Suppose that we have an imbalanced dataset in a binary classification problem, with many more datapoints with a true label of class $1$ than with a true label of class $0($e$.$g$.$ class $1$ is the "majority" class, and class $0$ is the "minority" class$).$ We would like to use a cost$-$sensitive method to train our model, which penalizes the model $3$ times as much for predicting a datapoint is in class $1$ when it is truly in class $0,$ than for predicting a datapoint is in class $0$ when it is truly in class $1($e$.$g$.,$ following the notation from the lecture videos, c$10=3*$c$01).$ The model is of course not penalized for making correct predictions.

In this case, suppose our trained model predicts that a particular test datapoint has probability p$=0\mathrm{.}7$ of being in class $1.$ True$/$False: The model assigns this test datapoint to class $1$ based on the optimal threshold that corresponds to the cost$-$sensitive method.