dml_core/
crossfit.rs

1use ndarray::{Array1, ArrayView1, ArrayView2, Axis};
2use rand::seq::SliceRandom;
3use rand::SeedableRng;
4use rand_chacha::ChaCha8Rng;
5
6use crate::data::gather;
7use crate::error::{DmlError, Result};
8use crate::learner::{Classifier, Regressor};
9
10/// One train/test split.
11#[derive(Debug, Clone, PartialEq)]
12pub struct Fold {
13    pub train: Vec<usize>,
14    pub test: Vec<usize>,
15}
16
17/// Shuffled K-fold splitter with a fixed seed, so results are reproducible.
18#[derive(Debug, Clone)]
19pub struct KFold {
20    n_folds: usize,
21    seed: u64,
22}
23
24impl KFold {
25    pub fn new(n_folds: usize, seed: u64) -> Result<Self> {
26        if n_folds < 2 {
27            return Err(DmlError::InvalidArgument(
28                "n_folds must be at least 2".into(),
29            ));
30        }
31        Ok(Self { n_folds, seed })
32    }
33
34    pub fn n_folds(&self) -> usize {
35        self.n_folds
36    }
37
38    /// Split `0..n` into `n_folds` disjoint test sets covering every index once.
39    pub fn split(&self, n: usize) -> Result<Vec<Fold>> {
40        if n < self.n_folds {
41            return Err(DmlError::InvalidArgument(format!(
42                "{n} observations is fewer than {} folds",
43                self.n_folds
44            )));
45        }
46        let mut idx: Vec<usize> = (0..n).collect();
47        idx.shuffle(&mut ChaCha8Rng::seed_from_u64(self.seed));
48
49        let base = n / self.n_folds;
50        let extra = n % self.n_folds;
51        let mut folds = Vec::with_capacity(self.n_folds);
52        let mut start = 0;
53        for k in 0..self.n_folds {
54            let size = base + usize::from(k < extra);
55            let mut test = idx[start..start + size].to_vec();
56            let mut train: Vec<usize> = idx[..start]
57                .iter()
58                .chain(&idx[start + size..])
59                .copied()
60                .collect();
61            test.sort_unstable();
62            train.sort_unstable();
63            folds.push(Fold { train, test });
64            start += size;
65        }
66        Ok(folds)
67    }
68}
69
70/// How observations are split for cross-fitting: `n_rep` independent
71/// repetitions of seeded K-fold splitting.
72#[derive(Debug, Clone, Copy, PartialEq, Eq)]
73pub struct SampleSplitting {
74    pub n_folds: usize,
75    pub n_rep: usize,
76    pub seed: u64,
77}
78
79impl Default for SampleSplitting {
80    fn default() -> Self {
81        Self {
82            n_folds: 5,
83            n_rep: 1,
84            seed: 0,
85        }
86    }
87}
88
89impl SampleSplitting {
90    /// Draw the folds for every repetition. Repetition `r` uses its own seed
91    /// derived from `seed`, so repetition 0 equals `KFold::new(n_folds, seed)`.
92    pub fn draw(&self, n: usize) -> Result<Vec<Vec<Fold>>> {
93        if self.n_rep == 0 {
94            return Err(DmlError::InvalidArgument("n_rep must be at least 1".into()));
95        }
96        (0..self.n_rep as u64)
97            .map(|r| {
98                // Golden-ratio increment keeps repetition seeds far apart.
99                let seed = self
100                    .seed
101                    .wrapping_add(r.wrapping_mul(0x9E37_79B9_7F4A_7C15));
102                KFold::new(self.n_folds, seed)?.split(n)
103            })
104            .collect()
105    }
106}
107
108/// Out-of-fold predictions: for each fold, fit a fresh copy of `learner` on the
109/// training rows and predict the held-out rows. Every row is predicted by a
110/// model that never saw it — the core of DML cross-fitting.
