Talks and Poster Presentations (without Proceedings-Entry):

C. Bors, M. Bögl, J. Bernard, T. Gschwandtner, S. Miksch:
"Quantifying Uncertainty in Time Series Data Processing";
Talk: VisInPractice Mini-Symposium on Visualizing Uncertainty, Berlin (invited); 2018-10-22.

English abstract:
Uncertainty visualization has become an integral part of many data analysis applications, aiding practitioners in making informed decisions, particularly when uncertain aspects are involved.
However, the assessment and quantification of uncertainty introduced by data processing methods is still neglected in application scenarios.
Using human motion and activity recognition as an example, different machine learning and data mining routines can be applied for data processing and analysis, which change the value domain of the underlying multivariate time series.

We identify value uncertainty as the information of changes done to the value domain, e.g., by cleansing/wrangling data to make them suited for further analysis.
For domain experts to appropriately account for uncertainties in their decision making process, they need to be quantified and externalized in the visualizations.
To accomplish this, we implemented a quantification of value uncertainties from
commonly used data processing routines for time series data (e.g., smoothing operations).
We also provide different aggregation methods of value uncertainties over consecutive routines in data processing pipelines by employing various uncertainty quantification techniques (e.g., statistical, bayesian, probabilistic).
This allows developers of data processing pipelines as well as users of the resulting visualization to consider the results with appropriate knowledge of value uncertainty that influenced the analysis outcome.
In our visual interactive environment different processing and segmentation routine results are juxtaposed, the most appropriate motion and activity recognition pipeline can be selected by the domain expert under the consideration of (a) segmentation accuracy, (b) value uncertainties introduced into the data, and (c) overall uncertainty of the result.

Related Projects:
Project Head Silvia Miksch:
Visuelle Segmentierung und Labeling multivariater Zeitserien

Created from the Publication Database of the Vienna University of Technology.