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    /// Stratified split: the rows of each stratum (e.g. treated / control) are
70    /// shuffled and dealt round-robin over the folds, so every fold holds about
71    /// the same share of each stratum. Fails if a stratum has fewer rows than
72    /// there are folds.
73    pub fn split_stratified(&self, strata: &[usize]) -> Result<Vec<Fold>> {
74        let n = strata.len();
75        if n < self.n_folds {
76            return Err(DmlError::InvalidArgument(format!(
77                "{n} observations is fewer than {} folds",
78                self.n_folds
79            )));
80        }
81        let mut groups: std::collections::BTreeMap<usize, Vec<usize>> = Default::default();
82        for (i, &s) in strata.iter().enumerate() {
83            groups.entry(s).or_default().push(i);
84        }
85        let mut rng = ChaCha8Rng::seed_from_u64(self.seed);
86        let mut tests: Vec<Vec<usize>> = vec![Vec::new(); self.n_folds];
87        let mut next = 0;
88        for (stratum, mut idx) in groups {
89            if idx.len() < self.n_folds {
90                return Err(DmlError::InvalidArgument(format!(
91                    "stratum {stratum} has {} observations, fewer than {} folds",
92                    idx.len(),
93                    self.n_folds
94                )));
95            }
96            idx.shuffle(&mut rng);
97            for i in idx {
98                tests[next].push(i);
99                next = (next + 1) % self.n_folds;
100            }
101        }
102        Ok(tests
103            .into_iter()
104            .map(|mut test| {
105                test.sort_unstable();
106                let mut in_test = vec![false; n];
107                for &i in &test {
108                    in_test[i] = true;
109                }
110                let train = (0..n).filter(|&i| !in_test[i]).collect();
111                Fold { train, test }
112            })
113            .collect())
114    }
115}
116
117/// How observations are split for cross-fitting: `n_rep` independent
118/// repetitions of seeded K-fold splitting.
119#[derive(Debug, Clone, Copy, PartialEq, Eq)]
120pub struct SampleSplitting {
121    pub n_folds: usize,
122    pub n_rep: usize,
123    pub seed: u64,
124}
125
126impl Default for SampleSplitting {
127    fn default() -> Self {
128        Self {
129            n_folds: 5,
130            n_rep: 1,
131            seed: 0,
132        }
133    }
134}
135
136impl SampleSplitting {
137    /// Draw the folds for every repetition. Repetition `r` uses its own seed
138    /// derived from `seed`, so repetition 0 equals `KFold::new(n_folds, seed)`.
139    pub fn draw(&self, n: usize) -> Result<Vec<Vec<Fold>>> {
140        if self.n_rep == 0 {
141            return Err(DmlError::InvalidArgument("n_rep must be at least 1".into()));
142        }
143        (0..self.n_rep as u64)
144            .map(|r| {
145                // Golden-ratio increment keeps repetition seeds far apart.
146                let seed = self
147                    .seed
148                    .wrapping_add(r.wrapping_mul(0x9E37_79B9_7F4A_7C15));
149                KFold::new(self.n_folds, seed)?.split(n)
150            })
151            .collect()
152    }
153}
154
155/// Out-of-fold predictions: for each fold, fit a fresh copy of `learner` on the
156/// training rows and predict the held-out rows. Every row is predicted by a
157/// model that never saw it — the core of DML cross-fitting.
158pub fn cross_fit_predict(
159    learner: &dyn Regressor,
160    x: ArrayView2<'_, f64>,
161    target: ArrayView1<'_, f64>,
162    folds: &[Fold],
163) -> Result<Array1<f64>> {
164    if target.len() != x.nrows() {
165        return Err(DmlError::DimensionMismatch(
166            "cross_fit_predict x/target rows".into(),
167        ));
168    }
169    cross_fit_with(x.nrows(), folds, |fold| {
170        let mut model = learner.fresh();
171        model.fit(
172            x.select(Axis(0), &fold.train).view(),
173            gather(target, &fold.train).view(),
174        )?;
175        model.predict(x.select(Axis(0), &fold.test).view())
176    })
177}
178
179/// Like [`cross_fit_predict`], but each model is trained only on the training
180/// rows where `keep[i]` is true (e.g. controls only); predictions still cover
181/// every test row.
