KAVis: Visual Interactive Space-Time Segmentation
We are hiring!
We are seeking highly motivated and talented individuals to join our dynamic research team for research in visual interactive segmentation of multivariate spatial and temporal data dealing with uncertainties. The successful candidates will conduct research in our team at TU Wien as a Post-Doc researcher and one PhD student. Details on the position in the News. Applications and questions to Markus Bögl.
Summary: Many application domains generate and analyze multivariate spatial time series and often face data quality issues. Segmentations are very useful for exploratory analysis of these varying data characteristics, but appropriate segmentations are currently challenging or frequently impossible. The project aims to integrate background and domain knowledge within a VA process to provide knowledge-assisted interactive segmentation of such multivariate spatial time series and deal with data quality issues (e.g., missing and uncertain values).

Motivation: Spatial time series, i.e., time series measured at geographical locations (points or areas), are a common data type in industry and scientific fields. Examples include but are not limited to commerce (e.g., sales of items in shops), meteorology (see Figure), demographics (e.g., household wealth), or epidemiology (e.g., Covid cases). Often, more than one variable is measured (multivariate data). In addition to being temporally and spatially distributed, the measured variables are often plagued by data quality issues, such as missing or extreme values, due to diverse reasons such as sensor failures. Measurements might also be uncertain, either because missing values were imputed or because the measurement process itself cannot precisely determine the value.

Detailed aims: Our project aims to enable knowledge-assisted interactive segmentation of such multivariate spatial time series with data quality issues (e.g., missing and uncertain values). While interactive segmentation approaches are common for multimedia data, the particular data characteristics we consider in our proposal define its novelty. Unlike images where data is evenly distributed on a grid, geographical locations will form sparsely and densely populated areas. Consequently, edges in the underlying graph will have non-uniform lengths, thus violating a core assumption of many existing approaches. These approaches also rarely deal with missing or uncertain data, as color information for pixels in an image is usually fully and perfectly defined. Further, while color images are technically multivariate data, we perceive the image as one entity. It contrasts geographic data, where variables may be considered individually or in user-defined groups. These characteristics also require the analyst to preprocess (wrangle) the dataset iteratively, e.g., by selecting variables, choosing temporal or spatial scales, imputing data points, or verifying assumptions of statistical methods. Every action influences the segments of time, space, or space-time that may be appropriate for the dataset. The segmentation procedure must be interactive to ensure these segments correspond to the user’s domain knowledge. Several alternatives might be appropriate and thus developed in parallel, with a final selection process at the end. Interactive visualizations are promising to support these complex tasks, but novel techniques must be designed.
- Markus Bögl
- Silvia Miksch
- 1 PostDoc position to be filled
- 1 PhD position to be filled
- Helwig Hauser, University of Bergen, Norway
- Cagatay Turkay, University of Warwick, UK
- Wolfgang Wagner, TU Wien, Austria
- VRVis, Austria
3 years, 2025-2028
Austrian Science Fund (FWF), grant [PAT 6437224]; Grant-DOI [10.55776/PAT6437224]
Many application domains generate and analyze multivariate spatial time series and often face data quality issues. Segmentations are very useful for exploratory analysis of these varying data characteristics, but appropriate segmentations are currently challenging or frequently impossible. The project aims to integrate background and domain knowledge within a VA process to provide knowledge-assisted interactive segmentation of such multivariate spatial time series and deal with data quality issues (e.g., missing and uncertain values).