Forschungsthemen
[MA] Investigation of Domain Transposition and Pattern Recognition Technologies for Correlation Detection and Analysis in the Context of Big Data
The ever-increasing size, dimensionality and complexity of current data caused by recent develop- ments in the fields of advanced data aggregation and the Internet of Things, require appropriate data analysis methods to automatically extract hidden knowledge and yield profit from large data. State-of the-Art Data Science technologies, however, come to their limits when applied to detect and analyze correlations <1> in Big Data contexts. Hence, there exists a demand for novel methods to determine the existence and characteristics of unknown data relations. A new approach assumes that the data can be transposed to a lower-dimensional and uniform representation, utilizing a customizable Multi-Dimensional Scaling (MDS) technique of low time complexity <2>, so that hidden correlations form characteristics patterns. In this way, these patterns can be detected automatically by known Pattern Recognition (PR), to conclude their existence and types. This thesis should design, implement, test and evaluate this approach. First, existing methods for large data have to be evaluated w.r.t. their result qualities and computing performance. With respect to known correlation types <1>, a provided MDS frame- work by <2> has to be customized and a generator for scalable data with meaningful, artificial correlations has to be developed. In the following, the customized MDS has to be applied to the generated data. The result patterns has to be evaluated with respect to the known cor- relations from the data generator and conventional correlation analysis techniques. Based on these insights, appropriate PR techniques has to be retrieved from literature and included in a framework extension, enabling automated correlation detection and analysis in this way. Finally, this new approach has to be compared to existing correlation analysis methods. <1> Andreas Hilbert. Zur Theorie der Korrelationsanalyse. PhD thesis, Universität Augsburg, 1998. <2> Karsten Wendt. Multi-Objective Optimization Utilizing Cluster Analysis Applied to Dimen- sional Transposed Problems. PhD thesis, Technische Universität Dresden, 2016.
Betreuer: Karsten Wendt-:#-#:- Patrick Zschech