Optimize the Large Scale SVM Social Spam Detection Model via Bistratal Reduction Method

Against the low efficiency of training on large-scale SVM, a reduction approach is proposed. This paper presents a new samples reduction method, called bistratal reduction method (BRM). BRM has two levels. The first level is coarse-grained reduction. It deletes the redundant clusters with KDC reduction. The second level is fine-grained reduction. It picks out the support vectors from the clusters remained and saves them as the final result set. In the traditional reduction method, the grain-size is single. There is still a contradiction between compression and accuracy, and can´t be solved perfectly. Bistratal reduction method changes the reduced intensity according to the number of redundant points remained. The experiments show that bistratal reduction method gives a higher compression ratio and accuracy. Apply the new method to the large scale SVM social spam detection model. The detection model speed up obviously.