kernelmoments 0.1.0

Data visualisation and estimation tools based on kernel regression

kernelmoments

Estimate and visualize conditional moments — means, variances, covariances, and correlations — using adaptive kernel regression. Built on kernelboost's KernelTree, which partitions data and fits local Nadaraya-Watson estimators with automatic bandwidth selection.

Designed for interactive data exploration. Typically fits in under a second on moderately sized datasets (a few thousand rows on CPU, more with GPU acceleration).

For a full tutorial and examples, see the GitHub repository.

Installation

pip install kernelmoments

For GPU acceleration (CUDA 12):

pip install kernelmoments[gpu]

Quick start

Plot from NumPy arrays with plot_relationship

from kernelmoments import plot_relationship

# Conditional mean with +-1.96*sqrt(Var) prediction bands
result = plot_relationship(x, y, moment="mean", bands=True)

# Conditional variance
result = plot_relationship(x, y, moment="variance")

# Conditional correlation between y and z given x
result = plot_relationship(x, y, z=z, moment="correlation")

Plot from DataFrames with Plotter

from kernelmoments import Plotter

p = Plotter(df, n_sample=5000)  # optional subsampling for faster fitting
p.fit(x="age", y="income", z="spending")  # pre-fit all moments

p.plot(x="age", y="income")                              # conditional mean
p.plot(x="age", y="income", moment="variance")            # conditional variance
p.plot(x="age", y="income", z="spending", moment="correlation")  # conditional correlation

Estimators directly

from kernelmoments import MeanEstimator, VarianceEstimator, CovarianceEstimator

mean_est = MeanEstimator().fit(X, y)
y_hat = mean_est.predict(X_new)

var_est = VarianceEstimator().fit(X, y)
var_hat = var_est.predict(X_new)

cov_est = CovarianceEstimator().fit(X, y, z)
cov_hat = cov_est.predict(X_new)
cov_est.fit_correlation()
corr_hat = cov_est.predict_correlation(X_new)

All estimators follow the scikit-learn fit / predict pattern. Constructor parameters (bandwidth bounds, kernel type, tree depth, etc.) are forwarded to KernelTree.

License

MIT