Module online_learning 

Online learning algorithms: Perceptron, Passive-Aggressive, OGD, and FTRL.

All algorithms process one sample at a time with O(d) memory where d is the number of features, making them suitable for streaming and large-scale applications where the full dataset cannot be held in memory.

Algorithms

• Perceptron: Classic binary classifier (Rosenblatt 1958)
• PassiveAggressive: PA, PA-I, PA-II variants (Crammer et al. 2006)
• OnlineGradientDescent: OGD with squared/hinge/logistic losses
• Ftrl: Follow the Regularized Leader-Proximal (McMahan et al. 2013)

Structs

Ftrl

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

OnlineGradientDescent

Online Gradient Descent for convex losses.

OnlineStats

Cumulative statistics for an online learning session.

OnlineUpdateResult

Result of a single online update step.

PassiveAggressive

Passive-Aggressive classifier (Crammer et al. 2006).

Perceptron

Binary Perceptron classifier (Rosenblatt 1958).

Enums

OGDLoss

Loss function for OnlineGradientDescent.

OnlineError

Errors that can arise in online learning routines.

PAVariant

Selects the PA update variant.

Traits

OnlineLearner

Trait for online learners that update one sample at a time.

Functions

online_evaluate

Evaluate an online learner sequentially on a dataset.