Abstract: Conventional one-class learning models perform poorly when data are multi-modal or multi-density. To address this problem, ensemble clustering and clustering stability analysis for one class learning are introduced. Firstly, identifying the number of clusters and their distributions are unified in one enhancing framework. Then multiple one-class learning models are constructed to describe clusters of the target class. Lastly these one-class learning models are fused following the maximum fusion volume method. Using classic support vector data description (SVDD) as an instance of one-class learning algorithm, an ensemble cluster based stable SVDD, ECS-SVDD, is proposed. Experimental results on UCI benchmark datasets and a real-world malware detection dataset show that the ECS-SVDD outperforms the single SVDD and some other related one-class learning algorithms. Besides, the method proposed can also enhance the abilities of handling multi-modal and multi-density data of other one-class learning algorithms that follow the volume set minimizing scheme.

Key words: one-class learning, outlier analysis, cluster analysis, cluster stability, support vector data description