Latu.ML.Evaluator (latu_ml v0.2.0)

A metric, as inert data: what to measure and on which columns.

The third kind of operator beside Latu.ML.Estimator and Latu.ML.Transformer, and the simplest — it is never fitted and never transforms, it just scores a frame. Spark has six of them, all under Latu.ML.Evaluation.

evaluator = Latu.ML.Evaluation.binary_classification_evaluator(metric_name: "areaUnderROC")

Latu.ML.evaluate/2 is what runs one, and it answers with the metric itself rather than with a handle: an evaluator caches nothing on the server. The frame it scores is one a model has already transformed — an evaluator reads prediction_col and label_col, not features.

Summary

Types

t()

Functions

new(class, opts \\ [])

An evaluator the registry does not know, by its JVM class name.

Types

t()

@type t() :: %Latu.ML.Evaluator{
  class: String.t(),
  known: [Latu.ML.Param.t()],
  params: [Latu.ML.Plan.param()],
  uid: String.t()
}

Functions

new(class, opts \\ [])

@spec new(String.t(), keyword()) :: t()

An evaluator the registry does not know, by its JVM class name.

See Latu.ML.Estimator.new/2; the same caveats apply.

Options

:params — the params to send, already in wire form. Defaults to [].
:uid — the uid to send. Defaults to a generated one in PySpark's shape.