Module Modelkit.Sgd_classifier

Incremental linear classification trained by stochastic gradient descent.

Penalties, learning-rate schedules, stopping reasons, batch semantics, and checkpoint immutability follow Sgd_regressor; the shared variant types are re-exported so schedule values interoperate. start additionally registers the complete class set for the stream: labels are sorted into ascending order, must be distinct, and must number at least two. Every positively weighted row of a later batch must carry a registered class or partial_fit returns a typed validation error; zero-weight rows are never inspected because they cannot change the parameters. fit registers the positively weighted classes of its single matrix, matching the other built-in classifiers.

Two registered classes train one model whose positive class is the higher label; three or more train one one-versus-rest model per ascending class. All models share the update counter, learning rate, and per-batch permutation, and each row updates every model before the counter advances. Hinge uses the margin loss max 0 (1 - y * score) with y in -1, +1; Log_loss uses the stable binomial deviance with gradient sigmoid score - y for y in 0, 1. Sample weights scale the loss gradient without normalizing individual updates; intercepts are never penalized. The reported objective is the weighted mean over rows of the summed per-model loss plus the declared penalty over all coefficient rows.

Sgd_classifier.coefficients has one row for binary problems and one row per class otherwise, and Sgd_classifier.decision_function returns one column per model in the same order. Binary prediction selects the higher class when its score is strictly positive; multiclass prediction takes the first maximum score, so exact ties select the lowest label. Sgd_classifier.binary_decision_function exposes the single binary score column as a vector for pipeline dispatch and returns a typed compatibility error for multiclass models. Sgd_classifier.predict_proba is available only for Log_loss: binary probabilities are the sigmoid of the score and its complement, and multiclass probabilities normalize the one-versus-rest sigmoids, becoming uniform when every sigmoid underflows.

One batch costs O(n * m * p) time and O(n + m * p) temporary storage for n samples, m models, and p features. Checkpoints and fitted values store O(m * p) data and are in-memory training state, not a persistent artifact format.