• PR curve added to the visualization options
• UI layout changed to accomodate multiple visualizations
• code rewriten with optimization in mind: load times signifficantly reduced
• minor corrections in function documentations

mewto currently consists of two functions:

This function launches a Shiny application where the user can interactively manipulate the threshold used in binary classification and view the associated metrics, confusion matrix and ROC curve. The app also allows for optimal threshold calculation according to a weighted version of Youden's J-statistic.

In R, simply call the function:

mewtoApp(actuals, probabilities)

actuals - Data of factor type with two levels: "yes" for positive and "no" for negative.

probabilities - Data of numeric type which should represent the probabilities of realization of the positive category.

This function calculates the optimal threshold according to a weighted version of Youden's J-statistic.

mewtoThresh(actuals, probabilities, weight)

actuals - Data of factor type with two levels: "yes" for positive and "no" for negative.

probabilities - Data of numeric type which should represent the probabilities of realization of the positive category.

weight - The importance attributed to sensitivity, or formulated differently, to the maximization of the true positives rate.

This example generates 100000 observations of actual data (labelled "no", "yes"), as well as predicted values (ranging between 0 and 1), and stores tham in a dataframe called "df". The user can then call the functions included in the mewto package to perform exploratory analysis or obtain the optimum threshold value according to the weighted Youden J-statistic.

# Generate dataset
set.seed(123)
nobs = 100000 # Select the number of observations to be generated
predicted <- runif(nobs, 0, 1) # Probabilities representing the predicted values
thresh <- runif(nobs, 0.2, 0.8) # Intermediary step to generate actuals 
df <- data.frame(predicted, thresh) # "predicted" and "thresh" combined in "df"
df$actuals <- c("no", "yes")[(df$predicted >= df$thresh)   1] # Actual data

# Call the mewto library
library(mewto)

# Run mewtoApp to launch the visual interface and experiment
mewtoApp(df$actuals, df$predicted)

# Run mewtoThresh with a weight of 0.5 to obtain the optimal threshold according to Youden's original J-statistic
mewtoThresh(df$actuals, df$predicted, weight=0.5)

In the calculation of the optimal threshold, a weighted version of Youden's J-statistic (Youden, 1950) is employed. The optimal cut-off is the threshold that maximizes the distance to the identity (diagonal) line. The function maximizes the metric:

w * sensitivity (1 - w) * specificity, where "w" is the "weight" parameter.

Youden's J-statistic has been modified by adding the weighting parameter "w". The statistic varies in the interval [0;1]. Given a weighting factor w = 0.5, the weighted optimization function produces the same result as Youden's original J statistic. This particular statistic has been chosen since it is well-suited for weighting, and it is also the default criterion used in the R package pROC.

You can download mewto directly from Github. To do so, you need to have the devtools pachage installed and loaded. Once you are in R, run the following commands:

install.packages("devtools")

library("devtools")

install_github("alexandrumonahov/mewto")

You may face downloading errors from Github if you are behind a forewall or there are https download restrictions. To avoid this, you can try running the following commands:

options(download.file.method = "libcurl")

options(download.file.method = "wininet")

Alternatively, if you cannot download the file through Github, you may also download the binary package file from the link below:

https://github.com/alexandrumonahov/zip/blob/main/mewto.zip

Place the downloaded file into the working directory of R. The do one of the following:

Option 1) Run the following command:

install.packages('mewto_1.0.zip', repos = NULL, type = "win.binary")

Option 2) In RStudio:

Go to the Packages tab in the bottom-right pane and click on "Install". In the pop-up window that appears, click on "Browse" and choose the package mewto.zip that you have just downloaded. Click on "Install".

Once the package is stalled, you can run it using the: library(mewto) command.

I would like to give special thanks to Prof. Stefan Bender, Dr. Jens Mehrhoff, Gabriela Alves Werb and the Bundesbank ICBD team for having inspired the creation of this package.

• mewto's application launch
• interactive threshold component added
• weighted optimization algorithm developped based on Youden's J-Statistic
• confusion matrix and performance metrics analysis included
• ROC curve visualization augmented to display user's threshold selection on the curve

Alexandru Monahov, 2021