The Good, The Bad, and The Average: Benchmarking of Reconstruction Based Multivariate Time Series Anomaly Detection

Reconstruction-based algorithms offer state-of-the-art performance in multivariate time series anomaly detection. But as always: there is no single best algorithm. To find the optimal solution, one has to compare different methods and tune their hyperparameters. This paper introduces a lightweight modular benchmarking framework for data scientists and researchers in the field. The framework can be easily set up and automatically create a visual summary of the relevant performance indicators and automatically selected examples to give insight into the behavior of the model and aid during the development.