MissLearn

No imputation. No listwise deletion. No fake data.

sklearn-style estimators with native missing-data support via Full Information Maximum Likelihood (FIML).

Pre-release. Version 0.9.2 is complete and tested (2,658 tests, 95% coverage) but has not yet been used by anyone outside its author. If you are reading this because you were asked to try it, that is what you are being asked to try: whether it works on your data, and whether the documentation tells you what you need. Please open an issue for anything that surprises you, including anything the guides fail to explain. There are templates under Issues, and CONTRIBUTING.md says what makes a report easy to act on.

MissLearn models accept NaN values directly in fit and predict. Every observation contributes exactly the information it contains: an observation with three of five predictors observed contributes a three-dimensional likelihood, and no value is ever invented, imputed, or assumed. The aim is simple; on incomplete data, match or beat the conventional strategies of dropping rows, dropping columns, or imputing.

All estimators follow the scikit-learn API (fit / predict / score / get_params / set_params) and work with sklearn pipelines and cross-validation utilities that do not internally strip NaN values. Pandas DataFrames and Series are accepted transparently.

Features

Model families

Family	Estimators	Task
MissLinear	MissLinear	FIML linear regression (joint multivariate normal)
MissLogistic	MissLogistic	FIML logistic regression
MissRidge	MissRidgeRegressor, MissRidgeClassifier, MissRidge (auto)	Ridge-penalized FIML
MissLASSO	MissLASSORegressor, MissLASSOClassifier, MissLASSO (auto)	LASSO-penalized FIML
MissNeighbors	MissNeighborsRegressor, MissNeighborsClassifier, MissNeighbors (auto)	K-nearest neighbours FIML
MissBayes	MissBayesRegressor, MissBayesClassifier, MissBayes (auto)	Naive Bayes FIML
MissSupport	MissSupportRegressor, MissSupportClassifier, MissSupport (auto)	Support vector FIML
MissGaussian	MissGaussianRegressor, MissGaussianClassifier, MissGaussian (auto)	Gaussian Process (marginalized kernel, exact Bayesian intervals). O(n^3): see the cost note under Benchmarks before using it on more than a few thousand rows
MissMixed	MissMixedRegressor, MissMixedClassifier, MissMixed (auto)	Random-intercept LME / GLMM for grouped and longitudinal data
MissEnsemble	MissEnsemble	Bootstrap-aggregated ensemble of MissLearn models, homogeneous or heterogeneous with weights and OOB scores
MissMulticlass	MissMulticlass	One-vs-Rest multi-class wrapper for any binary MissLearn classifier
MissPreprocessor	MissPreprocessor, prefit_check	Validation, FIML compatibility checking, categorical encoding with NaN preservation

All models accept copula=True (or 'auto') to apply a marginal Gaussian copula transform for skewed or heavy-tailed features. The transform assumes continuous margins, so columns with fewer than three distinct observed values are passed through untouched rather than mapped, and 'auto' decides from the margins alone. It is off by default: applying it to one arm of a comparison and not the others measures the transform rather than the missing-data treatment.

Tools

Tool	Purpose
MissImputer	Multiple imputation by draws from the FIML-estimated joint MVN; Rubin's rules combiner for downstream models that cannot accept NaN
MissDiagnostic	Missingness mechanism assessment: Little's MCAR test, MAR plausibility, pattern summary, missingness correlations
MissRecommender	Evidence-based model triage: ranks the families for a given incomplete dataset with the reasoning attached, flags columns to drop rather than impute, and names the required follow-up analyses
MissExplainer	SHAP explainability using the FIML model as the exact coalition value function; value SHAP and missingness SHAP (exact 2^p for p<=15, KernelSHAP above)
MissSensitivity	MNAR sensitivity analysis via delta-adjustment with tipping-point deltas
CV utilities	MissKFold, MissStratifiedKFold, miss_cross_val_score, miss_cross_validate: NaN-safe splitters and scorers with full NaN-in-y support

Installation

Requires Python 3.9 or later.

pip install git+https://github.com/amaxiom/MissLearn.git

To work from a clone, which is what you want if you plan to run the examples or the test suite:

git clone https://github.com/amaxiom/MissLearn.git
cd MissLearn
pip install -e ".[all]"

[all] adds pandas, matplotlib and the gradient-boosted tree learners that can be used as MissEnsemble members. The package itself needs only numpy, scipy and scikit-learn; pandas support is duck-typed rather than a dependency, so DataFrames work if pandas is present and nothing breaks if it is not. Add [test] for pytest and hypothesis if you want to run the suites.

Quick start

import numpy as np
from MissLearn import MissLinear

X = np.array([[1.0, np.nan], [2.0, 3.0], [np.nan, 4.0], [1.5, 2.5]])
y = np.array([0.5, 1.2, np.nan, 0.9])          # NaN in y is fine too

model = MissLinear().fit(X, y)                   # no imputation, no dropped rows
y_pred = model.predict(X)                        # NaN-in-X handled natively
lo, hi = model.predict_interval(X, alpha=0.05)   # 95% intervals; wider when more
model.summary()                                  # features are missing

Benchmarks

Headline results (July 2026): synthetic data, 25% MAR missingness, 5-fold CV, against Drop Rows / Drop Columns / Mean imputation / KNN imputation / MICE baselines.

