Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Visualizing Time Series Consistency for Feature Selection

: Cibulski, Lena; May, Thorsten; Preim, Bernhard; Bernard, Jürgen; Kohlhammer, Jörn


Journal of WSCG 27 (2019), No.2, pp.93-102
ISSN: 1213-6972
ISSN: 1213-6964
International Conference in Central Europe on Computer Graphics, Visualization and Computer Vision (WSCG) <27, 2019, Plzen, Czech Republic>
European Commission EC
H2020; 768892; CloudiFacturing
Cloudification of Production Engineering for Predictive Digital Manufacturing
Journal Article, Conference Paper
Fraunhofer IGD ()
Lead Topic: Digitized Work; Research Line: Computer graphics (CG); Research Line: Modeling (MOD); visual analytic; feature selection; consistency; multivariate time series

Feature selection is an effective technique to reduce dimensionality, for example when the condition of a system is to be understood from multivariate observations. The selection of variables often involves a priori assumptions about underlying phenomena. To avoid the associated uncertainty, we aim at a selection criterion that only considers the observations. For nominal data, consistency criteria meet this requirement: a variable subset is consistent, if no observations with equal values on the subset have different output values. Such a model-agnostic criterion is also desirable for forecasting. However, consistency has not yet been applied to multivariate time series. In this work, we propose a visual consistency-based technique for analyzing a time series subset’s discriminating ability w.r.t. characteristics of an output variable. An overview visualization conveys the consistency of output progressions associated with comparable observations. Interaction concepts and detail visualizations provide a steering mechanism towards inconsistencies. We demonstrate the technique’s applicability based on two real-world scenarios. The results indicate that the technique is open to any forecasting task that involves multivariate time series, because analysts could assess the combined discriminating ability without any knowledge about underlying phenomena.