datarust 0.6.1

Scikit-learn-style preprocessing and classical ML 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
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238

//! Integration tests for `linear_model::Ridge` and `linear_model::Lasso`.

use datarust::linear_model::{Lasso, LinearRegression, Ridge, RidgeSolver};
use datarust::traits::{Predictor, Regressor};
use datarust::Matrix;

fn approx(a: f64, b: f64, tol: f64) -> bool {
    (a - b).abs() < tol
}

/// Non-collinear feature matrix with a known linear signal.
fn sample_xy() -> (Matrix, Vec<f64>) {
    let rows: Vec<Vec<f64>> = (0..50)
        .map(|i| {
            let i = i as f64;
            vec![i.sin(), (i + 7.0).ln(), (i * 0.3 + 1.0).exp() * 0.01]
        })
        .collect();
    let y: Vec<f64> = rows
        .iter()
        .map(|r| 2.0 * r[0] - 1.0 * r[1] + 0.5 * r[2] + 3.0)
        .collect();
    (Matrix::new(rows).unwrap(), y)
}

// ============================== Ridge =====================================

#[test]
fn ridge_alpha_zero_matches_linear_regression() {
    let (x, y) = sample_xy();
    let mut lr = LinearRegression::new();
    lr.fit(&x, &y).unwrap();
    let mut ridge = Ridge::new().with_alpha(0.0);
    ridge.fit(&x, &y).unwrap();
    for (r, l) in ridge.coef().iter().zip(lr.coef().iter()) {
        assert!(approx(*r, *l, 1e-6), "ridge={r} lr={l}");
    }
    assert!(approx(ridge.intercept(), lr.intercept(), 1e-6));
}

#[test]
fn ridge_shrinkage_monotonic() {
    let (x, y) = sample_xy();
    let mut norms = Vec::new();
    for &alpha in &[0.01, 1.0, 100.0, 10_000.0] {
        let mut m = Ridge::new().with_alpha(alpha);
        m.fit(&x, &y).unwrap();
        let norm: f64 = m.coef().iter().map(|v| v * v).sum::<f64>().sqrt();
        norms.push(norm);
    }
    // L2 norm should be non-increasing as alpha grows.
    for w in norms.windows(2) {
        assert!(w[1] <= w[0] + 1e-9, "norms not decreasing: {norms:?}");
    }
}

#[test]
fn ridge_handles_collinear_features() {
    // Duplicate column → XᵀX singular; LinearRegression (Cholesky) fails, Ridge works.
    let x = Matrix::new(vec![
        vec![1.0, 1.0],
        vec![2.0, 2.0],
        vec![3.0, 3.0],
        vec![4.0, 4.0],
    ])
    .unwrap();
    let y = vec![2.0, 4.0, 6.0, 8.0];

    let mut lr = LinearRegression::new();
    assert!(lr.fit(&x, &y).is_err());

    let mut ridge = Ridge::new().with_alpha(1.0);
    ridge.fit(&x, &y).unwrap();
    let pred = ridge.predict(&x).unwrap();
    assert_eq!(pred.len(), 4);
}

#[test]
fn ridge_solvers_agree() {
    let (x, y) = sample_xy();
    let mut chol = Ridge::new().with_alpha(5.0);
    chol.fit(&x, &y).unwrap();
    let mut svd = Ridge::new().with_alpha(5.0).with_solver(RidgeSolver::Svd);
    svd.fit(&x, &y).unwrap();
    for (c, s) in chol.coef().iter().zip(svd.coef().iter()) {
        assert!(approx(*c, *s, 1e-6), "cholesky={c} svd={s}");
    }
    assert!(approx(chol.intercept(), svd.intercept(), 1e-6));
}

#[test]
fn ridge_score_high_for_clean_signal() {
    let (x, y) = sample_xy();
    let mut m = Ridge::new().with_alpha(0.01);
    m.fit(&x, &y).unwrap();
    let r2 = m.score(&x, &y).unwrap();
    assert!(r2 > 0.99, "r2={r2}");
}

#[test]
fn ridge_predict_new_data() {
    let rows: Vec<Vec<f64>> = (0..20)
        .map(|i| {
            let i = i as f64;
            vec![i]
        })
        .collect();
    let x = Matrix::new(rows).unwrap();
    let y: Vec<f64> = (0..20).map(|i| 2.0 * i as f64 + 1.0).collect();
    let mut m = Ridge::new().with_alpha(0.001);
    m.fit(&x, &y).unwrap();
    let new_x = Matrix::new(vec![vec![100.0]]).unwrap();
    let pred = m.predict(&new_x).unwrap();
    assert!(approx(pred[0], 201.0, 1e-1));
}

