forester 0.0.3

A crate for implementing various flavors of random forests and decision trees.
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
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328

//! Module for working with splits.

use std::f64;

use rand::{thread_rng, Rng};

use criterion::SplitCriterion;
use data::{SampleDescription, TrainingData};
use split_between::SplitBetween;

/// Parametric representation of a split.
///
/// A split consists of a data-set dependent parameter `theta` that corresponds to a feature, and
/// a threshold to split the feature into two half-spaces.
#[derive(Debug)]
pub struct Split<Theta, Threshold> {
    pub theta: Theta,
    pub threshold: Threshold,
}

/// Find split
pub trait SplitFinder
{
    /// Attempt to find a split for the given data set.
    fn find_split<Sample, Training>(&self, data: &mut Training)
        -> Option<Split<Sample::ThetaSplit, Sample::Feature>>
        where Sample: SampleDescription,
              Training: ?Sized + TrainingData<Sample>;
}

/// Find best random split.
///
/// This structure is used to find an optimal random split. It generates `n_splits` random splits
/// and selects the best one according to a data-dependent criterion. A higher number of splits
/// increases the chance of finding a more optimal split but reduces diversity and is
/// computationally more demanding.
pub struct BestRandomSplit {
    n_splits: usize,
}

impl BestRandomSplit {
    pub fn new(n_splits: usize) -> Self {
        BestRandomSplit {
            n_splits
        }
    }
}

impl SplitFinder for BestRandomSplit
{
    fn find_split<Sample, Training>(&self, data: &mut Training)
                                    -> Option<Split<Sample::ThetaSplit, Sample::Feature>>
        where Sample: SampleDescription,
              Training: ?Sized + TrainingData<Sample>
    {
        let n = data.n_samples() as f64;
        //let mut best_criterion = data.split_criterion();
        let mut best_criterion = Training::Criterion::from_dataset(data).get();
        let mut best_split = None;

        let mut rng = thread_rng();

        for _ in 0..self.n_splits {
            let theta = data.gen_split_feature();

            let (min, max) = data.feature_bounds(&theta);

            if max <= min {
                continue
            }

            let threshold = rng.gen_range(min, max);

            let split = Split{theta, threshold};
            let (left, right) = data.partition_data(&split);

            let left_crit = Training::Criterion::from_dataset(left).get_weighted();
            let right_crit = Training::Criterion::from_dataset(right).get_weighted();
            let criterion = (left_crit + right_crit) / n;

            if criterion <= best_criterion {
                best_criterion = criterion;
                best_split = Some(split);
            }

            // stop early if we find a perfect split
            // TODO: tolerance rather than exact comparison?
            if best_criterion == 0.0 {
                break
            }
        }

        best_split
    }
}

/// Find best split, in a number of randomly selected features.
///
/// Normally, exactly `n_features` are tested. However, there are two notable exceptions:
/// 1. if a "perfect split" is encountered (split criterion is exactly zero) the search stops early.
/// 2. if after no split has been found after `n_features` the search continues until it finds one
///    or runs out of features.
pub struct BestSplitRandomFeature {
    n_features: usize,
}

impl BestSplitRandomFeature {
    pub fn new(n_features: usize) -> Self {
        BestSplitRandomFeature {
            n_features
        }
    }
}

impl SplitFinder for BestSplitRandomFeature
{
    fn find_split<Sample, Training>(&self, data: &mut Training)
                                    -> Option<Split<Sample::ThetaSplit, Sample::Feature>>
        where Sample: SampleDescription,
              Training: ?Sized + TrainingData<Sample>
    {
        let mut best_criterion = Training::Criterion::from_dataset(data).get();
        let mut best_split = None;

        let mut features: Vec<_> = data
            .all_split_features()
            .expect("Dataset does not support iteration over features.")
            .collect();

        thread_rng().shuffle(&mut features);

        let mut n_to_check = self.n_features;

        for theta in features {
            if best_criterion == 0.0 {
                break
            }

            if n_to_check == 0 {
                if best_split.is_some() {
                    break
                }
            } else {
                n_to_check -= 1;
            }

            let (criterion, split) = find_best_split_for_feature(data, &theta);

            if criterion <= best_criterion {
                best_criterion = criterion;
                best_split = split;
            }

        }

        best_split
    }
}

/// Find best split.
///
/// Finds best split among all possible splits in all features. This only works for data sets that
/// implement `all_split_features` to return not `None`.
pub struct BestSplit {
}

