Unsupervised Anomaly Detection
- 1 min The aim of the project was to develop a more precise method using anomaly detection, which evaluates these loading processes on the basis of the log files and provides information on possible errors gives. In the choice of methods, the focus was on the broadest possible applicability for the future, so that even higher dimensional data sets can be processed. A particular challenge here was the evaluation of the methods used at the Base of unlabeled data. For this purpose we use an approach that has been developing since 2015 and on the basis of Excess-Mass and Mass-Volume curves statements about the quality of scoring functions meets.
The project showed that the methods iForest and Local Outlier Factor are suitable for to detect anomalies or to investigate observations. A suitable criterion for the evaluation of the methods and for the determination of the parameter values was the Excess-Mass-Volume criterion. These allow a Comparison between methods and can also handle unlabeled data.
