I can't tell you how many times I've plotted a roc curve for a multi-class problem from scratch. Too many times. I decided to make this gist to demonstrate how to implement a multi-class ROC (Receiver Operator Characteristic) plot in the most simple manner possible using Python.

multi-roc_curve.py

## import any sklearn models and collect predictions / probabilities beforehand

import matplotlib.pyplot as plt
from cycler import cycler

## Line color config -- rather than create a structure with a finite color palette, use your own to cycle through a list.
default_cycler = (cycler(color=['r', 'g', 'b', 'y']) + 
                  cycler(linestyle=['-', '--', ':', '-.']))

plt.rc('axes', prop_cycle = default_cycler)

## Set confusion metrics per class
fpr, tpr, thresh = {}, {}, {}
for index, class_name in enumerate(pipe_model.classes_):
    fpr, tpr, threshold = roc_curve(y_encoded, y_hat_prob[:,index], pos_label=index)
    plt.plot(fpr, tpr, label = f"Class - {class_name}")
    
plt.title('Multiclass ROC curve')
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive rate')
plt.legend(loc='best')

Also, encoder is an instance of sklearn.preprocessing.LabelEncoder.