Group anomaly detection based on Bayesian framework with genetic algorithm

Anomaly detection is an important application field of evolutionary algorithm. Unlike traditionly anomaly detection, group anomaly detection aims to discover the anomalous aggregate behaviors in data points. Over past decades, a large number of promising methods have been successfully applied for group anomaly detection. However, they inherently neglect the correlations among groups in data points, limiting their abilities. This paper presents a correlated hierarchical generative model, which can model the intricate correlations hidden in groups by introducing a logistic normal distribution to capture the correlations among groups. With the proposed model, we construct a full variational Bayesian framework, which can data-adaptively optimize the model parameters of the proposed model. The model is designed and trained using Genetic Algorithm (GA), which helps automating the use of generative model. Further, a new score function is proposed as an anomaly criterion to estimate final anomaly groups in data points. Several experiments on synthetic data and real astronomical star data from Sloan Digital Sky Survey demonstrate the effectiveness of proposed method compared with the-state-of-art methods, in terms of average accurac (AP) and area under the Receiver Operating Characteristic(ROC) curve(AUC).