Title: BIRCH based localized data analysis

Description:

With ever-growing data volumes in nearly all sectors of our lives, unsupervised anomaly de- tection is an important and multifaceted field of modern data science. In order to differentiate anomalous from normal data, one can apply distance based outlier detection techniques.

This thesis introduces two unsupervised methods incorporating the aforementioned ap- proach to identify anomalies. Both rely on the scalable clustering algorithm BIRCH, which partitions a given global data region into specialized subregions for expert models to enable lo- calized data analyses. While the first approach utilizes local linear Principal Component Analysis experts, the second method applies non-linear One-Class Support Vector Machine models within the identified localities. Moreover, an implemented evaluation framework is presented as a wrapper for the two expert systems. Finally, both methods are evaluated quantitatively and compared to their BIRCH-less, global counterparts, as well as to the Isolation Forest anomaly detection algorithm.

Supervisor: Boris Lorbeer

Type:  Bachelor Thesis

Duration: 4 months