Struct FtrlClassifier 

pub struct FtrlClassifier { /* private fields */ }

FTRL binary classifier with log loss.

Predicts P(y=1 | x) = sigmoid(w · x + b). The gradient of the log loss w.r.t. the logit simplifies to probability - target, so each feature’s gradient is (probability - target) * x_i.

The returned probability lies in [0, 1]. Extreme logits can produce exactly 0.0 or 1.0 after sigmoid; downstream consumers such as crate::loss::log_loss::BinaryLogLoss clip internally.

Examples

use rill_ml::models::{FtrlClassifier, FtrlConfig};
use rill_ml::sparse::SparseFeatures;
use rill_ml::SparseClassifier;

let mut model = FtrlClassifier::new(FtrlConfig::default()).unwrap();
let sf = SparseFeatures::from_sorted(vec![(0, 1.0)]).unwrap();
let _proba = model.predict_proba(&sf).unwrap();
model.learn(&sf, true).unwrap();

Implementations

impl FtrlClassifier

pub fn new(config: FtrlConfig) -> Result<Self, RillError>

Create a new FTRL classifier.

Returns an error if the configuration is invalid.

pub const fn config(&self) -> &FtrlConfig

The model configuration.

pub fn weights(&self) -> Vec<(FeatureId, f64)>

Return the current non-zero feature weights, sorted by FeatureId.

Features whose FTRL weight is exactly zero (due to L1 soft-thresholding or never having been updated) are excluded.

pub fn intercept(&self) -> f64

Compute the current intercept (bias) weight.

pub fn feature_count(&self) -> usize

Number of distinct features the model has seen.

pub fn resource_diagnostics(&self) -> FtrlResourceDiagnostics

Report feature-capacity and new-feature admission diagnostics.

Trait Implementations

impl Clone for FtrlClassifier

fn clone(&self) -> FtrlClassifier

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for FtrlClassifier

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl SparseClassifier for FtrlClassifier

fn samples_seen(&self) -> u64

How many training samples the model has seen so far.

fn predict_proba(&self, features: &SparseFeatures) -> Result<f64, RillError>

Predict the probability of the positive class.

fn learn( &mut self, features: &SparseFeatures, target: bool, ) -> Result<(), RillError>

Update the model using a single labeled sparse sample.

fn reset(&mut self)

Reset the model to its initial state.

fn predict(&self, features: &SparseFeatures) -> Result<bool, RillError>

Predict the boolean class label using a 0.5 threshold.