Performs ICS via a shiny app where the user can change the scatter matrices, explore the output and download graphs and components.
Also the ICS outlier detection framework, from the ICSOutlier package is available.
It is inspired from the Factoshiny application of the FactoMineR package.
ICSShiny(x, S1 = MeanCov, S2 = Mean3Cov4,
S1args = list(), S2args = list(), seed = NULL)data matrix or dataframe with at least two numeric variables. Please note that it can contain non-numeric variables, but ICS is only performed on numeric variables.
list with optional additional arguments when calling function S1.
list with optional additional arguments when calling function S2.
to fix a seed when needed in order to fix the thresholds. Default is NULL. See details for more information.
Returns several tabs on the navigator:
The scatterplot matrix of an ICS object for the parameters chosen on the left part (variables included/excluded, the location vectors and scatter matrices).
Three different subtabs to help the user to choose the interesting components. The first sub-tab is the screeplot of the eigenvalues of the ICS object followed by the summary of the analysis. The second sub-tab plots the kernel density of the ICS components. The third sub-tab suggests which components to select, starting from the highest and/or the lowest kurtosis, through different normality tests or simulations.
The default values of the slidebar in the left are obtained from "agostino.test" at 5%.
The two sub-tabs aim at identifying groups or outliers by using pairwise plots of invariant coordinates. It offers two ways of plotting them: only two invariant components or a scatterplot matrix with up to six invariant components. The left panel allows to color the groups identified by the user and label the observations.
This tab plots outlyingness values for each observation based on the selected components. These squared ICS distances are computed through the ics.distances function as the Euclidian distance of the observations to the origin using the selected centered components. The identification of the outliers can be based on different cut-offs: from Monte Carlo simulations as in dist.simu.test or by giving a percentage or a number of observations to identify.
This tab gives some descriptive statistics on different subsets of the data (for all the observations, for the observations from a given cluster, for the outlying observations) and enables to compare the sub-populations. The application includes a boxplot, a kernel density, an histogram and some basic statistics: Min, Q1, Mean, Median, Q3 and Max.
This tab contains the dataset with a nice display and the possibility to choose different sub-populations of the data: all the observations, the observations from a given cluster or the outlying observations.
This tab allows to display and save the data table of components and the summary of operations. The data frame contains the components kept in the analysis as well as the distance generated by these components. It also includes the cluster the observation belongs to whether the observation is defined as an outlier, as well as the variables used for labelling and categorizing the data. The data are saved in a csv format. The summary of operations contains a summary of all parameters that were used to obtain the current result, it may be useful for another user who may want to get the same result as the original user. It is saved in a txt format.
The "Close the session" button closes the application and saves the icsshiny object into the global environment.
The scatter matrices and their associated location estimators can be selected through the list out of the options: MeanCov, Mean3Cov4, MCD,
TM. It is also possible to run the application with your own functions as long as they are passed as an argument of the call to ICSShiny. However, in this case it is not possible to run the simulations' steps for now.
ICS is only performed on numeric variables. Only non-numeric variables are proposed for labelling and/or categorizing the observations.
For computing the kernel densities in the second sub-tab, the weight is given by the Gaussian function and the bandwidth follows the rule of thumb of Silverman (1986).
For the automatic selection of the Invariant Components (IC), the referenced normality tests are the same as in the comp.norm.test function: "jarque.test", "anscombe.test",
"bonett.test", "agostino.test", "shapiro.test". All the decisions are corrected from multiple testing by adjusting the levels as in comp.norm.test.
The number of components to keep can also be decided from Monte Carlo simulations trough the comp.simu.test function. This parallel analysis method
may need a very long time to compute, so it is used only if the user clicks on the 'Launch the test' button.
Nordhausen, K., Oja, H. and Tyler, D.E. (2008), Tools for exploring multivariate data: The package ICS, Journal of Statistical Software, 28, 1--31. <doi:10.18637/jss.v028.i06>.
Archimbaud, A., Nordhausen, K. and Ruiz-Gazen, A. (2016), Multivariate Outlier Detection With ICS, <https://arxiv.org/pdf/1612.06118.pdf>.
# NOT RUN {
if(interactive()){
library(ICSShiny)
# ICS with ICSShiny:
res.shiny <- ICSShiny(iris)
# Close the session by clicking on the button or closing the navigator's tab
# ICS on a result of an ICSshiny object
ICSShiny(res.shiny)
# ICS with ICSShiny and different parameters
res.shiny <- ICSShiny(iris, S1 = MCD, S1args=list(alpha=0.7), seed = 7587)
}
# }
Run the code above in your browser using DataLab