Anomaly detection in univariate time series incorporating active learning

In this research. we study anomaly detection in univariate time series and optimize according to a business objective using a novel active learning approach. The motivation is to detect anomalies while monitoring systems within an IT infrastructure, known as intrusion detection, specifically for hotel organizations. The proposed detector is based on moving averages in combination with a prediction interval, where parameters are optimized via an active learning component. By using prediction intervals, the results are easily interpretable for domain experts due to the white-box nature of the detector. Annotations originating from domain experts serve as input to acquire oracle parameters, which are obtained via Bayesian optimization using Gaussian process. The detector is tested on the Numenta Anomaly Benchmark (NAB) and is compared to commonly used black-box models.