Subconcept Based One Class Classification Method with Cluster Updating

One class classification is an effective way to perform anomaly detection or outlier detection. Previous work has empirically shown that complexity in the background (also known as target) class degrades the performance of one-class classifiers, and one source of data complexity is the presence of subconcepts in that class. Learning over subconcepts individually can mitigate the effects of domain complexity and improve one class classification performance. Unless a clustering could be derived from domain knowledge, the approach used to search for subconcepts is unsupervised clustering. However, in some cases, the examples belonging to some of the clusters may be too scattered, while, in others, clusters may be affected by the small disjunct problem where some clusters comprise only of a few examples that cannot be used to train a robust classifier. To handle these problems, this paper presents a method to update the clusters obtained by the clustering process prior to learning over them. In addition, it introduces the c-Nearest-Cluster (c-NC) method for combining the individual classifiers derived from each subconcept. Our experiments on classical outlier detection datasets and on cyber security datasets show that our method can improve upon the classification performance obtained by the recently proposed subconcept based one class classification method.