Anomaly Detection Towards Modelling with Outlier Robustness

Machine learning algorithms have evolved in numerous fields into an important enabler for today's society. Thereby, anomaly detection is one of the machine learning tasks which contributes to many application domains. Anomalies,  also known as outliers, can thereby be of advantageous but also of detrimental effect, so that outlier robustness is one of the desired properties when it comes to the design of performant machine learning algorithms. 
This dissertation elaborates on how a point anomaly detection algorithm needs to be designed to neglect negative impacts of anomalies during the learning phase in order to improve their capability to detect anomalies. Furthermore, it shows how an anytime approach allows to provide permanent response and performance improvements simultaneously for point anomaly detection. Moreover, a benchmark framework for reproducible point anomaly detection is proposed. In addition to that, traditional and deep learning approaches for group anomaly detection on mobility data are introduced, as well as a deep 3D convolutional neural network for robust soil parameter prediction on hyperspectral satellite image data.

Rechte

Nutzung und Vervielfältigung:


CC BY-SA 4.0

Bitte beachten Sie, dass einzelne Bestandteile der Publikation anderweitigen Lizenz- bzw. urheberrechtlichen Bedingungen unterliegen können.

Zitieren

Zitierform:
Lohrer, A., 2024. Anomaly Detection Towards Modelling with Outlier Robustness.
Zitierform konnte nicht geladen werden.