Uncertainty types in segmenting and labeling time series data
The segmenting and labeling of time series can be described as a three stage process.
Given a multivariate time series, multiple (pre-)processing steps are applied to make
the data useful for subsequent analysis. Secondly, appropriate segmenting and labeling
algorithms are utilized to provide meaningful segments and labels. Different algorithms
and parameter inputs generate different results and subsequently need to be explored for
the most adequate result. However, all steps are affected by uncertainty. Throughout the
process they are introduced and altered, thus affecting the reliability and trustworthiness
of the results.
In this work we identify four types of uncertainty that influence this process and fur-
thermore the latter decision making. (a) Value uncertainty is comprised of uncertainties
inherent in the data input and stemming from processing routines that affect the value
domain (e.g., noise reduction). (b) Result uncertainty stems from the likelihood of a
segment actually representing a particular label/result, and the definitive start and end
time, being contingent on the implicit transitions between labels. Uncertainty will be (c)
aggregated by various processing steps but also when being visualized, if multiple seg-
mentation results are shown and screen resolution is insufficient. (d) Cause and effect
uncertainty is implicit, stemming from consecutively adjusting and comparing different
algorithms and/or algorithm parametrizations.
In order for users to judge which algorithm and parameter configurations provide ade-
quate results while trading the influence of uncertainty on the outcome, it is necessary to
inspect different stages of analysis with uncertainty information available. To accomplish
this we derived different types of visual elements from  to effectively convey the afore-
mentioned types of uncertainty throughout the entire analysis pipeline. Depending on the
analysis task, the associated uncertainty needs to be visually communicated appropriately
within the visual representation of the data. During the (pre-)processing step, inherent
value uncertainty is visually externalized in an enhanced processing view. When selecting
potential algorithms and parameter configurations, the influence of result uncertainty is
shown for algorithms and parameters individually. In the last step of analysis, the effects
of result and aggregation uncertainties require attention to distinguish if segmenting and
labeling results were correctly determined, and how credible they are.
|Year of Publication||
|Funding projects (new)|