Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Visual guidance to find the right spot in the parameter space

Visuelle Unterstützung bei der Parameterauswahl
: Brakowski, Alexander
: Kuijper, Arjan; Maier, Sebastian

Darmstadt, 2015, 56 pp.
Darmstadt, TU, Bachelor Thesis, 2015
Bachelor Thesis
Fraunhofer IGD ()
Business Field: Visual decision support; Research Area: Human computer interaction (HCI); data exploration; feedback; Active Learning; visual data analysis

The last few decades brought upon a technological revolution that has been generating data by users with an ever increasing variety of digital devices, resulting in such an incredible volume of data, that we are unable to make any sense of it any more. This trend also does not seem to stop in the near future. This problem requires us to think about new methods to help us explore these large data sets. We already have a lot of Machine learning and Data mining algorithms, that are able to help us find something interesting in the data, but because of their high algorithmic complexity using them can make the data exploration extremely time consuming and frustrating. One solution to decrease the required execution time of these algorithms would be the preprocessing of the data by sampling it before starting the exploration process. That indeed does help, but one issue remains when using the available Machine Learning and Data Mining algorithms: they all have parameters. That is a big problem for most users, because a lot of these parameters require expert knowledge to be able to tune them. Even for expert users a lot of the parameter configurations highly depend on the data. In this work I will present a system that tackles that data exploration process from the angle of parameter space exploration. Here we use the active learning approach and iteratively try to query the user for their opinion of an algorithm execution. For that an end-user only has to express a preference for algorithm results presented to them in form of a visualisations. That way the system is iteratively learning the interest of the end-user, which results in good parameters at the end of the process. A good parametrisation is obviously very subjective here and only reflects the interest of an user. This solution has the nice ancillary property of omitting the requirement of expert knowledge when trying to explore an data set with Data Mining or Machine Learning algorithms. Optimally the end-user does not even know what kind of parameters the algorithms require. Furthermore I present an intuitive frontend that allows an end-user to configure the exploration process, which includes choosing the data and the algorithm that should be used for the exploration. In the end of this thesis I present some interesting use cases for this process.