182pub fn cross_fit_predict_subset(
183    learner: &dyn Regressor,
184    x: ArrayView2<'_, f64>,
185    target: ArrayView1<'_, f64>,
186    folds: &[Fold],
187    keep: &[bool],
188) -> Result<Array1<f64>> {
189    cross_fit_predict(learner, x, target, &subset_folds(folds, keep, x.nrows())?)
190}
191
192/// Like [`cross_fit_predict_proba`], but each model is trained only on the
193/// training rows where `keep[i]` is true; predictions cover every test row.
194pub fn cross_fit_predict_proba_subset(
195    learner: &dyn Classifier,
196    x: ArrayView2<'_, f64>,
197    labels: ArrayView1<'_, f64>,
198    folds: &[Fold],
199    keep: &[bool],
200) -> Result<Array1<f64>> {
201    cross_fit_predict_proba(learner, x, labels, &subset_folds(folds, keep, x.nrows())?)
202}
203
204/// Restrict every training set to the rows where `keep` is true.
205fn subset_folds(folds: &[Fold], keep: &[bool], n: usize) -> Result<Vec<Fold>> {
206    if keep.len() != n {
207        return Err(DmlError::DimensionMismatch("subgroup mask length".into()));
208    }
209    let subset: Vec<Fold> = folds
210        .iter()
211        .map(|f| Fold {
212            train: f.train.iter().copied().filter(|&i| keep[i]).collect(),
213            test: f.test.clone(),
214        })
215        .collect();
216    if subset.iter().any(|f| f.train.is_empty()) {
217        return Err(DmlError::InvalidArgument(
218            "a training fold has no observations in the requested subgroup".into(),
219        ));
220    }
221    Ok(subset)
222}
223
224/// Out-of-fold class-1 probabilities from a [`Classifier`].
225pub fn cross_fit_predict_proba(
226    learner: &dyn Classifier,
227    x: ArrayView2<'_, f64>,
228    labels: ArrayView1<'_, f64>,
229    folds: &[Fold],
230) -> Result<Array1<f64>> {
231    if labels.len() != x.nrows() {
232        return Err(DmlError::DimensionMismatch(
233            "cross_fit_predict_proba x/labels rows".into(),
234        ));
235    }
236    cross_fit_with(x.nrows(), folds, |fold| {
237        let mut model = learner.fresh();
238        model.fit(
239            x.select(Axis(0), &fold.train).view(),
240            gather(labels, &fold.train).view(),
241        )?;
242        model.predict_proba(x.select(Axis(0), &fold.test).view())
243    })
244}
245
246/// Run `fit_predict` on every fold and scatter each fold's test-set
247/// predictions into one length-`n` vector.
248///
249/// With the `parallel` feature (on by default) folds run on rayon's thread
250/// pool; results are identical to the sequential order because every fold
251/// writes to disjoint rows.
252pub(crate) fn cross_fit_with<F>(n: usize, folds: &[Fold], fit_predict: F) -> Result<Array1<f64>>
253where
254    F: Fn(&Fold) -> Result<Array1<f64>> + Sync,
255{
256    #[cfg(feature = "parallel")]
257    let preds: Vec<Array1<f64>> = {
258        use rayon::prelude::*;
259        folds.par_iter().map(&fit_predict).collect::<Result<_>>()?
