Categorical Data

Note

As of XGBoost 1.6, the feature is experimental and has limited features. Only the Python package is fully supported.

Starting from version 1.5, the XGBoost Python package has experimental support for categorical data available for public testing. For numerical data, the split condition is defined as \(value < threshold\), while for categorical data the split is defined depending on whether partitioning or onehot encoding is used. For partition-based splits, the splits are specified as \(value \in categories\), where categories is the set of categories in one feature. If onehot encoding is used instead, then the split is defined as \(value == category\). More advanced categorical split strategy is planned for future releases and this tutorial details how to inform XGBoost about the data type.

Training with scikit-learn Interface

The easiest way to pass categorical data into XGBoost is using dataframe and the scikit-learn interface like XGBClassifier. For preparing the data, users need to specify the data type of input predictor as category. For pandas/cudf Dataframe, this can be achieved by

X["cat_feature"].astype("category")

for all columns that represent categorical features. After which, users can tell XGBoost to enable training with categorical data. Assuming that you are using the XGBClassifier for classification problem, specify the parameter enable_categorical:

# Supported tree methods are `approx` and `hist`.
clf = xgb.XGBClassifier(tree_method="hist", enable_categorical=True, device="cuda")
# X is the dataframe we created in previous snippet
clf.fit(X, y)
# Must use JSON/UBJSON for serialization, otherwise the information is lost.
clf.save_model("categorical-model.json")

Once training is finished, most of other features can utilize the model. For instance one can plot the model and calculate the global feature importance:

# Get a graph
graph = xgb.to_graphviz(clf, num_trees=1)
# Or get a matplotlib axis
ax = xgb.plot_tree(clf, num_trees=1)
# Get feature importances
clf.feature_importances_

The scikit-learn interface from dask is similar to single node version. The basic idea is create dataframe with category feature type, and tell XGBoost to use it by setting the enable_categorical parameter. See Getting started with categorical data for a worked example of using categorical data with scikit-learn interface with one-hot encoding. A comparison between using one-hot encoded data and XGBoost’s categorical data support can be found Train XGBoost with cat_in_the_dat dataset.

Optimal Partitioning

New in version 1.6.

Optimal partitioning is a technique for partitioning the categorical predictors for each node split, the proof of optimality for numerical output was first introduced by [1]. The algorithm is used in decision trees [2], later LightGBM [3] brought it to the context of gradient boosting trees and now is also adopted in XGBoost as an optional feature for handling categorical splits. More specifically, the proof by Fisher [1] states that, when trying to partition a set of discrete values into groups based on the distances between a measure of these values, one only needs to look at sorted partitions instead of enumerating all possible permutations. In the context of decision trees, the discrete values are categories, and the measure is the output leaf value. Intuitively, we want to group the categories that output similar leaf values. During split finding, we first sort the gradient histogram to prepare the contiguous partitions then enumerate the splits according to these sorted values. One of the related parameters for XGBoost is max_cat_to_onehot, which controls whether one-hot encoding or partitioning should be used for each feature, see Parameters for Categorical Feature for details.

Using native interface

The scikit-learn interface is user friendly, but lacks some features that are only available in native interface. For instance users cannot compute SHAP value directly or use quantized DMatrix. Also native interface supports data types other than dataframe, like numpy/cupy array. To use the native interface with categorical data, we need to pass the similar parameter to DMatrix and the train function. For dataframe input:

# X is a dataframe we created in previous snippet
Xy = xgb.DMatrix(X, y, enable_categorical=True)
booster = xgb.train({"tree_method": "hist", "max_cat_to_onehot": 5}, Xy)
# Must use JSON for serialization, otherwise the information is lost
booster.save_model("categorical-model.json")

SHAP value computation:

SHAP = booster.predict(Xy, pred_interactions=True)

# categorical features are listed as "c"
print(booster.feature_types)

For other types of input, like numpy array, we can tell XGBoost about the feature types by using the feature_types parameter in DMatrix:

# "q" is numerical feature, while "c" is categorical feature
ft = ["q", "c", "c"]
X: np.ndarray = load_my_data()
assert X.shape[1] == 3
Xy = xgb.DMatrix(X, y, feature_types=ft, enable_categorical=True)

For numerical data, the feature type can be "q" or "float", while for categorical feature it’s specified as "c". The Dask module in XGBoost has the same interface so dask.Array can also be used for categorical data.

Miscellaneous

By default, XGBoost assumes input categories are integers starting from 0 till the number of categories \([0, n\_categories)\). However, user might provide inputs with invalid values due to mistakes or missing values in training dataset. It can be negative value, integer values that can not be accurately represented by 32-bit floating point, or values that are larger than actual number of unique categories. During training this is validated but for prediction it’s treated as the same as not-chosen category for performance reasons.

References

[1] Walter D. Fisher. “On Grouping for Maximum Homogeneity”. Journal of the American Statistical Association. Vol. 53, No. 284 (Dec., 1958), pp. 789-798.

[2] Trevor Hastie, Robert Tibshirani, Jerome Friedman. “The Elements of Statistical Learning”. Springer Series in Statistics Springer New York Inc. (2001).

[3] Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, Tie-Yan Liu. “LightGBM: A Highly Efficient Gradient Boosting Decision Tree.” Advances in Neural Information Processing Systems 30 (NIPS 2017), pp. 3149-3157.