Struct changeforest::ClassifierGain

pub struct ClassifierGain<T: Classifier> {
    pub classifier: T,
}

Fields

classifier: T

Trait Implementations

impl<T> ApproxGain for ClassifierGain<T> where
    T: Classifier, 

fn gain_approx(
    &self,
    start: usize,
    stop: usize,
    guess: usize,
    _: &[usize]
) -> ApproxGainResult

Return an approximation of the classifier-likelihood based gain when splitting segment [start, stop) for each split in split_candidates.

A single fit is generated with a split at guess.

impl<T> Gain for ClassifierGain<T> where
    T: Classifier, 

fn n(&self) -> usize

Total number of observations.

fn gain(&self, start: usize, stop: usize, split: usize) -> f64

Return classifier-likelihood based gain when splitting segment [start, stop) at split.

fn model_selection(
    &self,
    optimizer_result: &OptimizerResult
) -> ModelSelectionResult

Perform a permutation test.

We test whether the maximum observed gain from the first step in the TwoStepSearch optimizer is significant. Using the maximum gain from the first step instead allows us to do a proper permutation test with control of type I error without fitting additional classifiers.

In the first step of the TwoStepSearch optimizer, three gain curves and corresponding maximal gains are computed. The maximum gain of the first step of TwoStepSearch is the maximum of these three gains, which are available as the first three elements of optimizer_result.gain_results.

For each permutation, we shuffle the predictions (and thus the likelihoods) of each of the three initial classifier fits (using the same permutation), and compute the maximum of the three resulting maximal gains. We count the number of permutations where the resulting maximal gain was larger than the observed maximal gain to compute a p-value.

fn control(&self) -> &Control

Hyperparameters.

fn loss(&self, start: usize, stop: usize) -> f64

Loss of segment [start, stop).

fn gain_full(
    &self,
    start: usize,
    stop: usize,
    split_candidates: &[usize]
) -> FullGainResult

Gain when splitting segment [start, stop) at points in split_candidates.