Both approaches are not mutually exclusive
Via the Loss function
A loss function will try to:
- Minimize Calibration error
- Maximize sharpness
Via a post processing step.
Another way to fix calibration errors, is in a post processing step. Find some function, that maps the predictions onto more calibrated predictions. Computationally less expensive than to constantly tweak the loss function.
Example: Platt scaling method:
apply sigmoidal transformation to the models output whose parameters are learnt during maximum likelihood estimation.
