Issue on param class_weights in CatBoostClassifier

[erko]

Have some trouble using class weights for binary classes. Can't share dataset because of confidentiality, but looks like problem is in algorithm itself. Tried to track prob in source, but ended up at file catboost/app/mode_fit.cpp - can't find where is problem.

Using value [1, 13], 13 for positive class works (AUC ~ 0.91) but slightly worse than training without classweighting. XGBoost performs better ~0.93 with [1, 15] class weightings.
But, training process stuck on AUC 0.5. if I set 15 for positive class:

class_weights = np.array([1., 15.])
# class_weights = (class_weights / np.sum(class_weights))
class_weights = class_weights.tolist()
print(class_weights)

random_seed = 123456
model = cgb.CatBoostClassifier(iterations=500, learning_rate=0.1, depth=5, l2_leaf_reg=2, rsm=1, class_weights=class_weights,
                               auto_stop_pval=1e-4, random_seed=random_seed, eval_metric='AUC')

model.fit(train_X, train_Y, eval_set=(test_X, test_Y), verbose=True, plot=True)

[annaveronika]

Could you provide a reproducible example of this behavoir?

[annaveronika]

closing because of inactivity