Energy status data is typically time related and contains useful information to understand real world systems. Advanced measurement technology allows to collect data from such systems with high dimensionality and high volume. For example, in a production facility, modern smart meters allow to measure different physical quantities like voltage, frequency and harmonic distortion. With sample rates up to multiple measurements per second, these devices produce huge amounts of data. The data collected can give an indication about the behavior of machines and the quality of the electrical grid. In such a scenario, it is often of interest to detect unusual patterns which can reveal system misconfiguration or predict critical failures.

Most approaches to detect unusual patterns leave the interpretation of algorithmic results to the user. In addition, users need an advanced understanding of the detection methods to be able to adapt the algorithms to their needs. In the case of supervised learning, users must also provide the algorithm with additional ground truth information, such as the labeling of already known unusual patterns. The high dimensionality and the enormous volume of energy data makes this a challenging and time consuming task.

Focus of the doctoral thesis is to develop new approaches which utilize user feedback to advance the identification and interpretation of unusual patterns for high- dimensional data sets. There are two sides to this problem. First, the attention of a user is a scarce resource. Hence, mechanisms to request additional information from the user need to be designed carefully. Second, most methods for high-dimensional outlier detection are unsupervised and have no means to make use of labeled data. This calls for novel methods that consider both aspects. In addition, evaluation of novel approaches in this area is a challenge by itself. Not all identified anomalous patterns might be of interest to the user and the extent of interest can depend on individual preferences. This makes the quantification of interpretability and usefulness an open research task.

• Data Mining
• Outlier Detection in High Dimensional Data Sets

• Bachelorthesis: Improving Outlier Scores to Identify Suspicious Objects in Subspaces
• Masterthesis: Exploiting Neighborhood Relationships to Detect Suspicious Objects