svemnet — Self-Validated Ensemble Models (SVEM) for Python

SVEM regression for small-sample design of experiments (DOE): lasso / elastic-net and forward-selection ensembles tuned by validation-weighted information criteria, with bootstrap prediction intervals.

svemnet is the Python implementation of the Self-Validated Ensemble Model method (Lemkus, Gotwalt, Ramsey, and Weese 2021) by the author of the R package SVEMnet. SVEM is designed for the small-n, wide-model regime typical of designed experiments — mixture formulations, process optimization, chemometrics — where cross-validation is unstable: instead of holding data out, every run appears in both the training and validation roles through anti-correlated fractional random weights, and predictions are ensembled across bootstrap replicates.

Install

pip install svemnet            # core: numpy + scikit-learn
pip install "svemnet[formula]" # + R-style formula interface (formulaic, pandas)

Quickstart (formula interface, closest to R / JMP)

import pandas as pd
import svemnet

df = pd.read_csv("experiment.csv")  # columns: y, X1, X2, X3, ...

# SVEM with lasso / elastic-net base learners (the SVEMnet default)
model = svemnet.svem("y ~ (X1 + X2 + X3)**2 + I(X1**2)", df, seed=1)
preds = model.predict(new_df)

# SVEM with forward-selection base learners
model = svemnet.svem("y ~ (X1 + X2 + X3)**2", df, method="forward", seed=1)
model.selection_frequencies          # per-term bootstrap selection rates
model.coef_table()                   # coefficients + % bootstraps nonzero

# Bootstrap percentile intervals
out = model.predict(new_df, interval=True, level=0.90)
out["fit"], out["lwr"], out["upr"]

# Deterministic forward selection by AICc (single-model benchmark)
bench = svemnet.forward_aicc("y ~ (X1 + X2 + X3)**2", df)
bench.selected_terms

Categorical predictors: mark them as pandas Categorical (or leave them as strings) and they are treatment-coded automatically, like R factors. Their contrast columns enter and leave forward-selection paths together as whole effects.

(X1 + X2 + X3)**2 expands to main effects plus two-way interactions, matching R's (X1 + X2 + X3)^2. The helper svemnet.response_surface_formula("y", continuous=["X1","X2"], nominal=["F"]) builds standard response-surface formulas.

Quickstart (scikit-learn interface)

from svemnet import SVEMRegressor

est = SVEMRegressor(method="forward", n_boot=200, random_state=0)
est.fit(X, y)                        # numeric matrix, no intercept column
y_pred = est.predict(X_new)
y_pred, intervals = est.predict_interval(X_new, confidence_level=0.9)

SVEMRegressor is a conformant scikit-learn estimator: it works in Pipeline, cross_val_score, and GridSearchCV, and handles categoricals the scikit-learn way (ColumnTransformer / OneHotEncoder).

Coming from R SVEMnet or JMP

The Gaussian fitting mathematics match the published SVEM conventions used by the R package: anti-correlated FRW train/validation weights rescaled to mean one, validation-weighted wSSE/wAIC/wBIC path selection with the Kish effective-sample-size admissibility guardrail, intercept-only fallback, coefficient averaging across the ensemble, and optional linear debiasing. The forward-selection engines here and the R functions forward_aicc() / svem_forward() (SVEMnet ≥ 3.5.0) are ports of the same reference implementation and validate against each other numerically. The lasso / elastic-net base learners use scikit-learn's coordinate descent rather than glmnet, so individual path fits can differ slightly in lambda gridding; ensemble predictions agree closely in practice.

Scope. This package is intentionally minimal: Gaussian responses, lasso/elastic-net and forward-selection base learners, prediction with bootstrap uncertainty. For binomial responses, whole-model significance testing, mixture-constrained random-search optimization, and Thompson-sampling batch design, use the R package SVEMnet; JMP Pro users have SVEM built into Generalized Regression.

References

• Lemkus, T., Gotwalt, C., Ramsey, P., & Weese, M. L. (2021). Self-Validated Ensemble Models for design of experiments. Chemometrics and Intelligent Laboratory Systems, 219, 104439. doi:10.1016/j.chemolab.2021.104439
• Karl, A. T. (2024). A randomized permutation whole-model test for SVEM. Chemometrics and Intelligent Laboratory Systems, 249, 105122. doi:10.1016/j.chemolab.2024.105122
• Karl, A. T. (2026). SVEMnet: Self-Validated Ensemble Models in R. Chemometrics and Intelligent Laboratory Systems. doi:10.1016/j.chemolab.2026.105660
• Xu, L., Gotwalt, C., Hong, Y., King, C. B., & Meeker, W. Q. (2020). Applications of the fractional-random-weight bootstrap. The American Statistician, 74(4), 345–358. doi:10.1080/00031305.2020.1731599

If you use this package in published work, please cite Lemkus et al. (2021) for the method and Karl (2026) for the software.

License

MIT. Copyright (c) 2026 Andrew T. Karl.