missbgm 0.1.0

Missingness-aware data imputation with Bayesian generative modeling (MissBGM).

MissBGM

Missingness-aware data imputation with Bayesian generative modeling and uncertainty quantification.

MissBGM is a missingness-aware Bayesian generative model for imputing data with non-ignorable missingness (Missing Not At Random, MNAR). It jointly models the data-generating process and the missingness mechanism, enabling both accurate imputations and principled uncertainty quantification via posterior sampling.

Highlights

• Joint modeling data-generating process and missingness mechanisms
• MAP imputation for fast point estimates
• Posterior sampling (MCMC) for uncertainty quantification and prediction intervals
• Bridging Bayesian deep learning and missing data imputation

Model overview

Figure 1: MissBGM model overview. If the PDF does not render inline on GitHub, click the figure to open fig1.pdf.

Installation

Create a conda environment

conda create -n missbgm python=3.12 -y
conda activate missbgm

Install from source

git clone https://github.com/liuq-lab/MissBGM.git && cd MissBGM
pip install -e .

Install from PyPI

pip install missbgm

Dependencies

This project is tested with:

• Python 3.12
• TensorFlow 2.18.0
• TensorFlow Probability 0.25.0
• tf-keras==2.18.0
• numpy, pandas, pyyaml, scikit-learn

Quickstart (Python API)

Below is a minimal end-to-end example that mirrors the logic in main.py: simulate MNAR data, train MissBGM, read MAP imputations, and then draw posterior samples to produce uncertainty intervals.

import yaml

from missbgm.models import MissBGM
from missbgm.datasets import simulate_mnar_oracle_data

# Load experiment configuration
params = yaml.safe_load(open("configs/Sim_MNAR_oracle.yaml", "r"))

# Simulate MNAR-oracle data (synthetic benchmark)
data = simulate_mnar_oracle_data(
    n_samples=500,
    x_dim=50,
    n_anchor=5,
    missing_rate=0.5,
    alpha=0.05,
    seed=123,
)

# Instantiate the MissBGM model with the configuration
model = MissBGM(params, random_seed=42)

# Train the MissBGM model
model.fit(data=data["x_obs"], mask=data["mask"], x_true=data["x_full"], verbose=1)

# Get the MAP imputation
map_imputed = model.x_map_imputed_

# Make posterior predictions with uncertainty quantification
mcmc_imputed, intervals = model.predict(
    data=data["x_obs"],
    mask=data["mask"],
    x_true=data["x_full"],
    alpha=0.05,
    n_mcmc=1000,
    burn_in=1000,
    step_size=0.1,
    num_leapfrog_steps=5,
    seed=42,
    verbose=1,
)

Reproducing experiments with main.py

main.py is the primary experiment entrypoint. It reads a YAML config (-c) and runs:

• Synthetic simulation when dataset: Sim_MNAR_oracle
• Real-data benchmark otherwise (the code normalizes data, trains, and evaluates RMSE on missing entries)

1) Synthetic MNAR simulation

Use configs/Sim_MNAR_oracle.yaml:

python main.py -c configs/Sim_MNAR_oracle.yaml

2) Real-data benchmarks (4 datasets)

The following configs are provided:

• configs/Real_Wine.yaml (UCI Wine)
• configs/Real_Concrete.yaml (UCI Concrete)
• configs/Real_Breast.yaml (UCI Breast Cancer Wisconsin Original)
• configs/Real_Gisette.yaml (UCI Gisette (high-dimensional))

Run each dataset:

python main.py -c configs/Real_Wine.yaml
python main.py -c configs/Real_Concrete.yaml
python main.py -c configs/Real_Breast.yaml
python main.py -c configs/Real_Gisette.yaml

If you want to re-run even when cached staged files / outputs exist, add --force:

python main.py -c configs/Real_Wine.yaml --force

What main.py does (pipeline summary)

For each run:

• Loads config from YAML (e.g., configs/Real_Wine.yaml)
• Prepares data
  • Synthetic: calls simulate_mnar_oracle_data(...)
  • Real: calls prepare_real_benchmark_data(dataset_name, missing_rate=..., seed=..., force=...)
• Trains MissBGM.fit(...)
• Computes MAP imputation via model.x_map_imputed_
• Draws posterior samples via model.predict(...) to get:
  • mcmc_imputed: posterior mean imputation (from samples)
  • intervals: prediction intervals (controlled by alpha)
• Reports metrics such as RMSE on missing entries

Project structure

missbgm/
  datasets/        # synthetic simulators + real dataset staging / preprocessing
  models/          # BGM base model + MissBGM
  utils/           # masking, baselines, metrics, prediction intervals
configs/           # YAML configs for experiments
main.py            # experiment entrypoint (simulation + real benchmarks)

Citation

If you use MissBGM in your research, please cite our paper (arXiv link coming soon):

@misc{missbgm2026,
  title        = {MissBGM: Missingness-Aware Bayesian Generative Modeling for MNAR Imputation},
  author       = {TBD},
  year         = {2026},
  eprint       = {TBD},
  archivePrefix= {arXiv},
  primaryClass = {stat.ML}
}