Module ftrl 

FTRL-Proximal online learning for sparse features.

Implements the Follow-The-Regularized-Leader Proximal algorithm, which is well-suited for high-dimensional sparse data. L1 regularization produces sparse weight vectors, and the per-coordinate learning rate adapts to feature frequency.

See: McMahan et al., “Ad Click Prediction: a View from the Trenches” (KDD 2013).

Per-coordinate learning rate

eta_i = alpha / (beta + sqrt(n_i))

Weight computation

For feature i:

if |z_i| <= lambda1:
    w_i = 0
else:
    w_i = -(z_i - sign(z_i) * lambda1) / (lambda2 + (beta + sqrt(n_i)) / alpha)

The intercept uses lambda1 = 0 (no L1 regularization).

Structs

FtrlClassifier

FTRL binary classifier with log loss.

FtrlConfig

Configuration for FTRL models.

FtrlParam

Per-feature FTRL state.

FtrlRegressor

FTRL regressor with squared loss.