Morais–Lima–Martin Split

MoraisLimaMartinSplit starts from a Kennard–Stone selection and then randomly swaps a fraction of samples between the training and test sets. This introduces controlled stochasticity into an otherwise deterministic diversity-based split, useful when you need multiple independent splits from a small dataset.

How it works

1. Run Kennard–Stone to produce an initial train/test split.
2. Randomly select swap_frac × min(n_train, n_test) samples from each cohort.
3. Swap those samples between train and test.

The result retains most of the spatial coverage of Kennard–Stone while adding randomness that enables ensemble evaluation.

When to use it

• You need multiple different splits from the same small dataset (e.g. to estimate the variance of a model evaluation metric), but pure random splits lose the coverage guarantee.
• You want Kennard–Stone as a baseline with some stochastic perturbation to assess split sensitivity.
• You are building a bootstrap or ensemble of calibration models and need diverse but roughly coverage-preserving training sets.

Usage

using DataSplits

# Default: swap 10% of samples.
res = partition(X, MoraisLimaMartinSplit(); train = 0.8, test = 0.2,
                rng = MersenneTwister(42))
X_train, X_test = splitdata(res, X)

# Swap 5% — closer to pure Kennard–Stone.
res = partition(X, MoraisLimaMartinSplit(; swap_frac = 0.05);
                train = 0.8, test = 0.2, rng = MersenneTwister(1))

# Swap 20% — more random, less deterministic.
res = partition(X, MoraisLimaMartinSplit(; swap_frac = 0.20);
                train = 0.8, test = 0.2)

# Custom metric.
using Distances
res = partition(X, MoraisLimaMartinSplit(; metric = Cityblock());
                train = 0.8, test = 0.2)

The rng keyword controls the random swap — fix it for reproducibility, vary it to get different realisations.

Parameters

API reference

• MoraisLimaMartinSplit