rankit 0.1.4

Learning to Rank: differentiable ranking, LTR losses (RankNet, LambdaRank, ApproxNDCG, ListNet, ListMLE), trainers, and IR evaluation metrics
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

//! Property tests for core ranking operations.

use proptest::prelude::*;

proptest! {
    #[test]
    fn soft_rank_preserves_ordering(
        a in -100.0..100.0_f64,
        b in -100.0..100.0_f64,
        reg in 0.1..10.0_f64,
    ) {
        let values = vec![a, b];
        let ranks = crate::rank::soft_rank(&values, reg);

        if a > b + 1e-6 {
            prop_assert!(ranks[0] > ranks[1], "a={}, b={}, ranks={:?}", a, b, ranks);
        } else if b > a + 1e-6 {
            prop_assert!(ranks[1] > ranks[0], "a={}, b={}, ranks={:?}", a, b, ranks);
        }
    }

    #[test]
    fn soft_rank_in_range(
        values in prop::collection::vec(-100.0..100.0_f64, 1..20),
        reg in 0.1..10.0_f64,
    ) {
        let ranks = crate::rank::soft_rank(&values, reg);
        let n = values.len();

        for &r in &ranks {
            prop_assert!(r >= -0.01, "rank {} below 0", r);
            prop_assert!(r <= (n - 1) as f64 + 0.01, "rank {} above n-1={}", r, n - 1);
        }
    }

    #[test]
    fn sigmoid_bounded(x in -1000.0..1000.0_f64) {
        let s = crate::rank::sigmoid(x);
        prop_assert!((0.0..=1.0).contains(&s), "sigmoid({}) = {} out of [0,1]", x, s);
    }

    // -----------------------------------------------------------------------
    // Natural gradient property tests
    // -----------------------------------------------------------------------

    /// Fisher information matrix is symmetric for any probability distribution.
    #[test]
    fn fisher_is_symmetric(
        raw in prop::collection::vec(0.1_f64..10.0, 2..10),
    ) {
        // Normalize to valid probability distribution.
        let sum: f64 = raw.iter().sum();
        let probs: Vec<f64> = raw.iter().map(|&x| x / sum).collect();
        let n = probs.len();

        let fisher = crate::gradients::natural::fisher_information_softmax(&probs);
        for i in 0..n {
            for j in i+1..n {
                let diff = (fisher[i * n + j] - fisher[j * n + i]).abs();
                prop_assert!(diff < 1e-14, "Not symmetric at [{i}][{j}]: diff={diff}");
            }
        }
    }

    /// Fisher information matrix is PSD: v^T F v >= 0 for random v.
    #[test]
    fn fisher_is_psd(
        raw in prop::collection::vec(0.1_f64..10.0, 2..8),
        v in prop::collection::vec(-10.0..10.0_f64, 2..8),
    ) {
        if raw.len() != v.len() {
            return Ok(());
        }
        let sum: f64 = raw.iter().sum();
        let probs: Vec<f64> = raw.iter().map(|&x| x / sum).collect();
        let n = probs.len();

        let fisher = crate::gradients::natural::fisher_information_softmax(&probs);
        let mut vtfv = 0.0;
        for i in 0..n {
            for j in 0..n {
                vtfv += v[i] * fisher[i * n + j] * v[j];
            }
        }
        prop_assert!(vtfv >= -1e-10, "v^T F v = {vtfv} < 0");
    }

    /// Natural gradient is finite for all valid inputs.
    #[test]
    fn natural_gradient_is_finite(
        raw_probs in prop::collection::vec(0.1_f64..10.0, 2..8),
        grad in prop::collection::vec(-5.0..5.0_f64, 2..8),
    ) {
        if raw_probs.len() != grad.len() {
            return Ok(());
        }
        let sum: f64 = raw_probs.iter().sum();
        let probs: Vec<f64> = raw_probs.iter().map(|&x| x / sum).collect();

        let nat_grad = crate::gradients::natural::natural_gradient_softmax(&grad, &probs);

        for (i, &ng) in nat_grad.iter().enumerate() {
            prop_assert!(ng.is_finite(), "natural_grad[{i}] is not finite: {ng}");
        }
    }

    /// Fisher-natural gradient identity: g^T F^{-1} g = sum(g_i^2 / p_i) - (sum g_i)^2.
    /// Verify by computing g^T * nat_grad (which equals g^T F^{-1} g).
    #[test]
    fn natural_gradient_quadratic_form(
        raw_probs in prop::collection::vec(0.1_f64..10.0, 2..8),
        grad in prop::collection::vec(-5.0..5.0_f64, 2..8),
    ) {
        if raw_probs.len() != grad.len() {
            return Ok(());
        }
        let sum: f64 = raw_probs.iter().sum();
        let probs: Vec<f64> = raw_probs.iter().map(|&x| x / sum).collect();

        let nat_grad = crate::gradients::natural::natural_gradient_softmax(&grad, &probs);

        // g^T * nat_grad
        let dot: f64 = grad.iter().zip(nat_grad.iter()).map(|(&g, &ng)| g * ng).sum();

        // Expected: sum(g_i^2 / p_i) - (sum g_i)^2
        let sum_g: f64 = grad.iter().sum();
        let expected: f64 = grad.iter().zip(probs.iter()).map(|(&g, &p)| g * g / p).sum::<f64>() - sum_g * sum_g;

        prop_assert!(
            (dot - expected).abs() < 1e-6,
            "g^T nat_grad ({dot}) != expected ({expected})"
        );
    }

    #[test]
    fn ranknet_loss_nonnegative(
        predictions in prop::collection::vec(-10.0..10.0_f64, 2..10),
        relevance in prop::collection::vec(0.0..5.0_f64, 2..10),
    ) {
        if predictions.len() != relevance.len() {
            return Ok(());
        }

        let loss = crate::losses::ranknet_loss(&predictions, &relevance);
        prop_assert!(loss >= 0.0, "ranknet_loss should be >= 0, got {}", loss);
    }
}