Binary classification

Binary classification has two classes, conventionally 1 (positive) and 0 (negative). A single number $\hat{p}=P(y=1\mid \mathbf{x})$ describes the prediction; $P(y=0\mid \mathbf{x})=1-\hat{p}$.

Which class is 'positive' is a choice — usually the rarer or more interesting one (fraud, disease, spam). This matters because precision, recall and F1 are defined relative to the positive class.

Binary problems are often imbalanced: 1% fraud, 99% normal. Accuracy is then misleading. Handle it with suitable metrics (precision, recall, F1, AUC), class weights in the loss, resampling (oversample the minority, undersample the majority), or threshold tuning.