Probability calibration is the process of calibrating an ML model to return the true likelihood of an event. This is necessary when we need the probability of the event in question rather than its classification.
Image that you have two models to predict rainy days, Model A and Model B. Both models have an accuracy of 0.8. And indeed, for every 10 rainy days, both mislabelled two days.
But if we look at the probability connected to each prediction, we can see that Model A reports a probability of 80%, while Model B of 100%.
This means that model B is 100% sure that it will rain, even when it will not, while model A is only 80% sure. It appears that model B is overconfident with its prediction, while model A is more cautious.
And it’s this level of confidence in predictions that makes Model A a more reliable model with respect to Model B; Model A is better despite the two models having the same accuracy.
Model B offers a more yes-or-no prediction, while Model A tells us the true likelihood of the event. And in real life, when we look at the weather forecast, we get the prediction and its probability, leaving us to decide if, for example, a 30% risk of rain is acceptable or not.Continue reading “How Probability Calibration Works”