optuna-fast-fanova

optuna-fast-fanova provides Cython-accelerated version of FanovaImportanceEvaluator.

n_trials	n_params	n_trees	fANOVA (Optuna)	fast-fanova
1000	32	64	71.431s	2.963s (-95.9%)
1000	8	64	92.307s	2.315s (-97.5%)
1000	2	64	52.295s	1.297s (-97.5%)
100	32	64	1.668s	0.306s (-81.6%)
100	8	64	1.652s	0.138s (-91.7%)
100	2	64	1.242s	0.095s (-92.4%)

The benchmark script was run on my laptop (Macbook M1 Pro) so the times should not be taken precisely.

Installation

Supported Python versions are 3.7 or later.

$ pip install optuna-fast-fanova

Please note that this library depends on the scikit-learn's C-API (Cython pxd files). However, its ABI may contain breaking changes, even in patch releases. If you install optuna-fast-fanova with scikit-learn v1.1.1 and then upgrade scikit-learn to v1.1.2, optuna-fast-fanova will not work. Please reinstall optuna-fast-fanova if you update scikit-learn.

Usage

Usage is like this:

import optuna
from optuna_fast_fanova import FanovaImportanceEvaluator


def objective(trial):
    x = trial.suggest_float("x", -10, 10)
    y = trial.suggest_int("y", -10, 10)
    return x ** 2 + y


if __name__ == "__main__":
    study = optuna.create_study()
    study.optimize(objective, n_trials=1000)

    importance = optuna.importance.get_param_importances(
        study, evaluator=FanovaImportanceEvaluator()
    )
    print(importance)

You can use optuna-fast-fanova in only two steps.

1. Add an import statement: from optuna_fast_fanova import FanovaImportanceEvaluator.
2. Pass a FanovaImportanceEvaluator() object to an evaluator argument of get_param_importances() function.

How to cite fANOVA

This is a derived work of https://github.com/automl/fanova. For how to cite the original work, please refer to https://automl.github.io/fanova/cite.html.