Schmidl, SebastianWenig, PhillipPapenbrock, ThorstenKönig-Ries, BirgittaScherzinger, StefanieLehner, WolfgangVossen, Gottfried2023-02-232023-02-232023978-3-88579-725-8https://dl.gi.de/handle/20.500.12116/40327In many domains, such as data cleaning, machine learning, pattern mining, or anomaly detection, a system’s performance depends significantly on the selected configuration hyperparameters. However, manual configuration of hyperparameters is particularly difficult because it requires an in-depth understanding of the problem at hand and the system’s internal behavior. While automatic methods for hyperparameter optimization exist, they require labeled training datasets and many trials to assess a system’s performance before the system can be applied to production data. Hence, automatic methods just shift the human effort from parameter optimization to the effort of labelling datasets, which is still complex and time-consuming. In this paper, we, therefore, propose a novel hyperparameter optimization framework called HYPEX that learns promising default parameters and explainable parameter rules from synthetically generated datasets, without the need for manually labeled datasets. HYPEX’ learned parameter model enables the easy adjustment of a system’s configuration to new, unlabeled, and unseen datasets. We demonstrate the capabilities of HYPEX in the context of time series anomaly detection because anomaly detection algorithms suffer from a general lack of labeled datasets and they are particularly sensitive to parameter changes. In our evaluation, we show that our hyperparameter suggestions on unseen data significantly improve an algorithm’s performance compared to existing manual hyperparameter optimization approaches and often are competitive to the optimal performance achieved with Bayesian optimization.enTime Series Anomaly DetectionBayesian OptimizationCausal DiscoveryText/Conference Paper10.18420/BTW2023-22