#[test]
fn ridge_is_fitted_flag() {
    let (x, y) = sample_xy();
    let mut m = Ridge::new();
    assert!(!m.is_fitted());
    m.fit(&x, &y).unwrap();
    assert!(m.is_fitted());
}

// ============================== Lasso =====================================

#[test]
fn lasso_drives_irrelevant_feature_to_zero() {
    // Feature 1 is pure noise (random-ish), feature 0 carries the signal.
    let rows: Vec<Vec<f64>> = (0..60)
        .map(|i| {
            let i = i as f64;
            // feature 0 = i, feature 1 = unrelated noise.
            vec![i, (i * 17.0).fract()]
        })
        .collect();
    let y: Vec<f64> = rows.iter().map(|r| 3.0 * r[0]).collect();
    let x = Matrix::new(rows).unwrap();

    let mut model = Lasso::new().with_alpha(1.0).with_max_iter(2000);
    model.fit(&x, &y).unwrap();
    assert!(
        model.coef()[1].abs() < 1e-6,
        "irrelevant feature should be zeroed: {}",
        model.coef()[1]
    );
}

#[test]
fn lasso_zero_alpha_fits_signal() {
    let (x, y) = sample_xy();
    let mut m = Lasso::new()
        .with_alpha(0.0)
        .with_max_iter(5000)
        .with_tol(1e-10);
    m.fit(&x, &y).unwrap();
    let r2 = m.score(&x, &y).unwrap();
    assert!(r2 > 0.99, "r2={r2}");
}

#[test]
fn lasso_sparsity_increases_with_alpha() {
    let (x, y) = sample_xy();
    let mut zero_counts = Vec::new();
    for &alpha in &[0.01, 1.0, 10.0, 100.0] {
        let mut m = Lasso::new().with_alpha(alpha).with_max_iter(2000);
        m.fit(&x, &y).unwrap();
        let zeros = m.coef().iter().filter(|c| c.abs() < 1e-10).count();
        zero_counts.push(zeros);
    }
    // Sparsity (zero count) should be non-decreasing with alpha.
    for w in zero_counts.windows(2) {
        assert!(w[1] >= w[0], "sparsity not increasing: {zero_counts:?}");
    }
}

#[test]
fn lasso_n_iter_in_range() {
    let (x, y) = sample_xy();
    let mut m = Lasso::new().with_alpha(0.1).with_max_iter(500);
    m.fit(&x, &y).unwrap();
    assert!(m.n_iter() > 0 && m.n_iter() <= 500);
}

#[test]
fn lasso_predict_before_fit_errors() {
    let model = Lasso::new();
    let x = Matrix::new(vec![vec![1.0]]).unwrap();
    assert!(matches!(
        model.predict(&x).unwrap_err(),
        datarust::DatarustError::NotFitted(_)
    ));
}

#[test]
fn lasso_negative_alpha_rejected() {
    let x = Matrix::new(vec![vec![1.0], vec![2.0]]).unwrap();
    let mut m = Lasso::new().with_alpha(-0.5);
    assert!(matches!(
        m.fit(&x, &[1.0, 2.0]).unwrap_err(),
        datarust::DatarustError::InvalidConfig(_)
    ));
}

#[test]
fn lasso_intercept_recovered() {
    let rows: Vec<Vec<f64>> = (0..30).map(|i| vec![i as f64]).collect();
    let x = Matrix::new(rows).unwrap();
    let y: Vec<f64> = (0..30).map(|i| 2.0 * i as f64 + 5.0).collect();
    let mut m = Lasso::new()
        .with_alpha(0.001)
        .with_max_iter(5000)
        .with_tol(1e-10);
    m.fit(&x, &y).unwrap();
    assert!(approx(m.coef()[0], 2.0, 1e-2));
    assert!(approx(m.intercept(), 5.0, 1e-1));
}

#[test]
fn ridge_and_lasso_implement_regressor_trait() {
    // Compile-time check that both can be used through the trait.
    fn predict_via_trait<R: Regressor>(m: &R, x: &Matrix) -> usize {
        Predictor::predict(m, x).map(|v| v.len()).unwrap_or(0)
    }
    fn name_via_trait<R: Regressor>(m: &R) -> &'static str {
        m.name()
    }
    let (x, y) = sample_xy();
    let mut ridge = Ridge::new().with_alpha(0.1);
    ridge.fit(&x, &y).unwrap();
    let mut lasso = Lasso::new().with_alpha(0.1);
    lasso.fit(&x, &y).unwrap();
    assert_eq!(predict_via_trait(&ridge, &x), x.nrows());
    assert_eq!(predict_via_trait(&lasso, &x), x.nrows());
    assert_eq!(name_via_trait(&ridge), "Ridge");
    assert_eq!(name_via_trait(&lasso), "Lasso");
}