has a good performance in prediction accuracy. Furthermore, it does not use any tree structure so that it has a much better efficiency in high dimensional data and stability than the previous algorithms.