111pub fn cross_fit_predict(
112    learner: &dyn Regressor,
113    x: ArrayView2<'_, f64>,
114    target: ArrayView1<'_, f64>,
115    folds: &[Fold],
116) -> Result<Array1<f64>> {
117    if target.len() != x.nrows() {
118        return Err(DmlError::DimensionMismatch(
119            "cross_fit_predict x/target rows".into(),
120        ));
121    }
122    cross_fit_with(x.nrows(), folds, |fold| {
123        let mut model = learner.fresh();
124        model.fit(
125            x.select(Axis(0), &fold.train).view(),
126            gather(target, &fold.train).view(),
127        )?;
128        model.predict(x.select(Axis(0), &fold.test).view())
129    })
130}
131
132/// Like [`cross_fit_predict`], but each model is trained only on the training
133/// rows where `keep[i]` is true (e.g. controls only); predictions still cover
134/// every test row.
135pub fn cross_fit_predict_subset(
136    learner: &dyn Regressor,
137    x: ArrayView2<'_, f64>,
138    target: ArrayView1<'_, f64>,
139    folds: &[Fold],
140    keep: &[bool],
141) -> Result<Array1<f64>> {
142    cross_fit_predict(learner, x, target, &subset_folds(folds, keep, x.nrows())?)
143}
144
145/// Like [`cross_fit_predict_proba`], but each model is trained only on the
146/// training rows where `keep[i]` is true; predictions cover every test row.
147pub fn cross_fit_predict_proba_subset(
148    learner: &dyn Classifier,
149    x: ArrayView2<'_, f64>,
150    labels: ArrayView1<'_, f64>,
151    folds: &[Fold],
152    keep: &[bool],
153) -> Result<Array1<f64>> {
154    cross_fit_predict_proba(learner, x, labels, &subset_folds(folds, keep, x.nrows())?)
155}
156
157/// Restrict every training set to the rows where `keep` is true.
158fn subset_folds(folds: &[Fold], keep: &[bool], n: usize) -> Result<Vec<Fold>> {
159    if keep.len() != n {
160        return Err(DmlError::DimensionMismatch("subgroup mask length".into()));
161    }
162    let subset: Vec<Fold> = folds
163        .iter()
164        .map(|f| Fold {
165            train: f.train.iter().copied().filter(|&i| keep[i]).collect(),
166            test: f.test.clone(),
167        })
168        .collect();
169    if subset.iter().any(|f| f.train.is_empty()) {
170        return Err(DmlError::InvalidArgument(
171            "a training fold has no observations in the requested subgroup".into(),
172        ));
173    }
174    Ok(subset)
175}
176
177/// Out-of-fold class-1 probabilities from a [`Classifier`].
178pub fn cross_fit_predict_proba(
179    learner: &dyn Classifier,
180    x: ArrayView2<'_, f64>,
181    labels: ArrayView1<'_, f64>,
182    folds: &[Fold],
183) -> Result<Array1<f64>> {
184    if labels.len() != x.nrows() {
185        return Err(DmlError::DimensionMismatch(
186            "cross_fit_predict_proba x/labels rows".into(),
187        ));
188    }
189    cross_fit_with(x.nrows(), folds, |fold| {
190        let mut model = learner.fresh();
191        model.fit(
192            x.select(Axis(0), &fold.train).view(),
193            gather(labels, &fold.train).view(),
194        )?;
195        model.predict_proba(x.select(Axis(0), &fold.test).view())
196    })
197}
198
199/// Run `fit_predict` on every fold and scatter each fold's test-set
200/// predictions into one length-`n` vector.
201///
202/// With the `parallel` feature (on by default) folds run on rayon's thread
203/// pool; results are identical to the sequential order because every fold
204/// writes to disjoint rows.
205pub(crate) fn cross_fit_with<F>(n: usize, folds: &[Fold], fit_predict: F) -> Result<Array1<f64>>
206where
207    F: Fn(&Fold) -> Result<Array1<f64>> + Sync,
208{
209    #[cfg(feature = "parallel")]
210    let preds: Vec<Array1<f64>> = {
211        use rayon::prelude::*;
212        folds.par_iter().map(&fit_predict).collect::<Result<_>>()?
