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

 #![warn(missing_docs)]

 /*!
Measure classifier's performance using [Receiver Operating
Characteristic](https://en.wikipedia.org/wiki/Receiver_operating_characteristic)
(ROC) and [Precision-Recall](https://en.wikipedia.org/wiki/Precision_and_recall)
(PR) curves.

The curves themselves can be computed as well as trapezoidal areas under the curves.
 */

extern crate num_traits;
use num_traits::{Float, NumCast};

mod test;

/// Integration using the trapezoidal rule.
fn trapezoidal<F: Float>(x: &[F], y: &[F]) -> F {
    let mut prev_x = x[0];
    let mut prev_y = y[0];
    let mut integral = F::zero();

    for (&x, &y) in x.iter().skip(1).zip(y.iter().skip(1)) {
        integral = integral + (x - prev_x) * (prev_y + y) / NumCast::from(2.0).unwrap();
        prev_x = x;
        prev_y = y;
    }
    integral
}

/// Checks if all data-points are finite
fn check_data<F: Float>(v: &[(bool, F)]) -> bool {
    v.iter().any(|x| x.0) && v.iter().any(|x| !x.0) && v.iter().all(|x| x.1.is_finite())
}

/// Uses a provided closure to convert free-form data into a standard format
fn convert<T, X, F, I>(data: I, convert_fn: F) -> Vec<(bool, X)> where
        I: IntoIterator<Item=T>,
        F: Fn(T) -> (bool, X),
        X: Float {
    let data_it = data.into_iter();
    let mut v = Vec::with_capacity(data_it.size_hint().0);
    for i in data_it {
        v.push(convert_fn(i));
    }
    v
}

/// Computes a ROC curve of a given classifier sorting `pairs` in-place.
///
/// Returns `None` if one of the classes is not present or any values are non-finite.
/// Otherwise, returns `Some((v_x, v_y))` where `v_x` are the x-coordinates and `v_y` are the
/// y-coordinates of the ROC curve.
pub fn roc_mut<F: Float>(pairs: &mut [(bool, F)]) -> Option<(Vec<F>, Vec<F>)> {
    if !check_data(pairs) {
        return None;
    }
    pairs.sort_by(&|x: &(_, F), y: &(_, F)|
        match y.1.partial_cmp(&x.1) {
            Some(ord) => ord,
            None      => unreachable!(),
        });

    let mut s0 = F::nan();
    let (mut tp, mut fp) = (F::zero(), F::zero());
    let (mut tps, mut fps) = (vec![], vec![]);
    for &(t, s) in pairs.iter() {
        if s != s0 {
            tps.push(tp);
            fps.push(fp);
            s0 = s;
        }
        match t {
            false => fp = fp + F::one(),
            true =>  tp = tp + F::one(),
        }
    }
    tps.push(tp);
    fps.push(fp);

    // normalize
    let (tp_max, fp_max) = (tps[tps.len() - 1], fps[fps.len() - 1]);
    for mut tp in &mut tps {
        *tp = *tp / tp_max;
    }
    for mut fp in &mut fps {
        *fp = *fp / fp_max;
    }
    Some((fps, tps))
}

/// Computes a ROC curve of a given classifier.
///
/// `data` is a free-form `IntoIterator` object and `convert_fn` is a closure that converts each
/// data-point into a pair `(ground_truth, prediction).`
///
/// Returns `None` if one of the classes is not present or any values are non-finite.
/// Otherwise, returns `Some((v_x, v_y))` where `v_x` are the x-coordinates and `v_y` are the
/// y-coordinates of the ROC curve.
pub fn roc<T, X, F, I>(data: I, convert_fn: F) -> Option<(Vec<X>, Vec<X>)> where
        I: IntoIterator<Item=T>,
        F: Fn(T) -> (bool, X),
        X: Float {
    roc_mut(&mut convert(data, convert_fn))
}

