pkboost 2.2.0

Shannon-guided gradient boosting for extreme class imbalance with adaptive drift detection. Outperforms XGBoost/LightGBM on imbalanced data.
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

/// Loss functions for PKBoost
/// Provides gradient and hessian calculations for Newton-Raphson optimization
use rayon::prelude::*;
use serde::{Deserialize, Serialize};

/// Parallel threshold - avoid Rayon overhead for small arrays
const PARALLEL_THRESHOLD: usize = 10000;

#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct OptimizedShannonLoss;

impl OptimizedShannonLoss {
    pub fn new() -> Self {
        Self
    }

    pub fn init_score(&self, y: &[f64]) -> f64 {
        let pos = y.iter().filter(|&&v| v > 0.5).count() as f64;
        let neg = y.len() as f64 - pos;
        (pos / neg.max(1.0)).ln()
    }

    /// OPTIMIZED: Fused gradient and hessian calculation
    /// Computes both in a single pass, avoiding redundant sigmoid calculations
    #[inline]
    pub fn gradient_hessian(
        &self,
        y_true: &[f64],
        y_pred: &[f64],
        scale_pos_weight: f64,
    ) -> (Vec<f64>, Vec<f64>) {
        let n = y_true.len();

        if n >= PARALLEL_THRESHOLD {
            // Parallel path for large arrays
            let results: Vec<(f64, f64)> = y_pred
                .par_iter()
                .zip(y_true.par_iter())
                .map(|(&pred, &true_y)| {
                    let prob = 1.0 / (1.0 + (-pred).exp());
                    let weight = if true_y > 0.5 { scale_pos_weight } else { 1.0 };
                    let grad = weight * (prob - true_y);
                    let hess = weight * prob * (1.0 - prob).max(1e-6);
                    (grad, hess)
                })
                .collect();

            let mut grad = Vec::with_capacity(n);
            let mut hess = Vec::with_capacity(n);
            for (g, h) in results {
                grad.push(g);
                hess.push(h);
            }
            (grad, hess)
        } else {
            // Sequential path for small arrays - avoids Rayon overhead
            let mut grad = Vec::with_capacity(n);
            let mut hess = Vec::with_capacity(n);

            for (&pred, &true_y) in y_pred.iter().zip(y_true.iter()) {
                let prob = 1.0 / (1.0 + (-pred).exp());
                let weight = if true_y > 0.5 { scale_pos_weight } else { 1.0 };
                grad.push(weight * (prob - true_y));
                hess.push(weight * prob * (1.0 - prob).max(1e-6));
            }
            (grad, hess)
        }
    }

    /// Legacy: separate gradient (for backward compatibility)
    pub fn gradient(&self, y_true: &[f64], y_pred: &[f64], scale_pos_weight: f64) -> Vec<f64> {
        self.gradient_hessian(y_true, y_pred, scale_pos_weight).0
    }

    /// Legacy: separate hessian (for backward compatibility)
    pub fn hessian(&self, y_true: &[f64], y_pred: &[f64], scale_pos_weight: f64) -> Vec<f64> {
        self.gradient_hessian(y_true, y_pred, scale_pos_weight).1
    }

    pub fn sigmoid(&self, preds: &[f64]) -> Vec<f64> {
        if preds.len() >= PARALLEL_THRESHOLD {
            preds
                .par_iter()
                .map(|&p| 1.0 / (1.0 + (-p).exp()))
                .collect()
        } else {
            preds.iter().map(|&p| 1.0 / (1.0 + (-p).exp())).collect()
        }
    }
}

pub struct PoissonLoss;

impl PoissonLoss {
    /// Compute gradient and hessian for Poisson regression
    /// Gradient: exp(f) - y
    /// Hessian: exp(f)
    pub fn gradient_hessian(y_true: &[f64], y_pred: &[f64]) -> (Vec<f64>, Vec<f64>) {
        let mut grad = Vec::with_capacity(y_true.len());
        let mut hess = Vec::with_capacity(y_true.len());

        for (&y, &f) in y_true.iter().zip(y_pred.iter()) {
            let exp_f = f.exp().min(1e15); // Prevent overflow
            grad.push(exp_f - y);
            hess.push(exp_f.max(1e-6)); // Prevent zero hessian
        }
        (grad, hess)
    }

    /// Compute Poisson deviance loss
    pub fn loss(y_true: &[f64], y_pred: &[f64]) -> f64 {
        y_true
            .iter()
            .zip(y_pred.iter())
            .map(|(&y, &f)| {
                let exp_f = f.exp().min(1e15);
                exp_f - y * f
            })
            .sum::<f64>()
            / y_true.len() as f64
    }
}

pub struct MSELoss;

impl MSELoss {
    pub fn gradient_hessian(y_true: &[f64], y_pred: &[f64]) -> (Vec<f64>, Vec<f64>) {
        let grad: Vec<f64> = y_pred
            .iter()
            .zip(y_true.iter())
            .map(|(&pred, &true_y)| pred - true_y)
            .collect();
        let hess = vec![1.0; y_true.len()];
        (grad, hess)
    }

    pub fn loss(y_true: &[f64], y_pred: &[f64]) -> f64 {
        y_true
            .iter()
            .zip(y_pred.iter())
            .map(|(&y, &pred)| (y - pred).powi(2))
            .sum::<f64>()
            / y_true.len() as f64
    }
}

pub struct HuberLoss {
    pub delta: f64,
}

impl HuberLoss {
    pub fn new(delta: f64) -> Self {
        Self { delta }
    }

    pub fn gradient_hessian(&self, y_true: &[f64], y_pred: &[f64]) -> (Vec<f64>, Vec<f64>) {
        let mut grad = Vec::with_capacity(y_true.len());
        let mut hess = Vec::with_capacity(y_true.len());

        for (&y, &pred) in y_true.iter().zip(y_pred.iter()) {
            let residual = pred - y;
            if residual.abs() <= self.delta {
                grad.push(residual);
                hess.push(1.0);
            } else {
                grad.push(self.delta * residual.signum());
                hess.push(1e-6); // Small constant for outliers
            }
        }
        (grad, hess)
    }
}

#[derive(Debug, Clone, Copy)]
pub enum LossType {
    MSE,
    Huber,
    Poisson,
}

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

    #[test]
    fn test_poisson_gradient() {
        let y_true = vec![0.0, 1.0, 3.0, 2.0];
        let y_pred = vec![0.1, 0.5, 1.2, 0.8];
        let (grad, hess) = PoissonLoss::gradient_hessian(&y_true, &y_pred);

        // exp(0.1) - 0 ≈ 1.105
        assert!((grad[0] - 1.105).abs() < 0.01);
        // exp(0.5) - 1 ≈ 0.649
        assert!((grad[1] - 0.649).abs() < 0.01);

        // Hessian should be exp(f)
        assert!((hess[0] - 0.1_f64.exp()).abs() < 0.01);
    }

    #[test]
    fn test_poisson_loss() {
        let y_true = vec![1.0, 2.0, 3.0];
        let y_pred = vec![0.0, 0.69, 1.10]; // log(1), log(2), log(3)
        let loss = PoissonLoss::loss(&y_true, &y_pred);
        assert!(loss < 0.5); // Should be small for good predictions
    }
}