Skip to main content

Table 15 Performance Metrics for classification problem

From: Predicting academic success in higher education: literature review and best practices

Performance measures

How to express them

Interpretation

When to use

Accuracy

\( \frac{TP+ TN}{TP+ TN+ FP+ FN} \)

The number of all correct predictions made by the algorithm over all type of predictions made.

If the data is almost balanced.

Recall (Sensitivity/TP rate)

\( \frac{TP}{TP+ FN} \)

The proportion of successful students that classified correctly as “successful”, for all successful students

To concentrate on minimizing FN.

Precision

\( \frac{TP}{TP+ FP} \)

The proportion of successful students that classified correctly as “successful”, for all students predicted by the algorithm as a “successful” student.

To concentrate on minimizing FP.

Specificity (TN rate)

\( \frac{FP}{TN+ FP} \)

the proportion of non-successful students that are incorrectly considered as successful students, for all non-successful students.

To identify negative results.

F-Measure

\( \frac{2\times Precision\times Recall}{Precision+ Recall} \)

How precise your classifier is, as well as how robust it is

To find a balance between recall and precision.

ROC curve

Plotted at TP rate vs. FP rate where the TP rate is on the Y axis and the FP rate is on the X axis.

The area under the curve (AUC):

• If near to the 1, means the model has high class separation capacity.

• If near to the 0, means the model has no class separation capacity.

Used as a summary of the model’s skill.