maxtwin

Robust deterministic data splitting for model validation.

maxtwin splits data into training and testing sets by balancing two objectives:

1. Distribution preservation -- the test set has the same distribution as the full data (measured by energy distance)
2. Separation -- test points are far from training points, ensuring non-trivial predictions

The method is robust to paired and clustered data, where methods like random splitting or DUPLEX can fail.

Installation

# After CRAN release
install.packages("maxtwin")

Usage

library(maxtwin)

set.seed(42)
X <- matrix(runif(500), ncol = 5)
Y <- sin(X[,1]) + cos(X[,2]) + X[,3]^2

# Split with 20% test set
test_idx <- maxtwin(X, Y, splitratio = 0.2)

X_train <- X[-test_idx, ]
X_test  <- X[test_idx, ]
Y_train <- Y[-test_idx]
Y_test  <- Y[test_idx]

Factor responses and factor predictor columns are encoded using Helmert contrasts. Character columns must first be converted to factors.

# Classification with a factor response
test_idx <- maxtwin(
  X = iris[, 1:4],
  Y = iris$Species,
  splitratio = 0.2
)

When there is no separate response, splitting can be based on all columns in a data frame:

test_idx <- maxtwin(iris, splitratio = 0.2)

Step 1 only (without iterative separation improvement)

test_idx <- initialmaxtwin(X, Y, splitratio = 0.2)

Options

Parameter	Default	Description
splitratio	--	Fraction of data for test set. Must be 1/r for integer r >= 2 (0.5, 0.25, 0.2, 0.1, ...)
initialization	TRUE	Start twinning from the point closest to the centroid
standardization	TRUE	Standardize the augmented data before twinning

Algorithm

MaxTwin (Algorithm 1 in the paper) proceeds in two steps:

Step 1: Apply twinning to select candidate test points, then within each Voronoi cell of the twinning set, pick the point with the largest nearest-neighbor distance.

Step 2: If the average separation distance falls below the threshold from Proposition 1, iteratively replace test points to improve separation while staying within a ball constraint that preserves distribution quality.

Dependencies

• FNN -- KD-tree based nearest neighbor search
• twinning -- Energy distance minimization

Testing

devtools::test()

Tests cover:

• Output correctness (size, bounds, uniqueness, sorting)
• Determinism (same seed = same result)
• Edge cases (univariate, high-dimensional, paired, and categorical data)
• Separation quality vs random splitting
• Consistency with original implementation

Reference

Joseph, V.R. and Cho, Y. (in press). A Robust Deterministic Data Splitting Method for Model Validation. Journal of Computational and Graphical Statistics.

License

GPL (>= 3)