zzsi / gbt

gbt is a library for gradient boosted trees with minimal coding required. It is a thin wrapper around lightgbm.

What you need:

• a pandas dataframe,
• the target column to predict on,
• categorical feature columns (can be empty),
• numerical feature columns (can be empty, but you should have at least one categorical or numerical feature),
• params_preset: "binary", "multiclass", "mape", "l2" to specify what type of prediction objective and default hyperparameters to use.

You don't need to (though you are welcome to):

• normalize the numerical feature values
• construct the encoder to one-hot encode categorical features
• manage saving of artifacts for above feature transformation
• implement evaluation metrics

Install

pip install gbt

Quickstart

train_df = pd.DataFrame(
    {
        "a": [1, 2, 3, 4, 5, 6, 7],
        "b": ["a", "b", "c", None, "e", "f", "g"],
        "c": [1, 0, 1, 1, 0, 0, 1],
        "some_other_column": [0, 0, None, None, None, 3, 3],
    }
)

class DatasetBuilder:
    def training_dataset(self):
        return train_df
    
    def testing_dataset(self):
        return train_df  # TODO: use an actual test dataset

TrainingPipeline(
    params_preset="binary",  # one of mape, l2, binary, multiclass
    params_override={"num_leaves": 10},
    label_column="c",
    val_size=0.2,  # fraction of the validation split
    categorical_feature_columns=["b"],
    numerical_feature_columns=["a"],
).fit(DatasetBuilder())

Output of training

The output includes:

• the model file (decision trees and boosting coefficients),
• a feature transformer state file (in JSON)

Using the model to predict on new data

Under construction.