Useful Metrics for 2-Class Classification Problems

good-bad-opposite-choice

In machine learning, two-class classification problems are when we want to predict the class of an event given some attributes (e.g. spam/not spam, sick/not sick, pregnant/not pregnant, cancer/not cancer, etc.). We keep track of the prediction accuracy by constructing a confusion matrix and make a tally of the number of true positives, false positives, false negatives, and true negatives.

confusion-matrix

Here are some useful statistical metrics for two-class classification problems:

Accuracy: The proportion of all instances that are correctly predicted.

  • Accuracy = (TP + TN)/(TP + TN + FP + FN)

Specificity: The proportion of actual negatives (i.e. 0) that were correctly predicted as such (e.g., the percentage of people who are healthy who were correctly predicted as being healthy).

  • Specificity = (TN/(TN + FP))

Precision: The proportion of all positive predictions that were correct (e.g. the percentage of people who were predicted to have the disease and actually had the disease).

  • Precision = (TP/(TP + FP))

Recall: The proportion of actual positives that were correctly identified as such (e.g., the percentage of people who have the disease who were correctly predicted to have the disease).

  • Recall = (TP/(TP + FN))

Negative Predictive Value: The proportion of all negative predictions that were correct (e.g. the percentage of people who were predicted to be healthy who actually are healthy).

  • Negative Predictive Value = (TN/(TN + FN))

Miss Rate: The proportion of actual positives that were predicted to be negative (e.g. the percentage of people who have the disease that were predicted to be healthy).

  • Miss Rate = (FN/(FN + TP))  

Fall-Out: The proportion of actual negatives that were predicted to be positive (e.g. the percentage of people who are healthy that were predicted to have the disease).

  • Fall-Out = (FP/(FP + TN))

False Discovery Rate: The proportion of all positive predictions that were incorrect (e.g. the percentage of people who were predicted to have the disease that are actually healthy).

  • False Discovery Rate = (FP/(FP + TP))

False Omission Rate: The proportion of all negative predictions that were incorrect (e.g. the percentage of people who were predicted to be healthy that actually have the disease).

  • False Omission Rate = (FN / (FN + TN))

F1 Score: Measures prediction accuracy as a function of precision and recall. An F1 score of 1 is good…perfect precision and recall. An F1 score of 0 is bad, the worst-case precision and recall.

  • F1 Score = ((2TP)/(2TP + FP + FN))