Struct OnlineLearner 

pub struct OnlineLearner {
    pub algorithm: OnlineAlgorithm,
    pub weights: Vec<f64>,
    pub bias: f64,
    pub dims: usize,
    pub loss_fn: OlLossFunction,
    pub velocity: Vec<f64>,
    /* private fields */
}

Online / incremental learner supporting Perceptron, Passive-Aggressive, and SGD-with-Momentum algorithms.

The learner maintains a weight vector w ∈ ℝᵈ and a scalar bias, updated sample-by-sample via the selected OnlineAlgorithm.

Fields

algorithm: OnlineAlgorithm

The update algorithm in use.

weights: Vec<f64>

Current weight vector.

bias: f64

Scalar bias term.

dims: usize

Dimensionality (number of features).

loss_fn: OlLossFunction

Loss function for computing per-sample losses.

velocity: Vec<f64>

Velocity buffer for SGD-with-Momentum (zero for other algorithms).

Implementations

impl OnlineLearner

pub fn new( algorithm: OnlineAlgorithm, dims: usize, loss_fn: OlLossFunction, ) -> Self

Create a new OnlineLearner with zero-initialised weights.

Arguments

• algorithm — update rule to apply on each update() call.
• dims — feature dimensionality; all input vectors must have exactly dims elements.
• loss_fn — loss function used for reporting and SGD gradient computation.

Panics

Does not panic; returns a well-formed OnlineLearner even for dims == 0.

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

Compute the raw decision score: dot(weights, features) + bias.

Errors

Returns LearnerError::EmptyInput if features is empty when dims > 0, or LearnerError::DimensionMismatch if features.len() != dims.

pub fn predict_class(&mut self, features: &[f64]) -> Result<i32, LearnerError>

Return the predicted class (+1 or −1) for features.

The class is the sign of predict. A score of exactly zero is classified as +1.

This method also updates the internal prediction statistics.

Errors

Propagates errors from predict.

pub fn classify(&self, features: &[f64]) -> Result<i32, LearnerError>

A non-mutating variant of predict_class that does not update internal prediction statistics.

Useful for evaluation loops where you want to call accuracy() later without double-counting.

pub fn loss(&self, score: f64, label: f64) -> f64

Compute the loss for a given (score, label) pair using the learner’s configured OlLossFunction.

Loss	Formula
Hinge	max(0, 1 − label · score)
SquaredHinge	max(0, 1 − label · score)²
LogLoss	ln(1 + exp(−label · score)) (stable)

pub fn update(&mut self, sample: &TrainingSample) -> Result<f64, LearnerError>

Perform a single online update for sample and return the pre-update loss.

Errors

• LearnerError::EmptyInput — sample.features is empty but dims > 0.
• LearnerError::DimensionMismatch — feature length ≠ dims.
• LearnerError::InvalidLabel — label is not ±1.0 for Perceptron or Passive-Aggressive (binary classifiers).

pub fn batch_update( &mut self, samples: &[TrainingSample], ) -> Result<Vec<f64>, LearnerError>

Perform online updates for a batch of samples, returning the per-sample losses in the same order as samples.

Equivalent to calling update in sequence.

Errors

Returns the first error encountered, if any.

pub fn accuracy(&self, samples: &[TrainingSample]) -> Result<f64, LearnerError>

Compute the fraction of samples correctly classified without updating weights.

Classification is performed via classify so the internal total_predictions counter is not incremented.

Errors

• LearnerError::EmptyInput — samples is empty.
• Propagates dimension/label errors from classify.

pub fn reset(&mut self)

Reset weights, bias, velocity, and all accumulated statistics to zero.

pub fn l2_norm(&self) -> f64

Compute the L2 norm of the weight vector: √(Σ wᵢ²).

pub fn stats(&self) -> OnlineLearnerStats

Snapshot current training statistics.

pub fn average_loss( &self, samples: &[TrainingSample], ) -> Result<f64, LearnerError>

Compute the average loss over a slice of samples without updating weights.

Errors

Returns LearnerError::EmptyInput if samples is empty, or propagates dimension errors.

pub fn evaluate_losses( &self, samples: &[TrainingSample], ) -> Result<Vec<f64>, LearnerError>

Compute per-sample losses over samples without updating weights.

Errors

Returns LearnerError::EmptyInput if samples is empty.

pub fn record_prediction(&mut self, was_correct: bool)

Record a correct/incorrect prediction result into the running stats.

This is used internally when predict_class is called. Exposed publicly for external evaluation loops that use classify() and wish to manually feed outcomes back.

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

Return a reference to the current weight vector.

pub fn bias(&self) -> f64

Return the current bias value.

pub fn dims(&self) -> usize

Return the number of features this learner was constructed for.

Trait Implementations

impl Clone for OnlineLearner

fn clone(&self) -> OnlineLearner

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for OnlineLearner

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

Formats the value using the given formatter.