andruekonst / forest-self-attention

Forest Self-Attention

Installation

For the package installation, first install all the requirements and then install the sa_forst package.

$ pip install -r requirements.txt
$ python setup.py install

Usage

Self-Attention Forest model has a scikit-learn like interface. It is extended with optimize_weights method which can be executed with the same training data as used for an underlying forest training, or with a new data set.

Code example for model instantiation:

from sa_forest import (
    SAFParams,
    SelfAttentionForest,
    ForestKind,
    TaskType,
)

model = SelfAttentionForest(
    SAFParams(
        kind=ForestKind.EXTRA,
        task=TaskType.REGRESSION,
        eps=0.9,
        tau=1.0,
        gamma=0.9,
        sa_tau=1.0,
        sa_dist='y',
        forest=dict(
            n_estimators=200,
            max_depth=None,
            min_samples_leaf=5,
            random_state=12345,
        ),
    )
)

After the underlying forest should be trained:

model.fit(X_train, y_train)

And then weights are optimized:

model.optimize_weights(X_train, y_train)

In order to estimate weights optimization impact scores for model.predict_original(X_val) and model.predict(X_val) could be compared.