Catboost for Imbalanced Data Sets

[carlosr29]

Is there a parameter like "scale_pos_weight" in catboost package as there is in the xgboost package in python in order to handle imbalanced classes?

I know there is a parameter called "class_weights", but in the official Documentation (https://tech.yandex.com/catboost/doc/dg/concepts/python-reference_parameters-list-docpage/#python-reference_parameters-list it's not well explained if it helpts for the imbalanced problem, and how to set it.

Thanks in advance.

[annaveronika]

Hi, sorry for the long answer, we'll update documentation and add scale_pos_weight parameter.

[annaveronika]

For now you can use class_weights in the following way: set weight 1 for class 0 and weight scale_pos_weight for class 1. It will be equal to having scale_pos_weight parameter.

[annaveronika]

the parameter is added in last release

[annaveronika]

@carlosr29 We are currently working on improving quality for imbalance datasets with binary classification.
If you could share your dataset with us, it would be very helpful.

[elfwired]

@annaveronika is there a way to work with imbalanced datasets when solving regression problems?

[abhi070493]

@annaveronika Is there a way to downscale predicted scores after using class_weights? (Since I have noticed that the model over-predicts on using class_weights and I need to use the point estimates for my problem)

[The-Gupta]

I'm still getting the error, I tried scale_pos_weight as well-

classifier = CatBoostRegressor(class_weights = [0.8, 0.2])
Traceback (most recent call last):

  File "<ipython-input-49-9098f24b0f97>", line 1, in <module>
    classifier = CatBoostRegressor(class_weights = [0.8, 0.2])

TypeError: __init__() got an unexpected keyword argument 'class_weights'

Could you once check the version, if that could be the problem?

[Evgueni-Petrov-aka-espetrov]

Please use CatBoostClassifier.
Parameter class_weights is meaningful only for Logloss, MultiClass, and MultiClassOneVsAll loss functions, while CatBoostRegressor implies that loss function is RMSE, or MAE, or Quantile, or LogLinQuantile, Poisson, or MAPE.

[Sandy4321]

it is written above Hi, sorry for the long answer, we'll update documentation and add scale_pos_weight parameter
but in the documentation I see scale_pos_weight (alias for: class_weights) https://catboost.ai/docs/concepts/python-reference_parameters-list.html so what you added only synonym?
and still it is not clear how to use it : https://stackoverflow.com/questions/54437646/catboost-precision-imbalanced-classes do have a clear code example,
for example from this video data is inbalanced but model trained as for balanced data https://www.youtube.com/watch?v=xl1fwCza9C8&t=44s code https://github.com/catboost/tutorials/blob/master/events/pydata_moscow_oct_13_2018.ipynb

[annaveronika]

scale_pos_weigh sets weight for objects of the first class. This is equal to setting class_weights to [1, {scale_pos_weight value}]. To deal properly with inbalanced data, you can try experiment with either of these two parameters, or maybe do oversampling.

[sskarkhanis]

Question about class_weights for Multi-class problems

in the documentation for CB,here: https://catboost.ai/docs/concepts/python-reference_parameters-list.html

I see that to pass class_weights, we use a list; the documentation shows the example of binary classification as class_weights=[0.1, 4] which works fine in case of binary classification.

I know, I can pass a list of length equivalent to the #Classes but how does catboost assign these weights to the appropriate label in multi-class context?

I calculated the class weights using the sklearn utils as follows,

from sklearn.utils import class_weight
cw = list(class_weight.compute_class_weight('balanced',
                                             np.unique(df_train['Target']),
                                             df_train['Target']))

and get a list for e.g. [0.5, 4.5, 7.5, 3.4]
if I pass the list as-is, my model performance is worse compared to without the class_weights option.

How do I address this? would it be an option to allow class_weights to accept a dictionary?
e.g. class_weights = { "class_A": 3.5 , "class_B": 4.5, "class_C": 0.5 }

[annaveronika]

@sskarkhanis Could you please create a separate issue about this?