perpetual 2.0.0

A self-generalizing gradient boosting machine that doesn't need hyperparameter optimization
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

//! Hinge Loss objective for binary classification.
use crate::objective::ObjectiveFunction;
use crate::{metrics::evaluation::Metric, utils::fast_sum};
use serde::{Deserialize, Serialize};

#[derive(Default, Debug, Deserialize, Serialize, Clone)]
/// Hinge Loss objective.
/// `max(0, 1 - y_binary * yhat)`, where y_binary maps {0, 1} targets to {-1, 1}.
pub struct HingeLoss {}

impl ObjectiveFunction for HingeLoss {
    #[inline]
    fn loss(&self, y: &[f64], yhat: &[f64], sample_weight: Option<&[f64]>, _group: Option<&[u64]>) -> Vec<f32> {
        match sample_weight {
            Some(sample_weight) => y
                .iter()
                .zip(yhat)
                .zip(sample_weight)
                .map(|((y_, yhat_), w_)| {
                    let y_bin = if *y_ > 0.0 { 1.0 } else { -1.0 };
                    let diff = 1.0 - y_bin * yhat_;
                    let loss = if diff > 0.0 { diff } else { 0.0 };
                    (loss * w_) as f32
                })
                .collect(),
            None => y
                .iter()
                .zip(yhat)
                .map(|(y_, yhat_)| {
                    let y_bin = if *y_ > 0.0 { 1.0 } else { -1.0 };
                    let diff = 1.0 - y_bin * yhat_;
                    let loss = if diff > 0.0 { diff } else { 0.0 };
                    loss as f32
                })
                .collect(),
        }
    }

    #[inline]
    fn gradient(
        &self,
        y: &[f64],
        yhat: &[f64],
        sample_weight: Option<&[f64]>,
        _group: Option<&[u64]>,
    ) -> (Vec<f32>, Option<Vec<f32>>) {
        let len = y.len();
        let mut g = Vec::with_capacity(len);
        let mut h = Vec::with_capacity(len);

        // Hinge loss is piece-wise linear:
        // gradient is -y_bin if y_bin * yhat < 1, else 0.
        // hessian is 0 everywhere except undefined at 1. We use a smoothed approximation or a tiny constant.
        let h_val = 1e-6_f32;

        match sample_weight {
            Some(w) => {
                for i in 0..len {
                    let y_val = y[i] as f32;
                    let yhat_val = yhat[i] as f32;
                    let w_val = w[i] as f32;

                    let y_bin = if y_val > 0.0 { 1.0 } else { -1.0 };
                    let grad = if y_bin * yhat_val < 1.0 { -y_bin } else { 0.0 };

                    g.push(grad * w_val);
                    h.push(h_val * w_val);
                }
                (g, Some(h))
            }
            None => {
                for i in 0..len {
                    let y_val = y[i] as f32;
                    let yhat_val = yhat[i] as f32;

                    let y_bin = if y_val > 0.0 { 1.0 } else { -1.0 };
                    let grad = if y_bin * yhat_val < 1.0 { -y_bin } else { 0.0 };

                    g.push(grad);
                    h.push(h_val);
                }
                (g, Some(h))
            }
        }
    }

    #[inline]
    fn initial_value(&self, y: &[f64], sample_weight: Option<&[f64]>, _group: Option<&[u64]>) -> f64 {
        let mean_y = match sample_weight {
            Some(w) => {
                let mut ytot: f64 = 0.;
                let mut ntot: f64 = 0.;
                for i in 0..y.len() {
                    ytot += w[i] * if y[i] > 0.0 { 1.0 } else { 0.0 };
                    ntot += w[i];
                }
                ytot / ntot
            }
            None => {
                let ytot = fast_sum(
                    &y.iter()
                        .map(|&yi| if yi > 0.0 { 1.0 } else { 0.0 })
                        .collect::<Vec<f64>>(),
                );
                let ntot = y.len() as f64;
                ytot / ntot
            }
        };
        if mean_y > 0.5 {
            1.0 // Favor class 1
        } else if mean_y < 0.5 {
            -1.0 // Favor class -1
        } else {
            0.0
        }
    }

    fn default_metric(&self) -> Metric {
        Metric::LogLoss
    }

    fn requires_batch_evaluation(&self) -> bool {
        false
    }
}

impl HingeLoss {
    #[inline]
    pub fn loss_single(&self, y: f64, yhat: f64, sample_weight: Option<f64>) -> f32 {
        let y_bin = if y > 0.0 { 1.0 } else { -1.0 };
        let diff = 1.0 - y_bin * yhat;
        let l = if diff > 0.0 { diff } else { 0.0 };
        match sample_weight {
            Some(w) => (l * w) as f32,
            None => l as f32,
        }
    }
}

