Time-series Outlier Detection App
Authors: Mauro Gwerder, Maciej Dobrzynski, Marc-Antoine Jacques
Working with Shiny < 1 year
Abstract: This shiny-application was developed to facilitate working with time-series data, especially handling outliers. Although primarly made and used for handling microscopic video-data created in the Pertz lab, the application handles any kind of time-series data. There are currently four different modules that provide different tools that are suited for different kinds of time-series-data outliers.
Full Description: In the Pertz lab, visual signals produced by biomarkers in cells are measured over time, resulting in time-series data. Several kinds of outliers can occur, which can interfere with successful data analysis. As these outliers can vary strongly in their properties, only one method wouldn't be enough to handle all of them.
Currently, four modules are available:
- Quantile Trimming (author: Maciej Dobrzynski)
- Isolation Tree
- Rolling Window
- Interactive PCA
This module was developed by Maciej Dobrzynski. It visualizes the distribution of all measurements and offers to determine a cut-off, such that very extreme trajectories can be removed right away.
This module is called that way, because a single tree is created using hierarchical clustering. This tree will carry outliers on the outer branches, such that we can trim of one branch one by one to get through the outliers. This will also group similar outliers together.
One category of outliers are punctual spikes inside of an otherwise normal trajectory. A rolling window will slide across the trajectory and calculate a reasonable range of expected values depending on the current context of the trajectory. These spikes can be imputated to replace them with a more reasonable value, however, this is only possible for short outliers.
The additional possibility to extract features before analysing the data enables to specifically compare certain properties of the data. More features and an interactive boxplot will be added in the future.
Keywords: statistical analysis, time-series, outliers, data exploration
Shiny app: https://outlierapp.shinyapps.io/project/
RStudio Cloud: https://rstudio.cloud/spaces/57625/project/1047820