/// Computes a PR curve of a given classifier.
///
/// `data` is a free-form `IntoIterator` object and `convert_fn` is a closure that converts each
/// data-point into a pair `(ground_truth, prediction).`
///
/// Returns `None` if one of the classes is not present or any values are non-finite.
/// Otherwise, returns `Some((v_x, v_y))` where `v_x` are the x-coordinates and `v_y` are the
/// y-coordinates of the PR curve.
pub fn pr<T, X, F, I>(data: I, convert_fn: F) -> Option<(Vec<X>, Vec<X>)> where
        I: IntoIterator<Item=T>,
        F: Fn(T) -> (bool, X),
        X: Float {
    pr_mut(&mut convert(data, convert_fn))
}

/// Computes a PR curve of a given classifier sorting `pairs` in-place.
///
/// Returns `None` if one of the classes is not present or any values are non-finite.
/// Otherwise, returns `Some((v_x, v_y))` where `v_x` are the x-coordinates and `v_y` are the
/// y-coordinates of the PR curve.
pub fn pr_mut<F: Float>(pairs: &mut [(bool, F)]) -> Option<(Vec<F>, Vec<F>)> {
    if !check_data(pairs) {
        return None;
    }
    pairs.sort_by(&|x: &(_, F), y: &(_, F)|
        match y.1.partial_cmp(&x.1) {
            Some(ord) => ord,
            None      => unreachable!(),
        });

    let mut x0 = F::nan();
    let (mut tp, mut p, mut fp) = (F::zero(), F::zero(), F::zero());
    let (mut recall, mut precision) = (vec![], vec![]);

    // number of labels
    let ln = pairs.iter().fold(F::zero(), |a,b| a + if b.0 { F::one() } else { F::zero() });

    for &(l, x) in pairs.iter() {
        if x != x0 {
            recall.push(tp / ln);
            precision.push(if p == F::zero() { F::one() } else { tp / (tp + fp) });
            x0 = x;
        }
        p = p + F::one();
        if l { tp = tp + F::one(); }
        else { fp = fp + F::one(); }
    }
    recall.push(tp / ln);
    precision.push(tp / p);

    Some((precision, recall))
}

/// Computes the area under a PR curve of a given classifier.
///
/// `data` is a free-form `IntoIterator` object and `convert_fn` is a closure that converts each
/// data-point into a pair `(ground_truth, prediction).`
///
/// Returns `None` if one of the classes is not present or any values are non-finite.
/// Otherwise, returns `Some(area_under_curve)`.
pub fn pr_auc<T, X, F, I>(data: I, convert_fn: F) -> Option<X> where
        I: IntoIterator<Item=T>,
        F: Fn(T) -> (bool, X),
        X: Float {
    pr_auc_mut(&mut convert(data, convert_fn))
}

/// Computes the area under a PR curve of a given classifier sorting `pairs` in-place.
///
/// Returns `None` if one of the classes is not present or any values are non-finite.
/// Otherwise, returns `Some(area_under_curve)`.
pub fn pr_auc_mut<F: Float>(pairs: &mut [(bool, F)]) -> Option<F> {
    pr_mut(pairs).map(|curve| {
        trapezoidal(&curve.1, &curve.0)
    })
}

/// Computes the area under a ROC curve of a given classifier.
///
/// `data` is a free-form `IntoIterator` object and `convert_fn` is a closure that converts each
/// data-point into a pair `(ground_truth, prediction).`
///
/// Returns `None` if one of the classes is not present or any values are non-finite.
/// Otherwise, returns `Some(area_under_curve)`.
pub fn roc_auc<T, X, F, I>(data: I, convert_fn: F) -> Option<X> where
        I: IntoIterator<Item=T>,
        F: Fn(T) -> (bool, X),
        X: Float {
    roc_auc_mut(&mut convert(data, convert_fn))
}

/// Computes the area under a ROC curve of a given classifier sorting `pairs` in-place.
///
/// Returns `None` if one of the classes is not present or any values are non-finite.
/// Otherwise, returns `Some(area_under_curve)`.
pub fn roc_auc_mut<F: Float>(pairs: &mut [(bool, F)]) -> Option<F> {
    roc_mut(pairs).map(|curve| trapezoidal(&curve.0, &curve.1))
}