We should use a sigmoid function, like scipy.special.expit, to constrain the FIGS classifier's predict_proba between 0 and 1, instead of np.clip(preds, a_min=0., a_max=1.). This matches what XGBoost does, see binary:logistichere, and is nicer mathematically as it is a continuous function. If changed, we should also make a note of this in the documentation, and any future papers.
Users can further calibrate the output probabilities with CalibratedClassifierCV, or some other method, if they desire.
We should use a sigmoid function, like
scipy.special.expit, to constrain the FIGS classifier'spredict_probabetween 0 and 1, instead ofnp.clip(preds, a_min=0., a_max=1.). This matches what XGBoost does, seebinary:logistichere, and is nicer mathematically as it is a continuous function. If changed, we should also make a note of this in the documentation, and any future papers.Users can further calibrate the output probabilities with
CalibratedClassifierCV, or some other method, if they desire.