Diversity Aware-Based Sequential Ensemble Learning for Robust Anomaly Detection

Anomaly detection has become a popular topic in many domains because anomalies can provide valuable information. Recently, ensemble learning has been applied to improve the generalization ability of existing anomaly detection methods. In an anomaly ensemble framework, diversity is essential for building a powerful ensemble method to obtain impressive performance. However, most studies use heuristic techniques to improve the diversity of ensembles, and it generally leads to limited diversity. To obtain improved diversity, we propose a diversity aware-based sequential ensemble method (called D-SEM) for anomaly detection. Specifically, our proposed method divides the ensemble diversity into two parts: sample diversity and model diversity. For sample diversity, we introduce the subsampling technique to implement preliminary generation of diverse datasets for training. For model diversity, we design an ensemble-based optimization model to learn base classifiers with improved diversity. Furthermore, an unsupervised diversity measure is proposed to quantitatively assess diversity and an anomaly pruning strategy is utilized to successively eliminate pseudo-anomalies. Based on the inclusion of sample diversity and model diversity, the proposed D-SEM method obtains better generalization ability for anomaly detection. The experimental results based on real-world datasets suggest that the proposed method has superior performance compared with various state-of-the-art methods.