Non-stationary data domain description using weighted support vector novelty detector

Even though most of the classification methods deal with multiple classes, there is an objective need for classification methods that deal with a single class. This is particularly true when it is difficult or expensive to find examples for other classes. One-class classification (also called data domain description) is often used for outlier or novelty detection. These methods allow representation of the behavior of a system with few parameters compared to the number data points collected from the system. Methods with probability density assumptions have the weakness of applicability to real world applications. Very few one-class classification methods can handle non-stationary data. To the best of our knowledge, there exists no method that can handle non-stationary data without making stringent assumptions about the data distribution. This work proposes a data domain description method based on support vector machine principles for stationary as well as non-stationary data. Results from testing the proposed methods on several different datasets are very promising.