260    };
261    #[cfg(not(feature = "parallel"))]
262    let preds: Vec<Array1<f64>> = folds.iter().map(&fit_predict).collect::<Result<_>>()?;
263
264    let mut out = Array1::<f64>::from_elem(n, f64::NAN);
265    for (fold, pred) in folds.iter().zip(preds) {
266        if pred.len() != fold.test.len() {
267            return Err(DmlError::DimensionMismatch(
268                "learner returned wrong number of predictions".into(),
269            ));
270        }
271        for (&i, &p) in fold.test.iter().zip(pred.iter()) {
272            out[i] = p;
273        }
274    }
275    if out.iter().any(|v| v.is_nan()) {
276        return Err(DmlError::InvalidArgument(
277            "folds do not cover every observation, or a learner returned NaN".into(),
278        ));
279    }
280    Ok(out)
281}
282
283#[cfg(test)]
284mod tests {
285    use super::*;
286    use crate::learner::Ols;
287    use ndarray::Array2;
288
289    #[test]
290    fn folds_partition_all_indices() {
291        let folds = KFold::new(3, 7).unwrap().split(10).unwrap();
292        let mut seen: Vec<usize> = folds.iter().flat_map(|f| f.test.clone()).collect();
293        seen.sort_unstable();
294        assert_eq!(seen, (0..10).collect::<Vec<_>>());
295        for f in &folds {
296            assert!(f.test.iter().all(|i| !f.train.contains(i)));
297            assert_eq!(f.train.len() + f.test.len(), 10);
298        }
299        let sizes: Vec<usize> = folds.iter().map(|f| f.test.len()).collect();
300        assert_eq!(sizes, vec![4, 3, 3]);
301    }
302
303    #[test]
304    fn same_seed_same_split() {
305        let a = KFold::new(5, 42).unwrap().split(50).unwrap();
306        let b = KFold::new(5, 42).unwrap().split(50).unwrap();
307        let c = KFold::new(5, 43).unwrap().split(50).unwrap();
308        assert_eq!(a, b);
309        assert_ne!(a, c);
310    }
311
312    #[test]
313    fn repetitions_differ_and_first_matches_kfold() {
314        let reps = SampleSplitting {
315            n_folds: 4,
316            n_rep: 3,
317            seed: 9,
318        }
319        .draw(40)
320        .unwrap();
321        assert_eq!(reps.len(), 3);
322        assert_eq!(reps[0], KFold::new(4, 9).unwrap().split(40).unwrap());
323        assert_ne!(reps[0], reps[1]);
324        assert_ne!(reps[1], reps[2]);
325    }
326
327    #[test]
328    fn subset_training_uses_only_kept_rows() {
329        // Target is 0 on kept rows and 100 elsewhere: a mean model trained
330        // only on kept rows predicts ~0 everywhere.
331        let x = Array2::<f64>::zeros((20, 0));
332        let keep: Vec<bool> = (0..20).map(|i| i % 2 == 0).collect();
333        let y: Array1<f64> = keep.iter().map(|&k| if k { 0.0 } else { 100.0 }).collect();
334        let folds = KFold::new(4, 1).unwrap().split(20).unwrap();
335        let pred =
336            cross_fit_predict_subset(&Ols::new(), x.view(), y.view(), &folds, &keep).unwrap();
337        assert!(pred.iter().all(|p| p.abs() < 1e-12));
338    }
339
340    #[test]
341    fn stratified_folds_balance_each_stratum() {
342        let strata: Vec<usize> = (0..53).map(|i| usize::from(i % 5 == 0)).collect(); // 11 treated
343        let folds = KFold::new(5, 3).unwrap().split_stratified(&strata).unwrap();
344        let mut seen: Vec<usize> = folds.iter().flat_map(|f| f.test.clone()).collect();
345        seen.sort_unstable();
346        assert_eq!(seen, (0..53).collect::<Vec<_>>());
347        for f in &folds {
348            let treated = f.test.iter().filter(|&&i| strata[i] == 1).count();
349            assert!((2..=3).contains(&treated), "treated per fold: {treated}");
350            assert_eq!(f.train.len() + f.test.len(), 53);
351        }
352        assert!(KFold::new(5, 0)
353            .unwrap()
354            .split_stratified(&[0, 0, 0, 0, 0, 1])
355            .is_err());
356    }
357
358    #[test]
359    fn rejects_bad_settings() {
360        assert!(KFold::new(1, 0).is_err());
361        assert!(KFold::new(5, 0).unwrap().split(3).is_err());
362    }
363
364    #[test]
365    fn cross_fit_recovers_linear_signal() {
366        let n = 100;
367        let x = Array2::from_shape_fn((n, 2), |(i, j)| ((i * 7 + j * 3) % 11) as f64);
368        let y: Array1<f64> = x
369            .rows()
370            .into_iter()
371            .map(|r| 1.0 + 0.5 * r[0] - 2.0 * r[1])
372            .collect();
373        let folds = KFold::new(5, 1).unwrap().split(n).unwrap();
374        let pred = cross_fit_predict(&Ols::new(), x.view(), y.view(), &folds).unwrap();
375        assert!((&pred - &y).iter().all(|e| e.abs() < 1e-8));
376    }
377}
dml_core/crossfit.rs

dml_core/
crossfit.rs
