README.md

Important

PED-ANOVA is now available at Optuna. Note that the top-performance quantile $\gamma$ in the Optuna version of PED-ANOVA can be modified via baseline_quantile.

Note

See this branch for the IJCAI experiment repository.

PED-ANOVA

This repository is based on the paper PED-ANOVA: Efficiently Quantifying Hyperparameter Importance in Arbitrary Subspaces. Our method provides an easy-to-implement new f-ANOVA algorithm.

Simple example using Sci-Kit Learn

As mentioned in the paper, our method is very easy to implement. Here is an implementation example using Sci-Kit Learn.

from __future__ import annotations

import numpy as np
from sklearn.neighbors import KernelDensity


lower, upper = -5, 5

# Objective function for analysis
def sphere(x: np.ndarray) -> np.ndarray:
    coef = np.array([1, 5])
    x *= coef
    return np.sum(x ** 2, axis=-1)


def collect_observations(n_samples: int = 1000, dim: int = 2) -> tuple[np.ndarray, np.ndarray]:
    # Define the search space and collect observations using random search
    n_samples, dim = 1000, 2
    X = np.random.random((n_samples, dim)) * (upper - lower) + lower
    return X, sphere(X)


class PEDANOVAEvaluator:
    def __init__(self, X_train: np.ndarray, y_train: np.ndarray) -> None:
        # Define the top quantile and compute the local/global PDFs for each dimension
        # Note that we are using Eq. (14) to calculate the KDEs rather tahn Eq. (15), which is a quicker version
        quantile = 0.1
        (n_samples, dim) = X_train.shape
        X_sorted = X_train[np.argsort(y_train)]
        self.local_kdes = {
            f"x{d}": KernelDensity(kernel="gaussian").fit(X_sorted[:int(quantile * n_samples), d, None])
            for d in range(dim)
        }
        self.global_kdes = {
            f"x{d}": KernelDensity(kernel="gaussian").fit(X_sorted[:, d, None])
            for d in range(dim)
        }

    def __call__(self, X: np.ndrray) -> list[float]:
        hpi_list = []
        # Compute the global HPI of each dimension based on Eq. (16)
        for d in range(X.shape[-1]):
            log_pdf_global = self.global_kdes[f"x{d}"].score_samples(X[:, d, None])
            log_pdf_local = self.local_kdes[f"x{d}"].score_samples(X[:, d, None])
            # Equivalent to `(pdf_local / pdf_global - 1) ** 2`
            integrand = np.expm1(log_pdf_local - log_pdf_global) ** 2
            hpi_list.append(float(np.exp(log_pdf_global) @ integrand))

        hpi_sum = sum(hpi_list)
        return [hpi / hpi_sum for hpi in hpi_list]


X, y = collect_observations()
evaluator = PEDANOVAEvaluator(X, y)
# Compute the global HPI of each dimension based on Eq. (16)
dX = np.broadcast_to(np.linspace(lower, upper, 100)[:, None], (100, X.shape[-1]))
# Print the global HPI
print(evaluator(dX))

Note

This example does not use the discretization trick, so it would take much more time to compute the global HPI compared to what we can expect from the KDEs built by Eq. (15) if n_samples is a large number such as $10^9$.

Citations

For the citation, use the following format:

@article{watanabe2023ped,
  title={{PED-ANOVA}: Efficiently Quantifying Hyperparameter Importance in Arbitrary Subspaces},
  author={S. Watanabe and A. Bansal and F. Hutter},
  journal={International Joint Conference on Artificial Intelligence},
  year={2023}
}