A Hierarchical Probabilistic Deep Learning Approach for Contextual Anomaly Detection in Mixed-Type Tabular Data

Contextual anomaly is a subtype of anomaly that, when observed in isolation, may not have the characteristics of an anomaly but becomes an anomaly when observed within a given context. Contextual anomaly detection is applied in several areas, such as industrial production process control, computer network security, and financial fraud detection. Probabilistic approaches to solving this problem are often neglected or insufficiently researched compared to numerous other anomaly detection techniques, and efficient algorithms for mixed-type tabular data are generally lacking. In this paper, we present a Hierarchical Variational Autoencoder for contextual anomaly detection, a new promising approach based on probabilistic modeling with a hierarchy of two levels: contextual and common. The trained model predicts the conditional likelihood of an event with respect to the observed context, which is used to distinguish between anomalies and regular data. We experimentally confirm the improvement of the results compared to a benchmark model that does not consider contextual information.