Support vector one-class classification for multiple-distribution data

One-class support vector algorithms such as One-Class Support Vector Machine (OCSVM) and Support Vector Data Description (SVDD) often perform poorly with multi-distributed data. Because in the one-class classification context, only the target class is well represented, the classification problem is ill-posed and the task is more a class description or a class density estimation problem. To deal with multi-distributed data, we propose in this paper the Multi-Cluster One-Class Support Vector Machine (MCOS) algorithm, which first clusters the data and then applies a one-class support vector algorithm on each cluster separately. A test sample is then classified by using the corresponding local description. K-means clustering and a dendogram based clustering methods are tested and classification results are presented for synthetic and real world data by using the MCOS. Experiments show that in many cases, MCOS outperforms the OCSVM algorithm.