Classifier Metrics

The metrics used for evaluating machine learning systems are based on the task and the domain.

Classification

For a set of examples $X=(x_1,...x_N)$, for each $x_i$, predict a category $c_i$ from the set $C=c_1,...,c_K$. An example of classification is the object recognition, where given an image, you try to predict the object. If $K=2$, the task is called binary classification task, otherwise a multi-class classification task.

Accuracy

$$Acc=\frac{CorrectPredictions}{N}$$

Ratio of how many predictions the ML system got right. For binary classification, the accuracy can be defined as:

Error Rate

$$ErrorRate=\frac{IncorrectPredictions}{N}=(1-Acc)$$

References

• Perf Metrics