On Kernel Search Based Gaussian Process Anomaly Detection

Anomaly detection becomes more important with increasing automation. Especially for time series data, prevalent in industry, there are numerous methods that have been well researched. In this work we provide a proof of concept for a novel approach using the interpretability of Gaussian processes. To detect an abnormal section, the data is split into equally sized segments which are then interpreted individually using separate kernel searches. The resulting kernels can then be compared and clustered by one of multiple presented methods. The segments that contain an anomaly end up in their own cluster.

To test all possible configurations of our proposed approach, we applied them to a subset of the SIGKDD 2021 anomaly dataset mutliple times and evaluated the results. Almost all configurations were able to succeed, although not yet reliably reproducible. The results of our performance evaluation indicate that kernel searches are in principle applicable to anomaly detection in univariate time series data.