Introduction to ROCnGO"
In ROCnGO: Fast Analysis of ROC Curves

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

ROCnGO is an R package which allows to analyze the performance of a classifier by using receiver operating characteristic ($ROC$) curves. Conventional $ROC$ based analyses just tend to use area under $ROC$ curve ($AUC$) as a metric of global performance, besides this functionality, the package allows deeper analysis options by calculating partial area under $ROC$ curve ($pAUC$) when prioritizing local performance is preferred.

Furthermore, ROCnGO implements different $pAUC$ transformations described in literature which:

Make local performance interpretation easier.
Allow to work with $ROC$ curves which are not completely concave or not at all (improper).
Provide additional discrimination power when comparing classifiers with identical local performance (equal $pAUC$).

This document provides an introduction to ROCnGO tools and workflow to study the global and local performance of a classifier.

Prerequisites

In order to reproduce the example, following packages are needed:

library(ROCnGO)
library(dplyr)
library(forcats)

Data

To explore basic tools in the package we will be using iris dataset. The dataset contains 5 variables for 150 flowers of 3 different species: setosa, versicolor and virginica.

For the purpose of simplicity, we will only work with a subset of iris, considering only setosa and virginica species. In the following sections, performance of different variables to classify cases in the different species will be evaluated.

# Filter cases of versicolor species
iris_subset <- as_tibble(iris) %>% filter(Species != "versicolor")
iris_subset

Global performance

Calculate ROC curve

The foundation of this type of analyses implies to plot the $ROC$ curve of a classifier. This type of curves represent a classifier probability of correctly classify a case with a condition of interest, also known as true positive rate or $\text{Sensitivity}$ ($TPR$), and the complementary probability of correctly classify a case without the condition; also known as false positive rate, $1 - \text{Specificity}$, or $1 - TNR$, ($FPR$).

When working with a classifier that returns a series of numeric values, it can be complex to say when it is classifying a case as having the condition of interest (positive) or not (negative). To solve this problem, $ROC$ curves represent $(FPR, TPR)$ points considering hypothetical thresholds ($c$) where a case is considered as positive if its value is higher than the defined threshold ($X > c$).

These curve points can be calculated by using roc_points(). As most functions in the package, it takes a dataset, a data frame, as its first argument. The second and third argument refer to variables in the data frame, corresponding the variable that will be used as a classifier (predictor) and the response variable we want to predict (response).

For example, we can calculate $ROC$ points for Sepal.Length as a classifier of setosa species.

# Calculate ROC points for Sepal.Lenght
points <- roc_points(
  data = iris_subset,
  predictor = Sepal.Length,
  response = Species
)
points

# Plot points
plot(points$fpr, points$tpr)

As we may see, Sepal.Length doesn't perform very well predicting when a flower is from setosa species, in fact it's the other way around, the lower the Sepal.Length the more probable to be working with a setosa flower. This can be tested if we change the condition of interest to virginica.

Changing condition of interest

By default, condition of interest is automatically set to the first value in levels(response), so we can change this value by changing the order of levels in data.

# Check response levels
levels(iris_subset$Species)

# Set virginica as first value in levels
iris_subset$Species <- fct_relevel(iris_subset$Species, "virginica")
levels(iris_subset$Species)

# Plot ROC curve
points <- roc_points(
  data = iris_subset,
  predictor = Sepal.Length,
  response = Species
)
plot(points$fpr, points$tpr)

Local performance

Sometimes a certain task may requiere prioritize e.g. high sensitivity over global performance. In these scenarios, it's preferable to work in specific regions of $ROC$ curve.

We can calculate points in a specific region using calc_partial_roc_points(). Function uses same arguments as roc_points() but adding lower_threshold, upper_threshold and ratio, which delimit region in which we want to work.

For example, if we require to work in high sensitivity conditions, we could check points in region $(0.9, 1)$ of $TPR$.

# Calc partial ROC points
p_points <- calc_partial_roc_points(
  data = iris_subset,
  predictor = Sepal.Length,
  response = Species,
  lower_threshold = 0.9,
  upper_threshold = 1,
  ratio = "tpr"
)
p_points

# Plot partial ROC curve
plot(p_points$fpr, p_points$tpr)

Automating analysis

Performance metrics

When working with a high number of classifiers, it can be difficult to check each $ROC$ individually. In these scenarios, metrics such as $AUC$ and $pAUC$ may present more interest. Thus, by using the function summarize_predictor() we can obtain an overview of the performance of a classifier.

For example, we could consider the performance of Sepal.Length over a high sensitivity region, $TPR \in (0.9, 1)$, and high specificity region, $FPR \in (0, 0.1)$.

# Summarize predictor in high sens region
summarize_predictor(
  data = iris_subset,
  predictor = Sepal.Length,
  response = Species,
  threshold = 0.9,
  ratio = "tpr"
)

# Summarize predictor in high spec region
summarize_predictor(
  data = iris_subset,
  predictor = Sepal.Length,
  response = Species,
  threshold = 0.1,
  ratio = "fpr"
)

Besides $AUC$ and $pAUC$, function also returns other partial indexes derived from $pAUC$ which provide a better interpretation of performance than $pAUC$.

Furthermore, if we are interested in computing these metrics simultaneously for several classifiers summarize_dataset() can be used, which also provides some metrics of analysed classifiers.

summarize_dataset(
  data = iris_subset,
  response = Species,
  threshold = 0.9,
  ratio = "tpr"
)

Plotting

As we have seen, by using the output of roc_points() we can plot $ROC$ curve. Nevertheless, these plots can also be generated using plot_*() and add_*() functions, which provide further options to customize plot for classifier comparison.

For example, we can plot $ROC$ points of Sepal.Length in this way.

# Plot ROC points of Sepal.Length
sepal_length_plot <- plot_roc_points(
  data = iris_subset,
  predictor = Sepal.Length,
  response = Species
)
sepal_length_plot

Now by using + operator we can add further options to the plot. For example, including chance line, adding further $ROC$ points of other classifiers, etc.

sepal_length_plot +
  add_roc_curve(
    data = iris_subset,
    predictor = Sepal.Width,
    response = Species
  ) +
  add_roc_points(
    data = iris_subset,
    predictor = Petal.Width,
    response = Species
  ) +
  add_partial_roc_curve(
    data = iris_subset,
    predictor = Petal.Length,
    response = Species,
    ratio = "tpr",
    threshold = 0.7
  ) +
  add_threshold_line(
    threshold = 0.7,
    ratio = "tpr"
  ) +
  add_chance_line()