Our research focuses on deep learning methods for anomaly detection in and prediction of structured, imbalanced, and often sparsely labeled time series data such as climate data, environmental data, and structured data such as data from Enterprise Ressource Planning (ERP) systems.
We research and adapt deep learning methods to support and improve climate predictions. For handling extremly skewed variable distributions common in climate variables, we develop a universally applicable cost-sensitive learning method called DenseLoss: it automatically weighs samples by their rarity using kernel density estimators. We use this method to improve Model Output Statistics for precipitation prediction in climate models. Additionally, we research deep learning for ODEs and PDEs such as NeuralODEs to learn parametrisations for climate models.
For anomaly detection on ultra sparsely labeled data, we work on Semi-unsupervised learning algorithms. Semi-unsupervised learning algorithms learn to classify sparsely labeled data and expand the set of classes to accommodate unlabeled data that fits non of the known classes. We apply our algorithms to multivariate time series data of smart bee hives and to cell-gene expressions.
We research anomaly detection in transaction data, e.g. for finding financial fraud in ERP data supporting auditors, or detecting attacks and security breaches in IT networks. On the one hand, we address the highly specific challenges of these application domains with a specialized neural network architecture: the Improved Neural Arithmetic Logic Unit, our extension of the NALU which learns complex mathematical relations within our data. On the other hand, we research post-hoc explainable AI methods for anomaly detection which we use to ensure acceptable and comprehensible AI behavior in sensitive application fields such as fraud detection.