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

    #[test]
    fn test_hinge_loss_init() {
        let y = vec![1.0, 1.0, 0.0]; // Mean y = 2/3 > 0.5
        let loss_fn = HingeLoss::default();
        assert_eq!(loss_fn.initial_value(&y, None, None), 1.0);

        let y2 = vec![1.0, 0.0, 0.0]; // Mean y = 1/3 < 0.5
        assert_eq!(loss_fn.initial_value(&y2, None, None), -1.0);
    }

    #[test]
    fn test_hinge_loss() {
        let y = vec![1.0, 0.0]; // binary: 1, -1
        let yhat = vec![0.5, 0.5];
        let loss_fn = HingeLoss::default();

        let l = loss_fn.loss(&y, &yhat, None, None);
        // Loss[0]: max(0, 1 - 1*0.5) = 0.5
        // Loss[1]: max(0, 1 - (-1)*0.5) = 1.5
        assert_eq!(l, vec![0.5, 1.5]);

        let w = vec![2.0, 0.5];
        let lw = loss_fn.loss(&y, &yhat, Some(&w), None);
        assert_eq!(lw, vec![1.0, 0.75]);
    }

    #[test]
    fn test_hinge_gradient() {
        let y = vec![1.0, 0.0];
        let yhat = vec![0.5, 2.0];
        let loss_fn = HingeLoss::default();

        let (g, h) = loss_fn.gradient(&y, &yhat, None, None);
        // grad[0]: y_bin*yhat = 0.5 < 1 -> -y_bin = -1.0
        // grad[1]: y_bin*yhat = -1 * 2 = -2 < 1 -> -y_bin = 1.0
        assert_eq!(g, vec![-1.0, 1.0]);
        assert_eq!(h.unwrap(), vec![1e-6, 1e-6]);

        // Weighted
        let w = vec![2.0, 0.5];
        let (gw, hw) = loss_fn.gradient(&y, &yhat, Some(&w), None);
        assert_eq!(gw, vec![-2.0, 0.5]);
        assert_eq!(hw.unwrap(), vec![2e-6, 0.5e-6]);

        // Initial value weighted
        assert_eq!(loss_fn.initial_value(&[1.0, 0.0], Some(&[3.0, 1.0]), None), 1.0);
        assert_eq!(loss_fn.initial_value(&[1.0, 0.0], Some(&[1.0, 3.0]), None), -1.0);
        assert_eq!(loss_fn.initial_value(&[1.0, 0.0], Some(&[1.0, 1.0]), None), 0.0);
    }

    #[test]
    fn test_hinge_gradient_and_loss() {
        let y = vec![1.0, 0.0];
        let yhat = vec![0.5, 2.0];
        let loss_fn = HingeLoss::default();
        let (g, h, l) = loss_fn.gradient_and_loss(&y, &yhat, None, None);
        assert_eq!(g, vec![-1.0, 1.0]);
        assert_eq!(h.unwrap(), vec![1e-6, 1e-6]);
        assert_eq!(l, vec![0.5, 3.0]);
    }

    #[test]
    fn test_hinge_gradient_and_loss_into() {
        let y = vec![1.0, 0.0];
        let yhat = vec![0.5, 2.0];
        let loss_fn = HingeLoss::default();
        let mut grad = vec![0.0; 2];
        let mut hess = Some(vec![0.0; 2]);
        let mut loss = vec![0.0; 2];
        loss_fn.gradient_and_loss_into(&y, &yhat, None, None, &mut grad, &mut hess, &mut loss);
        assert_eq!(grad, vec![-1.0, 1.0]);
        assert_eq!(hess.unwrap(), vec![1e-6, 1e-6]);
        assert_eq!(loss, vec![0.5, 3.0]);
    }

    #[test]
    fn test_hinge_loss_single() {
        let loss_fn = HingeLoss::default();
        assert_eq!(loss_fn.loss_single(1.0, 0.5, None), 0.5);
        assert_eq!(loss_fn.loss_single(1.0, 0.5, Some(2.0)), 1.0);
    }
}