Novel Class Detection Using Hybrid Ensemble

Abstract

In this research, we propose a hybrid meta-classifier for novel class detection. It is able to efficiently detect the arrival of novel unseen classes as well as tackle real-time data stream classification. Specifically, the proposed hybrid meta-classifler includes three ensemble models, i.e. class-specific, cluster-specific and complementary boosting ensemble classifiers. Distinctive training strategies are also proposed for the generation of effective and diversified ensemble classifiers. The weights of the above ensemble models and the threshold of the novel class confidence are subsequently optimized using a modified Firefly Algorithm, to enhance performance. The above proposed ensemble and optimization algorithms cooperate with each other to conduct the detection of novel unseen classes. Several UCI databases are employed for evaluation, i.e. The KDD Cup, Image Segmentation, Soybean Large, Glass and Iris databases. In comparison with the baseline meta-algorithms such as Boosting, Bagging and Stacking, our approach shows significantly enhanced performance with the increment of the number of novel classes for all the test data sets, which poses great challenges to the existing baseline ensemble methods.