blockwise 0.1.0

Blockwise Reduced Modeling (BRM) for tabular data with blockwise missing patterns

blockwise

Blockwise Reduced Modeling (BRM) for tabular data with blockwise missing patterns — a scikit-learn-compatible estimator.

Install

pip install blockwise

Quickstart

import pandas as pd
from sklearn.ensemble import GradientBoostingRegressor
from blockwise import BRM, simulate_blockwise_missing, datasets

bike = datasets.load_bike()
bike_miss = simulate_blockwise_missing(
    bike,
    blocks=[["windspeed", "hum", "weekday"],
            ["hr", "temp", "weathersit"]],
    prop_missing=0.30,
    noise=0.05,
)

X = bike_miss.drop(columns=["cnt"])
y = bike_miss["cnt"]

brm = BRM(estimator=GradientBoostingRegressor()).fit(X, y)
y_hat = brm.predict(X)

BRM is learner-agnostic: pass any estimator with fit(X, y) / predict(X) (and optionally predict_proba(X) for classification). Each block's model is a fresh clone of estimator.

What BRM does

BRM partitions the training data into overlapping subsets based on per-row feature-missing patterns, pre-trains one model per subset on only the observed columns of that subset, and at prediction time routes each test row to the subset model whose missingness pattern most closely matches.

See notebooks/ for worked examples on the bike (regression), adult (binary classification), and house (regression) datasets.

Citation

If you use this package, please cite the paper that introduced the method:

Srinivasan, K., Currim, F., and Ram, S. (2025). A Reduced Modeling Approach for Making Predictions With Incomplete Data Having Blockwise Missing Patterns. INFORMS Journal on Data Science.

A machine-readable CITATION.cff is included at the repo root.

License

GPL-3.0-or-later. See LICENSE.