Multiclass classification
Performance metrics
We have already discussedbinary classificationat some length, and multiclass classification is similar, so here we will concentrate on the differences. The primary difference is that there are not just two possible outcomes, but three or more.
Assuming that all of the possible outcomes are equally interesting to you, with no special emphasis on any of them, it's reasonable to chooseAccuracyas your performance metric, and to prefer the model with the highestAccuracy.
When there areNpossible outcomes, the confusion matrix hasN x Nelements. The correct predictions are on the diagonal, and incorrect predictions are off-diagonal. Although the summary performance table in theModel Comparisononly mentionsAccuracyandClassification Error, you can still find the analogues ofPrecisionandRecallby looking at thePerformancetab for each individual model.Precisionis computed for each row;Recallis computed for each column.
| Actually A | Actually B | Actually C | Precision | |
|---|---|---|---|---|
| Predicted A | True A | True A / Predicted A | ||
| Predicted B | True B | True B / Predicted B | ||
| Predicted C | True C | True C / Predicted C | ||
| Recall | True A / Actually A | True B / Actually B | True C / Actually C |
In the example below, a multiclass classification problem with three possible outcomes, the model made two wrong predictions, indicated by a red mark in the confusion matrix, when applied to the test set. Consequently, theRecallfor the second column (5/7) and thePrecisionfor the third row (6/8) are less than 100%.