213    };
214    #[cfg(not(feature = "parallel"))]
215    let preds: Vec<Array1<f64>> = folds.iter().map(&fit_predict).collect::<Result<_>>()?;
216
217    let mut out = Array1::<f64>::from_elem(n, f64::NAN);
218    for (fold, pred) in folds.iter().zip(preds) {
219        if pred.len() != fold.test.len() {
220            return Err(DmlError::DimensionMismatch(
221                "learner returned wrong number of predictions".into(),
222            ));
223        }
224        for (&i, &p) in fold.test.iter().zip(pred.iter()) {
225            out[i] = p;
226        }
227    }
228    if out.iter().any(|v| v.is_nan()) {
229        return Err(DmlError::InvalidArgument(
230            "folds do not cover every observation, or a learner returned NaN".into(),
231        ));
232    }
233    Ok(out)
234}
235
236#[cfg(test)]
237mod tests {
238    use super::*;
239    use crate::learner::Ols;
240    use ndarray::Array2;
241
242    #[test]
243    fn folds_partition_all_indices() {
244        let folds = KFold::new(3, 7).unwrap().split(10).unwrap();
245        let mut seen: Vec<usize> = folds.iter().flat_map(|f| f.test.clone()).collect();
246        seen.sort_unstable();
247        assert_eq!(seen, (0..10).collect::<Vec<_>>());
248        for f in &folds {
249            assert!(f.test.iter().all(|i| !f.train.contains(i)));
250            assert_eq!(f.train.len() + f.test.len(), 10);
251        }
252        let sizes: Vec<usize> = folds.iter().map(|f| f.test.len()).collect();
253        assert_eq!(sizes, vec![4, 3, 3]);
254    }
255
256    #[test]
257    fn same_seed_same_split() {
258        let a = KFold::new(5, 42).unwrap().split(50).unwrap();
259        let b = KFold::new(5, 42).unwrap().split(50).unwrap();
260        let c = KFold::new(5, 43).unwrap().split(50).unwrap();
261        assert_eq!(a, b);
262        assert_ne!(a, c);
263    }
264
265    #[test]
266    fn repetitions_differ_and_first_matches_kfold() {
267        let reps = SampleSplitting {
268            n_folds: 4,
269            n_rep: 3,
270            seed: 9,
271        }
272        .draw(40)
273        .unwrap();
274        assert_eq!(reps.len(), 3);
275        assert_eq!(reps[0], KFold::new(4, 9).unwrap().split(40).unwrap());
276        assert_ne!(reps[0], reps[1]);
277        assert_ne!(reps[1], reps[2]);
278    }
279
280    #[test]
281    fn subset_training_uses_only_kept_rows() {
282        // Target is 0 on kept rows and 100 elsewhere: a mean model trained
283        // only on kept rows predicts ~0 everywhere.
284        let x = Array2::<f64>::zeros((20, 0));
285        let keep: Vec<bool> = (0..20).map(|i| i % 2 == 0).collect();
286        let y: Array1<f64> = keep.iter().map(|&k| if k { 0.0 } else { 100.0 }).collect();
287        let folds = KFold::new(4, 1).unwrap().split(20).unwrap();
288        let pred =
289            cross_fit_predict_subset(&Ols::new(), x.view(), y.view(), &folds, &keep).unwrap();
290        assert!(pred.iter().all(|p| p.abs() < 1e-12));
291    }
292
293    #[test]
294    fn rejects_bad_settings() {
295        assert!(KFold::new(1, 0).is_err());
296        assert!(KFold::new(5, 0).unwrap().split(3).is_err());
297    }
298
299    #[test]
300    fn cross_fit_recovers_linear_signal() {
301        let n = 100;
302        let x = Array2::from_shape_fn((n, 2), |(i, j)| ((i * 7 + j * 3) % 11) as f64);
303        let y: Array1<f64> = x
304            .rows()
305            .into_iter()
306            .map(|r| 1.0 + 0.5 * r[0] - 2.0 * r[1])
307            .collect();
308        let folds = KFold::new(5, 1).unwrap().split(n).unwrap();
309        let pred = cross_fit_predict(&Ols::new(), x.view(), y.view(), &folds).unwrap();
310        assert!((&pred - &y).iter().all(|e| e.abs() < 1e-8));
311    }
312}
dml_core/crossfit.rs

dml_core/
crossfit.rs
