Human-Interpretable Rules for Anomaly Detection in Time-series
Anomaly detection in time series is a widely studied issue in many areas. Anomalies can be detected using rule-based approaches and human-interpretable rules for anomaly detection refer to rules presented in a format that is intelligible to analysts. Learning these rules is a challenge but only a few works address the issue of detecting different types of anomalies in time-series.
This paper presents an extended decision tree based on patterns to generate a minimized set of human comprehensible rules for anomaly detection in univariate times-series. This method uses Bayesian optimization to avoid manual tuning of hyper-parameters.
We define a quality measure to evaluate both the accuracy and the intelligibility of the produced rules. Experiments show that our approach generates rules that outperforms the state of- the-art anomaly detection techniques.