Visualization of Anomalies Using Mixture Models

Anomaly detection is important to learn from major past events and to prepare for future crises. We propose a new anomaly detection method that visualizes multivariate data in a 2- or 3-dimensional space based on the probability of belonging to a mixture component and the probability of not belonging to any components. It helps to visually understand not only the magnitude of anomalies but also the relationships among anomalous and normal samples. This may provide new knowledge in the data, since we can see it from a different viewpoint. We show the validity of the proposed method by using both an artificial and an economic time series.