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

impl SplitFinder for BestSplit
{
    fn find_split<Sample, Training>(&self, data: &mut Training)
                                    -> Option<Split<Sample::ThetaSplit, Sample::Feature>>
        where Sample: SampleDescription,
              Training: ?Sized + TrainingData<Sample>
    {
        let mut best_criterion = Training::Criterion::from_dataset(data).get();
        let mut best_split = None;

        let features = data.all_split_features().expect("Dataset does not support iteration over features.");

        for theta in features {
            if best_criterion == 0.0 {
                break
            }

            let (criterion, split) = find_best_split_for_feature(data, &theta);

            if criterion <= best_criterion {
                best_criterion = criterion;
                best_split = split;
            }
        }

        best_split
    }
}

/// find optimal split for given feature
fn find_best_split_for_feature<Sample, Training>(data: &mut Training, theta: &Sample::ThetaSplit)
                                                 -> (f64, Option<Split<Sample::ThetaSplit, Sample::Feature>>)
    where Sample: SampleDescription,
          Training: ?Sized + TrainingData<Sample>
{
    data.sort_data(theta);

    let mut best_criterion = f64::INFINITY;
    let mut best_split = None;

    let mut left_crit = Training::Criterion::from_dataset(data);
    let mut right_crit = Training::Criterion::new();

    let mut prev_sf: Option<Sample::Feature> = None;
    data.visit_samples(|sample| {
        let sf = sample.sample_as_split_feature(theta);

        // don't try to split between two samples of same feature value
        if let Some(ref psf) = prev_sf {
            if psf == &sf {
                left_crit.remove_sample(sample);
                right_crit.add_sample(sample);
                return
            }
        }

        let criterion = (left_crit.get_weighted() + right_crit.get_weighted()) / data.n_samples() as f64;

        if criterion <= best_criterion {
            if let Some(ref psf) = prev_sf {
                best_criterion = criterion;
                best_split = Some(Split {
                    theta: theta.clone(),
                    threshold: psf.split_between(&sf)
                });
            }
        }

        left_crit.remove_sample(sample);
        right_crit.add_sample(sample);

        prev_sf = Some(sf);
    });

    (best_criterion, best_split)
}


#[cfg(test)]
mod tests {
    use super::*;
    use split::BestRandomSplit;
    use testdata::Sample;

    #[test]
    fn best_random_split() {
        let data: &mut [_] = &mut [
            Sample::new(&[0.0], 1.0),
            Sample::new(&[1.0], 2.0),
        ];
        let spl = BestRandomSplit::new(1);
        let split = spl.find_split(data).unwrap();
        assert_eq!(split.theta, 0);
        assert!(split.threshold >= 0.0);
        assert!(split.threshold <= 1.0);

        let data: &mut [_] = &mut [
            Sample::new(&[41.0, 0.0], 1.0),
            Sample::new(&[41.0, 1.0], 2.0),
            Sample::new(&[43.0, 2.0], 1.0),
            Sample::new(&[43.0, 3.0], 2.0),
            Sample::new(&[41.0, 4.0], 11.0),
            Sample::new(&[41.0, 5.0], 12.0),
            Sample::new(&[43.0, 6.0], 11.0),
            Sample::new(&[43.0, 7.0], 12.0),
        ];
        let spl = BestRandomSplit::new(100);
        let split = spl.find_split(data).unwrap();
        assert_eq!(split.theta, 1);
        assert!(split.threshold >= 3.0);
        assert!(split.threshold <= 4.0);
    }

    #[test]
    fn best_split() {
        let data: &mut [_] = &mut [
            Sample::new(&[0.0, 41.0, 0.0], 1.0),
            Sample::new(&[0.0, 41.0, 1.0], 2.0),
            Sample::new(&[0.0, 43.0, 2.0], 1.0),
            Sample::new(&[0.0, 43.0, 3.0], 2.0),
            Sample::new(&[0.0, 41.0, 4.0], 11.0),
            Sample::new(&[0.0, 41.0, 5.0], 12.0),
            Sample::new(&[0.0, 43.0, 6.0], 11.0),
            Sample::new(&[0.0, 43.0, 7.0], 12.0),
            Sample::new(&[0.0, 42.0, 8.0], 11.0),
            Sample::new(&[0.0, 42.0, 9.0], 12.0),
        ];
        let spl = BestSplit::new();
        let split = spl.find_split(data).unwrap();
        assert_eq!(split.theta, 2);
        assert!(split.threshold >= 3.0);
        assert!(split.threshold <= 4.0);
    }

    #[test]
    fn best_split_random_feature() {
        let data: &mut [_] = &mut [
            Sample::new(&[0.0, 41.0, 0.0], 1.0),
            Sample::new(&[0.0, 41.0, 1.0], 2.0),
            Sample::new(&[0.0, 43.0, 2.0], 1.0),
            Sample::new(&[0.0, 43.0, 3.0], 2.0),
            Sample::new(&[0.0, 41.0, 4.0], 11.0),
            Sample::new(&[0.0, 41.0, 5.0], 12.0),
            Sample::new(&[0.0, 43.0, 6.0], 11.0),
            Sample::new(&[0.0, 43.0, 7.0], 12.0),
            Sample::new(&[0.0, 42.0, 8.0], 11.0),
            Sample::new(&[0.0, 42.0, 9.0], 12.0),
        ];
        let spl = BestSplitRandomFeature::new(3);
        let split = spl.find_split(data).unwrap();
        assert_eq!(split.theta, 2);
        assert!(split.threshold >= 3.0);
        assert!(split.threshold <= 4.0);
    }
}