Package {blockwise}

Type:	Package
Title:	Reduced Modeling for Tabular Data with Blockwise Missingness
Version:	0.1.2
Description:	Supervised learning on tabular data with blockwise missing patterns, using the Blockwise Reduced Modeling (BRM) method of Srinivasan, Currim, and Ram (2025) <doi:10.1287/ijds.2022.9016>. BRM partitions the training data into overlapping subsets based on per-row feature-missing patterns, fits one user-supplied learner per subset with minimal imputation, and at prediction time routes each test instance to the best-matching subset model. The interface is learner-agnostic: any fit-and-predict pair can be plugged in, and convenience specifications are provided for linear models, tree models, random forests, and gradient boosting.
License:	GPL-3
Language:	en-US
Encoding:	UTF-8
LazyData:	true
RoxygenNote:	7.3.3
Depends:	R (≥ 3.6.0)
Imports:	stats, VIM, withr
Suggests:	testthat (≥ 3.0.0), knitr, rmarkdown, rpart, ranger, gbm, ggplot2
VignetteBuilder:	knitr
Config/testthat/edition:	3
URL:	https://github.com/KarAnalytics/blockwise
BugReports:	https://github.com/KarAnalytics/blockwise/issues
NeedsCompilation:	no
Packaged:	2026-06-18 16:03:18 UTC; k639s258
Author:	Karthik Srinivasan [aut, cre], Faiz Currim [aut], Sudha Ram [aut]
Maintainer:	Karthik Srinivasan <karthiks@ku.edu>
Repository:	CRAN
Date/Publication:	2026-06-24 08:00:27 UTC

blockwise: Reduced Modeling for Tabular Data with Blockwise Missingness

Description

Implements the Blockwise Reduced Modeling (BRM) method of Srinivasan, Currim, and Ram (2025). The main entry points are:

• brm — fit a BRM ensemble.

• predict.brm — predict on new data.

• choose_num_blocks — elbow heuristic for choosing k.

• learner and siblings — learner-agnostic fit/predict spec.

• simulate_blockwise_missing — mask complete data with a blockwise missingness pattern.

Author(s)

Maintainer: Karthik Srinivasan karthiks@ku.edu (ORCID)

Authors:

• Faiz Currim

• Sudha Ram

See Also

Useful links:

• https://github.com/KarAnalytics/blockwise

• Report bugs at https://github.com/KarAnalytics/blockwise/issues

UCI Adult income classification dataset

Description

Census-based binary classification dataset: predict whether a person's annual income exceeds $50,000. Used as the classification demonstration in Srinivasan, Currim, and Ram (2025).

Usage

adult

Format

A data.frame with roughly 32,561 rows including salary (the 0/1-valued response) and typical demographic/employment predictors.

Source

UCI Machine Learning Repository: https://archive.ics.uci.edu/dataset/2/adult

Capital Bikeshare hourly demand data

Description

Hourly count of rental bikes between 2011 and 2012 in the Capital Bikeshare system with the corresponding weather and seasonal information. Used as the regression demonstration in Srinivasan, Currim, and Ram (2025).

Usage

bike

Format

A data.frame with roughly 17,380 rows and the following columns:

season, mnth, hr, weekday, weathersit

Temporal and weather covariates.

temp, hum, windspeed

Numeric weather covariates.

cnt

Response: count of total rental bikes for that hour.

Source

UCI Machine Learning Repository: https://archive.ics.uci.edu/dataset/275/bike+sharing+dataset

Fit a Blockwise Reduced Modeling (BRM) ensemble

Description

BRM partitions the training data into n_blocks subsets based on per-row patterns of feature-missingness, fits one instance of the supplied learner per subset using only the features observed in that subset, and at prediction time routes each test row to the subset model whose training-time missingness pattern is closest.

Usage

brm(
  X,
  y,
  learner = learner_lm(),
  n_blocks = NULL,
  low_threshold = 0.05,
  n_restarts = 5L,
  overlap = TRUE
)

Arguments

Details

The learner interface is intentionally minimal: any fit / predict pair can be plugged in via learner(), and convenience specs are provided for common families (learner_lm, learner_rpart, etc.).

Value

An object of class "brm".

References

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

Examples

  data(bike, package = "blockwise")
  bike_miss <- simulate_blockwise_missing(
    bike,
    blocks       = list(c("hum", "windspeed", "weekday"),
                        c("hr", "temp", "weathersit")),
    prop_missing = 0.3
  )
  X <- bike_miss[, setdiff(names(bike_miss), "cnt")]
  y <- bike_miss$cnt
  fit <- brm(X, y, learner = learner_lm())
  preds <- predict(fit, X)

Estimate the number of blocks via the elbow heuristic

Description

Applies k-means to the binary missingness-indicator matrix for k = 1, ..., k_max and records, for each k, the fraction of still-missing cells after dropping columns that are sparse within each cluster. The curve is monotone-decreasing; BRM picks the k at its elbow.

Usage

choose_num_blocks(X, low_threshold = 0.05, n_restarts = 10, k_max = NULL)

Arguments

Details

Typically called internally by brm when n_blocks = NULL.

Value

A list with n_blocks (the chosen k) and missing_curve (numeric vector of length k_max).

References

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

King County, WA house sales

Description

The King County house-sales dataset. Used as a regression demonstration in Srinivasan, Currim, and Ram (2025).

Usage

house

Format

A data.frame with roughly 21,600 rows including price (the response) and typical property covariates such as bedrooms, bathrooms, sqft_living, sqft_lot, grade, and yr_built.

Source

Kaggle "House Sales in King County, USA" dataset.

Learner specification for BRM

Description

BRM trains one model per overlapping subset. The learner interface makes that choice user-controlled: supply a fit function that takes (X, y) and returns a fitted model, and a predict function that takes (model, X_new) and returns a numeric prediction vector (or a positive-class probability for binary classification).

Usage

learner(fit, predict, type = c("regression", "classification"))

learner_lm()

learner_glm_binomial()

learner_rpart(method = "anova", ...)

learner_ranger(...)

learner_gbm(distribution = "gaussian", n.trees = 500, ...)

Arguments

Value

An object of class "brm_learner".

Examples

my_learner <- learner(
  fit     = function(X, y) lm(y ~ ., data = cbind(X, y = y)),
  predict = function(m, X_new) predict(m, newdata = X_new),
  type    = "regression"
)

Predict from a fitted BRM ensemble

Description

Each row of newdata is routed to the block whose training-time missingness center is closest (Euclidean) to the row's missingness pattern. The corresponding block model then predicts on that row, using only that block's feature columns. Any NAs remaining in those columns are filled by simple mean/mode imputation against the training reference.

Usage

## S3 method for class 'brm'
predict(object, newdata, ...)

Arguments

Value

A numeric vector of length nrow(newdata).

Simulate a blockwise missing pattern on otherwise complete data

Description

Joint-masks groups of columns on randomly chosen rows, optionally adding light column-wise random-NA noise on top. Useful for benchmarking BRM on complete datasets; the default arguments reproduce the simulation design used in Srinivasan, Currim, and Ram (2025).

Usage

simulate_blockwise_missing(
  data,
  blocks,
  prop_missing,
  noise = 0.05,
  seed = NULL
)

Arguments

Value

A data.frame of the same shape as data, with NAs introduced in the specified pattern.

Examples

df <- data.frame(a = 1:100, b = 1:100, c = 1:100, d = 1:100)
simulate_blockwise_missing(df,
  blocks       = list(c("a", "b"), c("c", "d")),
  prop_missing = 0.3,
  seed         = 1234L)