perpetual 2.1.0

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

use crate::data::FloatData;
use crate::metrics::evaluation::EvaluationMetric;

pub struct LogLossMetric {}
impl EvaluationMetric for LogLossMetric {
    fn calculate_metric(y: &[f64], yhat: &[f64], sample_weight: &[f64], _group: &[u64], _alpha: Option<f32>) -> f64 {
        log_loss(y, yhat, sample_weight)
    }
    fn maximize() -> bool {
        false
    }
}

pub struct BrierLossMetric {}
impl EvaluationMetric for BrierLossMetric {
    fn calculate_metric(y: &[f64], yhat: &[f64], sample_weight: &[f64], _group: &[u64], _alpha: Option<f32>) -> f64 {
        brier_loss(y, yhat, sample_weight)
    }
    fn maximize() -> bool {
        false
    }
}

pub struct AUCMetric {}
impl EvaluationMetric for AUCMetric {
    fn calculate_metric(y: &[f64], yhat: &[f64], sample_weight: &[f64], _group: &[u64], _alpha: Option<f32>) -> f64 {
        roc_auc_score(y, yhat, sample_weight)
    }
    fn maximize() -> bool {
        true
    }
}

pub fn log_loss(y: &[f64], yhat: &[f64], sample_weight: &[f64]) -> f64 {
    let mut w_sum = 0.;
    let res = y
        .iter()
        .zip(yhat)
        .zip(sample_weight)
        .map(|((y_, yhat_), w_)| {
            w_sum += *w_;
            let yhat_ = f64::ONE / (f64::ONE + (-*yhat_).exp());
            -(*y_ * yhat_.ln() + (f64::ONE - *y_) * ((f64::ONE - yhat_).ln())) * *w_
        })
        .sum::<f64>();
    res / w_sum
}

pub fn brier_loss(y: &[f64], yhat: &[f64], sample_weight: &[f64]) -> f64 {
    let mut w_sum = 0.;
    let res = y
        .iter()
        .zip(yhat)
        .zip(sample_weight)
        .map(|((y_, yhat_), w_)| {
            w_sum += *w_;
            let p = f64::ONE / (f64::ONE + (-*yhat_).exp());
            (*y_ - p).powi(2) * *w_
        })
        .sum::<f64>();
    res / w_sum
}

fn trapezoid_area(x0: f64, x1: f64, y0: f64, y1: f64) -> f64 {
    (x0 - x1).abs() * (y0 + y1) * 0.5
}

pub fn roc_auc_score(y: &[f64], yhat: &[f64], sample_weight: &[f64]) -> f64 {
    let mut indices = (0..y.len()).collect::<Vec<_>>();
    indices.sort_unstable_by(|&a, &b| yhat[b].total_cmp(&yhat[a]));
    let mut auc: f64 = 0.0;

    let mut label = y[indices[0]];
    let mut w = sample_weight[indices[0]];
    let mut fp = (1.0 - label) * w;
    let mut tp: f64 = label * w;
    let mut tp_prev: f64 = 0.0;
    let mut fp_prev: f64 = 0.0;

    for i in 1..indices.len() {
        if yhat[indices[i]] != yhat[indices[i - 1]] {
            auc += trapezoid_area(fp_prev, fp, tp_prev, tp);
            tp_prev = tp;
            fp_prev = fp;
        }
        label = y[indices[i]];
        w = sample_weight[indices[i]];
        fp += (1.0 - label) * w;
        tp += label * w;
    }

    auc += trapezoid_area(fp_prev, fp, tp_prev, tp);
    if fp <= 0.0 || tp <= 0.0 {
        auc = 0.0;
        fp = 0.0;
        tp = 0.0;
    }

    auc / (tp * fp)
}
#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_log_loss_metric() {
        let y = vec![1.0, 0.0];
        let yhat = vec![1.0, -1.0]; // p approx 0.73, 0.27
        let w = vec![1.0, 1.0];
        let loss = log_loss(&y, &yhat, &w);
        // p = 1/(1+e^-1) = 0.7310585
        // loss = -(1*ln(p) + 0*ln(1-p)) = 0.313261
        // For two samples with same diff: 0.313261
        assert!((loss - 0.3132617).abs() < 1e-6);
    }

    #[test]
    fn test_brier_loss_metric() {
        let y = vec![1.0, 0.0];
        let yhat = vec![0.0, 0.0]; // p = 0.5
        let w = vec![1.0, 1.0];
        let loss = brier_loss(&y, &yhat, &w);
        assert_eq!(loss, 0.25);
    }

    #[test]
    fn test_auc_metric() {
        let y = vec![1.0, 0.0, 1.0, 0.0];
        let yhat = vec![0.8, 0.2, 0.7, 0.1];
        let w = vec![1.0, 1.0, 1.0, 1.0];
        let auc = roc_auc_score(&y, &yhat, &w);
        assert_eq!(auc, 1.0);

        let yhat2 = vec![0.1, 0.9, 0.2, 0.8];
        let auc2 = roc_auc_score(&y, &yhat2, &w);
        assert_eq!(auc2, 0.0);
    }

    #[test]
    fn test_auc_metric_tied() {
        let y = vec![1.0, 0.0];
        let yhat = vec![0.5, 0.5];
        let w = vec![1.0, 1.0];
        let auc = roc_auc_score(&y, &yhat, &w);
        assert_eq!(auc, 0.5);
    }

    #[test]
    fn test_auc_all_same() {
        let y = vec![1.0, 1.0];
        let yhat = vec![0.5, 0.5];
        let w = vec![1.0, 1.0];
        assert!(roc_auc_score(&y, &yhat, &w).is_nan());

        let y2 = vec![0.0, 0.0];
        assert!(roc_auc_score(&y2, &yhat, &w).is_nan());
    }

    #[test]
    fn test_metric_traits() {
        let y = vec![1.0, 0.0];
        let yhat = vec![0.0, 0.0];
        let w = vec![1.0, 1.0];
        let g: Vec<u64> = vec![0, 0];

        assert!(!LogLossMetric::maximize());
        assert!(LogLossMetric::calculate_metric(&y, &yhat, &w, &g, None::<f32>) > 0.0);

        assert!(!BrierLossMetric::maximize());
        assert_eq!(BrierLossMetric::calculate_metric(&y, &yhat, &w, &g, None::<f32>), 0.25);

        assert!(AUCMetric::maximize());
        assert_eq!(AUCMetric::calculate_metric(&y, &yhat, &w, &g, None::<f32>), 0.5);
    }
}