Struct Ftrl 

pub struct Ftrl { /* private fields */ }

Follow the Regularized Leader — Proximal (McMahan et al. 2013).

FTRL-Proximal maintains per-feature adaptive learning rates and supports L1 + L2 regularization. L1 induces sparsity: features with |z_i| ≤ l1 are zeroed out, which makes FTRL popular for large-scale sparse models.

Update equations (per coordinate i):

g_i       = gradient of logistic loss on this sample
z_i      += g_i − (√(n_i + g_i²) − √n_i) / α · w_i
n_i      += g_i²
if |z_i| ≤ l1:
    w_i = 0
else:
    w_i = −(z_i − sign(z_i)·l1) / ((β + √n_i) / α + l2)

Implementations

impl Ftrl

pub fn new(n_features: usize) -> Self

Create a new FTRL learner with n_features dimensions.

Defaults: α = 0.1, β = 1.0, l1 = 0.0, l2 = 0.0.

pub fn with_alpha(self, alpha: f64) -> Self

Set the learning rate α.

pub fn with_l1_l2(self, l1: f64, l2: f64) -> Self

Set L1 and L2 regularization coefficients.

pub fn stats(&self) -> &OnlineStats

Reference to running statistics.

Trait Implementations

impl Clone for Ftrl

fn clone(&self) -> Ftrl

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Ftrl

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

Formats the value using the given formatter.

impl OnlineLearner for Ftrl

fn update( &mut self, features: &[f64], label: f64, ) -> Result<OnlineUpdateResult, OnlineError>

Update model on a single (features, label) pair. Returns update stats.

fn predict(&self, features: &[f64]) -> Result<f64, OnlineError>

Predict class label or regression value for features.

fn n_updates(&self) -> usize

Number of updates seen so far.

fn weights(&self) -> &[f64]

Current weight vector (without bias).