Mixtures of Probabilistic Principal Component Analyzers for Anomaly Detection

Description/Abstract

Anomaly detection tools have been increasingly used in recent years to generate predictive insights on rare events. The typical challenges encountered in such applications include a large number of data dimensions and absence of labeled data. An anomaly detection strategy for these scenarios is dimensionality reduction followed by clustering in the reduced space, with the degree of anomaly of an event or observation quantified by statistical distance from the clusters. However, most research efforts so far are focused on single abrupt anomalies, while the correlation between observations is completely ignored. In this paper, we address the problem of detection of both abrupt and sustained anomalies with high dimensions. The task becomes more challenging than only detecting abrupt outliers because of the gradual and indiscriminant changes in sustained anomalies. We utilize a mixture model of probabilistic principal component analyzers to quantify each observation by probabilistic measures. A statistical process control method is then used to monitor both abrupt and gradual changes. On the other hand, the mixture model can be regarded as a trade-off strategy between linear and nonlinear dimensionality reductions in terms of computational efficiency. This compromise is particularly important in real-time deployment. The proposed method is evaluated on simulated and benchmark data, as well as on data from wide-area sensors at a truck weigh station test-bed.