model-selection-rs 0.1.0

//! Random-permutation train/test splitting (Monte-Carlo cross-validation).

use rand::rngs::StdRng;
use rand::seq::SliceRandom;
use rand::SeedableRng;

use super::CvSplitter;
use crate::error::{ModelSelectionError, Result};

/// How to size a train or test subset: an absolute count or a fraction of the
/// dataset.
#[derive(Debug, Clone, Copy, PartialEq)]
pub enum SubsetSize {
    /// A fixed number of samples.
    Count(usize),
    /// A fraction in `(0.0, 1.0)` of the total sample count.
    Fraction(f64),
}

impl SubsetSize {
    /// Resolve to an absolute sample count against `n_samples`.
    pub(crate) fn resolve(self, n_samples: usize) -> usize {
        match self {
            SubsetSize::Count(c) => c,
            SubsetSize::Fraction(f) => (f * n_samples as f64).round() as usize,
        }
    }
}

/// Random-permutation cross-validation.
///
/// Yields `n_splits` independent train/test splits, each formed by permuting the
/// samples and slicing off a test set and a training set. Unlike
/// [`KFold`](super::KFold), the splits are **not** guaranteed to partition the
/// dataset — the same sample may be tested in several splits or none, and (if
/// `train_size + test_size < n`) some samples may be left out of a given split
/// entirely.
///
/// ```
/// use model_selection_rs::splitters::{CvSplitter, ShuffleSplit, SubsetSize};
///
/// let ss = ShuffleSplit::new(5)
///     .with_test_size(SubsetSize::Fraction(0.25))
///     .with_seed(0);
/// let splits = ss.split(40).unwrap();
/// assert_eq!(splits.len(), 5);
/// assert_eq!(splits[0].1.len(), 10); // 25% of 40
/// ```
#[derive(Debug, Clone)]
pub struct ShuffleSplit {
    n_splits: usize,
    test_size: SubsetSize,
    train_size: Option<SubsetSize>,
    seed: u64,
}

impl ShuffleSplit {
    /// Create a `ShuffleSplit` with `n_splits` splits, a default test size of
    /// 10%, and the remainder used for training.
    #[must_use]
    pub fn new(n_splits: usize) -> Self {
        Self {
            n_splits,
            test_size: SubsetSize::Fraction(0.1),
            train_size: None,
            seed: 0,
        }
    }

    /// Set the test-subset size.
    #[must_use]
    pub fn with_test_size(mut self, test_size: SubsetSize) -> Self {
        self.test_size = test_size;
        self
    }

    /// Set the train-subset size (defaults to "everything not in test").
    #[must_use]
    pub fn with_train_size(mut self, train_size: SubsetSize) -> Self {
        self.train_size = Some(train_size);
        self
    }

    /// Set the base RNG seed.
    #[must_use]
    pub fn with_seed(mut self, seed: u64) -> Self {
        self.seed = seed;
        self
    }

    /// Resolve `(n_train, n_test)` for a dataset of `n_samples`, validating that
    /// both are non-empty and fit.
    pub(crate) fn resolve_sizes(&self, n_samples: usize) -> Result<(usize, usize)> {
        let n_test = self.test_size.resolve(n_samples);
        let n_train = match self.train_size {
            Some(ts) => ts.resolve(n_samples),
            None => n_samples.saturating_sub(n_test),
        };
        if n_test == 0 || n_train == 0 {
            return Err(ModelSelectionError::InvalidSplitCount {
                msg: format!(
                    "resolved train={n_train}, test={n_test}; both must be >= 1 \
                     (n_samples={n_samples})"
                ),
            });
        }
        if n_train + n_test > n_samples {
            return Err(ModelSelectionError::NotEnoughSamples {
                needed: n_train + n_test,
                got: n_samples,
            });
        }
        Ok((n_train, n_test))
    }
}

impl CvSplitter for ShuffleSplit {
    fn split(&self, n_samples: usize) -> Result<Vec<(Vec<usize>, Vec<usize>)>> {
        let (n_train, n_test) = self.resolve_sizes(n_samples)?;
        let mut splits = Vec::with_capacity(self.n_splits);
        for i in 0..self.n_splits {
            // A distinct but deterministic permutation per split.
            let mut rng = StdRng::seed_from_u64(self.seed.wrapping_add(i as u64));
            let mut indices: Vec<usize> = (0..n_samples).collect();
            indices.shuffle(&mut rng);
            let test: Vec<usize> = indices[..n_test].to_vec();
            let train: Vec<usize> = indices[n_test..n_test + n_train].to_vec();
            splits.push((train, test));
        }
        Ok(splits)
    }

    fn n_splits(&self) -> usize {
        self.n_splits
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use std::collections::HashSet;

    #[test]
    fn sizes_honoured_and_disjoint() {
        let ss = ShuffleSplit::new(4)
            .with_test_size(SubsetSize::Count(5))
            .with_train_size(SubsetSize::Count(10))
            .with_seed(3);
        for (train, test) in ss.split(30).unwrap() {
            assert_eq!(train.len(), 10);
            assert_eq!(test.len(), 5);
            let tr: HashSet<_> = train.iter().collect();
            let te: HashSet<_> = test.iter().collect();
            assert!(tr.is_disjoint(&te));
        }
    }

    #[test]
    fn fraction_test_size() {
        let ss = ShuffleSplit::new(2).with_test_size(SubsetSize::Fraction(0.2));
        let splits = ss.split(50).unwrap();
        assert!(splits.iter().all(|(_, te)| te.len() == 10));
    }

    #[test]
    fn deterministic_for_seed() {
        let a = ShuffleSplit::new(3).with_seed(11).split(20).unwrap();
        let b = ShuffleSplit::new(3).with_seed(11).split(20).unwrap();
        assert_eq!(a, b);
    }

    #[test]
    fn errors_when_sizes_dont_fit() {
        let ss = ShuffleSplit::new(2)
            .with_test_size(SubsetSize::Count(20))
            .with_train_size(SubsetSize::Count(20));
        assert!(matches!(
            ss.split(30),
            Err(ModelSelectionError::NotEnoughSamples { .. })
        ));
    }
}