An improved classification model based on covering algorithm and SVM

In order to overcome some shortages of SVM, an improved classification model is introduced in this paper. For the first problem about isolated points or noises mixed in training data sets which will cause overfitting problem and decrease the capability of generalization for SVM, we proposed modified covering algorithm to find out the isolated points and deal with it by the definition of covering sample density. As for the second problem, time cost for training SVM on large data sets usually is high; we introduce modified CA as the pre-classification step to reduce the training sample scale, by constructing a series of covers and deleting the isolated points, and then use the centroids of the rest covers as the new training data sets for SVM training. By the experiments on the real world data sets, results show the training time can drop significantly, and the accuracy is very close to Lib-SVM. So, CA-SVM is an efficient classification model.