Uncertainty-informed dynamic threshold for time series anomaly detection

As time series data continues to be collected across various fields, the importance of automated anomaly detection systems is steadily increasing. A key challenge in anomaly detection lies in setting an optimal threshold for anomaly scores to distinguish anomalies from normal data. Most existing studies use a fixed threshold, often resulting in misclassification of ambiguous data. Therefore, defining a dynamic and optimal threshold is crucial for improving detection performance. We aim to quantify uncertainty as a metric that determines the degree of ambiguity in the data. Because our models are trained only on normal data, anomalies exhibiting patterns divergent from the normal data entail higher uncertainty. Accordingly, in this study, we propose a dynamic thresholding method that better aligns with the nature of the data through uncertainty quantification. Through experimentation with synthetic datasets and five benchmark datasets for time series anomaly detection, we demonstrate the efficacy of our proposed method. Our proposed method outperforms both the fixed threshold and existing dynamic thresholding methods, achieving an average F1-score improvement of over 0.06 across benchmark datasets. In particular, the performance improvement is more significant when the distributions of normal data and anomalies are more similar. The source code can be accessed at https://github.com/jungminkr9195/UDT.