What is the misclassification rate?
The misclassification rate is the share of all predictions that a classifier gets wrong. It is also called the error rate or classification error. For a two-class model it is false positives plus false negatives divided by the total number of cases, which makes it one minus accuracy. For a model with several classes it is the sum of every off-diagonal cell in the confusion matrix divided by the total.
Misclassification rate in practice
A city routes free-text 311 service requests to nine departments, from streets and sanitation to water and parks. In one month, 1,400 of 20,000 requests went to the wrong department, a misclassification rate of 0.07. Each misrouted request sat in the wrong queue until a dispatcher reassigned it. The single figure gave the city a clear target for the next model version, and the confusion matrix showed where to look: most errors were water and sewer requests sent to the streets department.
What to watch for
- The rate counts every mistake the same, so a misrouted water main break weighs no more than a misrouted graffiti report. Assign a cost to each type of error when the difference matters.
- Compare the rate with the null error rate, because on imbalanced data a model can post a low error rate while missing every rare case.
- Report the rate for each class, since a good overall figure can hide a class the model almost never gets right.