Anomaly Detection Research For Imbalanced Classes

Anomaly detection can be solved as class imbalanced problems which have attract much attention from pattern recognition and machine learning areas. Recently, in class imbalanced problems, researchers have made some significant progress, and applied the progress in such areas as spam filtering, network intrusion detection, text mining and so on.Support Vector Machines (SVM) is an excellent classifier, which has gained some impressive effects in such areas as pattern recognition and machine learning. SVM has also been applied in class imbalanced problems. However, because of its vulnerability as well as the speciality of class imbalanced problems, SVM can not solve theses problems very well. Asymmetric Support Vector Machine (ASVM) which incorporates the ideas of SVM and One-class SVM (OCSVM) can solve class imbalanced problems relatively well. However, ASVM ignores the importance of within-class structure information. Inspired by ASVM, we develop a Structured Asymmetric Support Vector Machine (StASVM) to maximize the class-margin as well as the margin between the origin and one of the classes. Meanwhile, StASVM directly embeds the within-class structure information into the ASVM and maximizes the within-class tightness to raise its performance. By this method, StASVM not only focuses on the between-class scatter, but also takes the within-class scatter information into account. Experiment results show that by incorporating more prior knowledge from data, StASVM has a better generation than ASVM.ASVM and StASVM have gained good performances in class imbalanced problems. However, they both take much training time. Incremental learning algorithms can improve the time efficiency significantly. However, incremental algorithms for imbalanced classification are still relatively rare. In this paper, based on the reduction of the imbalance ratio, we apply a simple incremental learning algorithm called as MinOver to imbalanced classification problem and achieve good performances in some UCI databases. Besides, by introducing the active learning strategy to our algorithm, we further propose an incremental asymmetric support vector machines based on active learning (IASVM), which significantly reduces the time consuming while still keeps the comparable performance to the Minover algorithm.