dml-core 0.4.0

Double / debiased machine learning for causal inference in Rust
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151

use ndarray::{Array1, ArrayView1, ArrayView2};

use super::{center, Centered, Regressor};
use crate::error::{DmlError, Result};
use crate::linalg::cholesky_solve;

/// Ordinary least squares with an intercept and optional ridge penalty.
///
/// `ridge` (default 0) is added to the diagonal for slopes only, which keeps
/// the fit well defined when covariates are collinear.
#[derive(Debug, Clone, Default)]
pub struct Ols {
    ridge: f64,
    coef: Option<Array1<f64>>,
    intercept: f64,
}

impl Ols {
    pub fn new() -> Self {
        Self::default()
    }

    pub fn with_ridge(ridge: f64) -> Self {
        Self {
            ridge: ridge.max(0.0),
            ..Self::default()
        }
    }

    pub fn coef(&self) -> Option<&Array1<f64>> {
        self.coef.as_ref()
    }

    pub fn intercept(&self) -> f64 {
        self.intercept
    }
}

impl Ols {
    fn fit_impl(
        &mut self,
        x: ArrayView2<'_, f64>,
        y: ArrayView1<'_, f64>,
        w: Option<ArrayView1<'_, f64>>,
    ) -> Result<()> {
        let p = x.ncols();
        // Centre to absorb the intercept, then solve (Xc'Xc + λI) b = Xc'yc.
        let Centered {
            x_mean,
            y_mean,
            xc,
            yc,
        } = center(x, y, w, false)?;
        let mut xtx = xc.t().dot(&xc);
        for j in 0..p {
            xtx[[j, j]] += self.ridge;
        }
        let xty = xc.t().dot(&yc);
        let coef = if p == 0 {
            Array1::zeros(0)
        } else {
            cholesky_solve(&xtx, &xty)?
        };
        self.intercept = y_mean - x_mean.dot(&coef);
        self.coef = Some(coef);
        Ok(())
    }
}

impl Regressor for Ols {
    fn fit(&mut self, x: ArrayView2<'_, f64>, y: ArrayView1<'_, f64>) -> Result<()> {
        self.fit_impl(x, y, None)
    }

    /// Weighted least squares: minimises `Σ wᵢ (yᵢ − b − xᵢᵀβ)² + ridge·‖β‖²`.
    fn fit_weighted(
        &mut self,
        x: ArrayView2<'_, f64>,
        y: ArrayView1<'_, f64>,
        w: ArrayView1<'_, f64>,
    ) -> Result<()> {
        self.fit_impl(x, y, Some(w))
    }

    fn predict(&self, x: ArrayView2<'_, f64>) -> Result<Array1<f64>> {
        let coef = self.coef.as_ref().ok_or(DmlError::NotFitted)?;
        if x.ncols() != coef.len() {
            return Err(DmlError::DimensionMismatch(
                "Ols::predict feature count".into(),
            ));
        }
        Ok(x.dot(coef) + self.intercept)
    }

    fn fresh(&self) -> Box<dyn Regressor> {
        Box::new(Ols::with_ridge(self.ridge))
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use ndarray::array;

    #[test]
    fn recovers_exact_linear_fit() {
        let x = array![[1.0, 0.0], [0.0, 1.0], [1.0, 1.0], [2.0, 1.0], [0.5, 3.0]];
        let y: Array1<f64> = x
            .rows()
            .into_iter()
            .map(|r| 1.5 + 2.0 * r[0] - 0.5 * r[1])
            .collect();
        let mut m = Ols::new();
        m.fit(x.view(), y.view()).unwrap();
        let c = m.coef().unwrap();
        assert!((c[0] - 2.0).abs() < 1e-10 && (c[1] + 0.5).abs() < 1e-10);
        assert!((m.intercept() - 1.5).abs() < 1e-10);
    }

    #[test]
    fn predict_before_fit_errors() {
        assert_eq!(
            Ols::new().predict(array![[1.0]].view()).unwrap_err(),
            DmlError::NotFitted
        );
    }

    #[test]
    fn integer_weights_equal_row_duplication() {
        let x = array![[0.0], [1.0], [2.0], [3.0]];
        let y = array![1.0, 0.5, 2.5, 2.0];
        let mut weighted = Ols::new();
        weighted
            .fit_weighted(x.view(), y.view(), array![1.0, 2.0, 1.0, 3.0].view())
            .unwrap();
        let xd = array![[0.0], [1.0], [1.0], [2.0], [3.0], [3.0], [3.0]];
        let yd = array![1.0, 0.5, 0.5, 2.5, 2.0, 2.0, 2.0];
        let mut dup = Ols::new();
        dup.fit(xd.view(), yd.view()).unwrap();
        assert!((weighted.coef().unwrap()[0] - dup.coef().unwrap()[0]).abs() < 1e-12);
        assert!((weighted.intercept() - dup.intercept()).abs() < 1e-12);
    }

    #[test]
    fn ridge_handles_collinearity() {
        let x = array![[1.0, 2.0], [2.0, 4.0], [3.0, 6.0]];
        let y = array![1.0, 2.0, 3.0];
        assert!(Ols::new().fit(x.view(), y.view()).is_err());
        assert!(Ols::with_ridge(1e-6).fit(x.view(), y.view()).is_ok());
    }
}