ROC curves for multiclass classification in R

Question

I have a dataset with 6 classes and I would like to plot a ROC curve for a multiclass classification. The first answer in this thread given by Achim Zeileis is a very good one.

ROC curve in R using rpart package?

But this works only for a binomial classification. And the error i get is Error in prediction, Number of classes is not equal to 2. Any one who has done this for a multi-class classification?

Here is a simple example of what I am trying to do. data <- read.csv("colors.csv")

let's say data$cType has 6 values (or levels) as (red, green, blue, yellow, black and white)

Is there anyway to plot a ROC curve for these 6 classes? Any working example for a class of more than 2 would be appreciated.

Answer 1

I know this is an old question, but the fact that the only answer is written using Python bothers me a lot, given that the question specifically asks for an R solution.

As you can see from the code below, I am using pROC::multiclass.roc() function. The only requirement to make it work is that the names of the columns of the predictions matrix match the true classes (real_values).

The first example generates random predictions. The second one generates a better prediction. The third one generates the perfect prediction (i.e., always assigning the highest probability to the true class.)

library(pROC)
set.seed(42)
head(real_values)
real_values <- matrix( c("class1", "class2", "class3"), nc=1 )

# [,1]    
# [1,] "class1"
# [2,] "class2"
# [3,] "class3"

# Random predictions
random_preds <- matrix(rbeta(3*3,2,2), nc=3)
random_preds <- sweep(random_preds, 1, rowSums(a1), FUN="/")
colnames(random_preds) <- c("class1", "class2", "class3")


head(random_preds)

#       class1    class2    class3
# [1,] 0.3437916 0.6129104 0.4733117
# [2,] 0.6016169 0.4700832 0.9364681
# [3,] 0.6741742 0.8677781 0.4823129

multiclass.roc(real_values, random_preds)
#Multi-class area under the curve: 0.1667



better_preds <- matrix(c(0.75,0.15,0.5,
                         0.15,0.5,0.75,
                         0.15,0.75,0.5), nc=3)
colnames(better_preds) <- c("class1", "class2", "class3")

head(better_preds)

#       class1 class2 class3
# [1,]   0.75   0.15   0.15
# [2,]   0.15   0.50   0.75
# [3,]   0.50   0.75   0.50

multiclass.roc(real_values, better_preds)
#Multi-class area under the curve: 0.6667


perfect_preds <- matrix(c(0.75,0.15,0.5,
                          0.15,0.75,0.5,
                          0.15,0.5,0.75), nc=3)
colnames(perfect_preds) <- c("class1", "class2", "class3")
head(perfect_preds)

multiclass.roc(real_values, perfect_preds)
#Multi-class area under the curve: 1

Answer 2

Answering an old question while having the same requirement - I've found the scikit documentation explains a few approaches well.

http://scikit-learn.org/stable/auto_examples/model_selection/plot_roc.html

The approaches mentioned include:

• "binarizing" i.e. converting the problem to binary classification, using either macro-averaging or micro-averaging
• Draw multiple ROC curves, one per label
• One vs. All

Copying example from the above link, which illustrates one vs. all and micro averaging using their libs:

print(__doc__)

import numpy as np
import matplotlib.pyplot as plt
from itertools import cycle

from sklearn import svm, datasets
from sklearn.metrics import roc_curve, auc
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import label_binarize
from sklearn.multiclass import OneVsRestClassifier
from scipy import interp

# Import some data to play with
iris = datasets.load_iris()
X = iris.data
y = iris.target

# Binarize the output
y = label_binarize(y, classes=[0, 1, 2])
n_classes = y.shape[1]

# Add noisy features to make the problem harder
random_state = np.random.RandomState(0)
n_samples, n_features = X.shape
X = np.c_[X, random_state.randn(n_samples, 200 * n_features)]

# shuffle and split training and test sets
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=.5,
                                                    random_state=0)

# Learn to predict each class against the other
classifier = OneVsRestClassifier(svm.SVC(kernel='linear', probability=True,
                                 random_state=random_state))
y_score = classifier.fit(X_train, y_train).decision_function(X_test)

# Compute ROC curve and ROC area for each class
fpr = dict()
tpr = dict()
roc_auc = dict()
for i in range(n_classes):
    fpr[i], tpr[i], _ = roc_curve(y_test[:, i], y_score[:, i])
    roc_auc[i] = auc(fpr[i], tpr[i])

# Compute micro-average ROC curve and ROC area
fpr["micro"], tpr["micro"], _ = roc_curve(y_test.ravel(), y_score.ravel())
roc_auc["micro"] = auc(fpr["micro"], tpr["micro"])

I'm actually looking for a Javascript solution (using https://github.com/mljs/performance) so I haven't implemented it with the above library, but it's been the most illuminating example I found so far.