Metrics for classification (label prediction and probability)

Having discussed the metrics for regression problems, we are going now to illustrate the metrics for classification problems, starting from the binary classification problems (when you have to predict between two classes), moving to the multi-class (when you have more than two classes), and then to the multi-label (when the classes overlap).

Accuracy

When analyzing the performance of a binary classifier, the most common and accessible metric that is used is accuracy. A misclassification error is when your model predicts the wrong class for an example. The accuracy is just the complement of the misclassification error and it can be calculated as the ratio between the number of correct numbers divided by the number of answers:

This metric has been used, for instance, in Cassava Leaf Disease Classification (https://www.kaggle.com/c/cassava-leaf-disease-classification) and Text Normalization Challenge...