Struct GroupKFold 

pub struct GroupKFold<G> { /* private fields */ }

Group K-fold cross-validation.

Guarantees that the samples of any one group (e.g. a patient id, a user id) never straddle the train/test boundary within a fold. Group leakage — the same entity contributing rows to both training and evaluation — is an easy-to-miss correctness bug in real ML work; this splitter makes it structurally impossible, complementing what StratifiedKFold does for class balance.

Groups are assigned whole to folds using a greedy largest-group-first heuristic (the same idea scikit-learn uses): repeatedly place the largest remaining group into the fold that currently holds the fewest samples. This keeps fold sizes close without ever splitting a group.

The number of splits must not exceed the number of distinct groups.

use ndarray::array;
use model_selection_rs::splitters::{CvSplitter, GroupKFold};

let groups = array![1, 1, 2, 2, 3, 3, 4, 4];
let gkf = GroupKFold::new(2, &groups).unwrap();
let splits = gkf.split(groups.len()).unwrap();
assert_eq!(splits.len(), 2);

Implementations

impl<G: Eq + Hash + Clone> GroupKFold<G>

pub fn new(n_splits: usize, groups: &Array1<G>) -> Result<Self>

Create a GroupKFold over the per-sample groups.

Errors

Returns ModelSelectionError::InvalidSplitCount if n_splits < 2.

Trait Implementations

impl<G: Clone> Clone for GroupKFold<G>

fn clone(&self) -> GroupKFold<G>

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl<G: Eq + Hash + Clone> CvSplitter for GroupKFold<G>

fn split(&self, n_samples: usize) -> Result<Vec<(Vec<usize>, Vec<usize>)>>

Produce every (train_indices, test_indices) pair for n_samples rows.

fn n_splits(&self) -> usize

Number of splits this strategy yields.

impl<G: Debug> Debug for GroupKFold<G>

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.