Classification accuracy (medium dataset): MissLearn vs best baseline:

Model	MissLearn	Best baseline
MissBayes	0.939	0.889
MissSupport	0.940	0.927
MissNeighbors	0.933	0.921
Linear family (Logistic/Ridge/LASSO)	0.914	0.911

Regression R²: linear family at parity (~0.70); MissNeighbors 0.52 vs 0.50; MissSupport 0.64 vs 0.17.

Fast: classifier fits in ~0.2 s and the LASSO regressor in ~1.3 s at n=600, p=8. The correctness suites are 2,658 tests (1,600 unit, 1,042 conformance, 16 property) at 95.0% coverage, and run in about 25 minutes on an idle machine, considerably longer under load. Full per-family results, sweeps and plots are in benchmarks/, described in benchmarks/BENCHMARKS.md.

The exception is MissGaussian. Exact Gaussian-process inference is O(n^3) in the number of rows, and this family is the one place where that cost is visible rather than theoretical. Measured at n = 90, p = 3 on an idle machine, a single fit takes about 1.2 s for MissGaussianRegressor and about 11 s for MissGaussianClassifier. The classifier is roughly nine times the regressor because it runs a Laplace mode-finder of up to max_iter_newton Newton steps inside every objective evaluation, and the optimiser repeats that for each of 1 + n_restarts restarts. Cubed in n and multiplied by four nested loops, a few thousand rows will look like a hang.

Nothing in the estimator itself caps the row count. MissRecommender will veto the family above gp_max_n (default 1,000) and say why, and the benchmark harness applies its own max_n = 900, but neither of those is consulted when you call MissGaussian directly, so on your own data the ceiling is yours to set. If a fit is taking longer than you expected, in rough order of how much they buy: lower n_restarts (the default of 3 means four optimiser runs), subsample the rows, lower max_iter_newton on the classifier, or use another family. MissSupport and MissNeighbors cover much of the same ground with a kernel or a distance instead of a full posterior.

Documentation

Document	Contents
docs/USER_GUIDE.md	Practical guide to every model and tool
docs/COMPUTATIONAL_GUIDE.md	Complexity, scaling, and performance guidance
docs/INTERPRETATION_GUIDE.md	Interpreting coefficients, intervals, and SHAP output
docs/MissLearn_User_Guide.pdf	Combined user guide (PDF)
docs/METHODS_GUIDE.md	Statistical methodology: missing-data theory, FIML derivations, model design, inference
examples/EXAMPLES.md	Guide to the worked example notebooks
tests/TEST.md	Guide to the unit, performance, and benchmark test suites
benchmarks/BENCHMARKS.md	Guide to the synthetic benchmarks, and what makes each comparison fair
CONTRIBUTING.md	Development setup, style, and the deprecation policy
CODE_OF_CONDUCT.md	Contributor Covenant 2.1, and how to report a concern
CITATION.cff	Citation metadata, read by GitHub and by reference managers
CHANGELOG.md	Release history
docs/ROADMAP.md	Planned work

Requirements

• Python 3.9+
• numpy, scipy, scikit-learn
• Optional: pandas (DataFrame support is duck-typed; no hard dependency), matplotlib (plots)

Repository layout

MissLearn/
├── MissLearn/        # the package (models, tools, utilities); single source of truth
├── benchmarks/       # per-family benchmark and sweep notebooks + generators
├── docs/             # methodology, user guides
├── examples/         # worked example notebooks (see examples/EXAMPLES.md)
├── tests/            # unit / conformance / performance suites (see tests/TEST.md)
├── pyproject.toml    # packaging, builds directly from MissLearn/
├── CHANGELOG.md
└── LICENSE

There is deliberately no separate staging copy of the package. One existed and drifted: three library fixes made in a single session reached only the source tree and would have shipped missing. A second copy that has to be kept in sync by hand is a defect in itself, so packaging now builds from MissLearn/ directly. The previous staging area, including the verified 0.9.1 artefacts, is kept under _archive/ rather than deleted.

Citation

If you use MissLearn in published work, please cite the software. There is no paper yet; this section will be updated when there is one.

@software{barnard2026misslearn,
  author  = {Barnard, Amanda S.},
  title   = {{MissLearn}: full-information machine learning from data with
             meaningful missingness},
  year    = {2026},
  version = {0.9.2},
  url     = {https://github.com/amaxiom/MissLearn},
  note    = {Python package, MIT licence}
}

Or in text:

Barnard, A. S. (2026). MissLearn: full-information machine learning from data with meaningful missingness (Version 0.9.2) [Computer software]. https://github.com/amaxiom/MissLearn

CITATION.cff in the repository root carries the same metadata in a machine readable form, which is what GitHub reads to offer Cite this repository in the sidebar.

License

MIT. See LICENSE.