Package 'blockwise'

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

X	A data.frame of predictors. May contain NA. Categorical predictors should be factors.
y	A numeric vector (regression) or a 0/1 numeric vector (binary classification) of length nrow(X).
learner	A learner specification. Defaults to learner_lm().
n_blocks	Integer number of blocks; if NULL, chosen automatically by choose_num_blocks.
low_threshold	Column-density threshold for including a predictor in a block's model. Default 0.05.
n_restarts	k-means restarts for block assignment. Default 5.
overlap	If TRUE (default) subsets are enlarged via the set-theoretic inclusion rule; if FALSE, the non-overlapping variant is used.

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

X	A data.frame of predictors; may contain NA.
low_threshold	Fraction below which a column is considered absent in a candidate subset. Default 0.05.
n_restarts	Number of k-means restarts per k. Default 10.
k_max	Upper bound on k. Default min(ncol(X), 50).

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

fit	A function of the form function(X, y) -> model.
predict	A function of the form function(model, X_new) -> numeric.
type	Either "regression" or "classification".
method	rpart split method; one of "anova", "class", etc.
...	Additional arguments passed to the underlying fitter.
distribution	gbm distribution (e.g. "gaussian", "bernoulli", "poisson").
n.trees	Number of trees.

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

object	A fitted brm object.
newdata	A data.frame of predictors. May contain NA.
...	Unused.

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

data	A data.frame.
blocks	A list of character vectors; each vector names the columns masked jointly for one block of rows.
prop_missing	Proportion of rows affected per block. Either a scalar (applied to each block) or a numeric vector of length length(blocks).
noise	Extra per-column random-NA rate applied on top, restricted to columns named in any block. Default 0.05. Set to 0 to disable.
seed	Optional integer base seed for reproducibility. If NULL (the default), the current RNG state is used and not modified. If supplied, the seed is applied locally via with_seed so that the caller's RNG state is preserved.

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)

---