The result is different with Pandas 'category' and 'object' types

[khrisanfov]

When I create the Pool by passing Pandas DataFrame and the list of categorical features indexes into it, the result of learning is different when categorical features in DataFrame have types 'category' and 'object'. In the 2nd case, the result is much better, and the importance of categorical features is higher than in the 1st case.

catboost version: 0.14.2
Operating System: Ubuntu 18.04
CPU: Intel Xeon E3

[annaveronika]

Category is not supported yet, please use object for now.

[annaveronika]

From the duplicate issue:

Are SKLEARN API estimators can recognize difference between
'ordinal feature (applying pandas.api.types.CategoricalDtype(data, ordered=True))'
and
'nominal feature (applying pandas.api.types.CategoricalDtype(data, ordered=False))'

[andrey-khropov]

I wrote a quick test and cannot reproduce the difference in results between 'object', 'category' and default column types:

#!/usr/bin/env python3

import sys

import numpy as np

import catboost
from catboost import datasets
from sklearn.model_selection import train_test_split


print ('py version', sys.version)
print ('CatBoost version', catboost.__version__)

(train_df, test_df) = catboost.datasets.amazon()
y = train_df.ACTION
X = train_df.drop('ACTION', axis=1)

X_train, X_valid, y_train, y_valid = train_test_split(X, y, test_size=0.2)

model = catboost.CatBoostClassifier(iterations=30)
model.fit(X_train, y_train, eval_set=(X_valid, y_valid))
preds = model.predict(X_valid)

print ('with category')

model = catboost.CatBoostClassifier(iterations=30)
model.fit(X_train.astype('category'), y_train, eval_set=(X_valid.astype('category'), y_valid))
preds_from_df_with_category = model.predict(X_valid.astype('category'))

assert np.allclose(preds, preds_from_df_with_category)

print ('with object')

model = catboost.CatBoostClassifier(iterations=30)
model.fit(X_train.astype('object'), y_train, eval_set=(X_valid.astype('object'), y_valid))
preds_from_df_with_category = model.predict(X_valid.astype('object'))

assert np.allclose(preds, preds_from_df_with_category)

Maybe you mean that the result is different when you pass DataFrame with columns with 'category' type, but do not specify 'cat_features' parameter of Pool constructor? This is possible because at the moment CatBoost does not treat pandas.DataFrame columns with dtype='category' as categorical features automatically treating them as numeric features by default instead. You still have to specify 'cat_features' parameter if you have any categorical features. We will add special check for this case because it is usually an error.

[khrisanfov]

@andrey-khropov this is old issue, in the last version category type was supported by CatBoost, but I not tested it yet.

[annaveronika]

Closing the issue then, let us know if there will be problems with this type.