Regression Operator

The Regression Operator is a low-code operator for supervised tabular regression. It trains a model from a training dataset, optionally evaluates on labeled held-out test data, scores a separate unlabeled batch dataset, and writes a consistent set of artifacts such as predictions, metrics, an HTML report, and a serialized model bundle. Batch prediction does not require model registration or deployment.

Overview

Required inputs

The current implementation requires:

  • training_data

  • target_column

All columns in training_data except target_column are treated as features.

Optional inputs

The operator also supports:

  • test_data for held-out evaluation

  • prediction_data and prediction_output for unlabeled batch scoring; all prediction input columns pass through by default

  • output_directory for artifact location

  • column_types to override automatic type inference

  • model_kwargs to control explicit model runs

  • save_and_deploy_to_md for Model Catalog registration and Model Deployment creation

Supported models

The supported model values are:

  • auto

  • linear_regression

  • random_forest

  • knn

  • xgboost

auto performs cross-validation across the explicit model families and selects the best one for the configured metric. Explicit models use Optuna-based tuning by default.

Preprocessing

By default, the operator:

  • infers numeric, categorical, and date columns

  • imputes missing numeric values with the median

  • imputes missing categorical values with the mode

  • one-hot encodes categorical columns

  • expands date columns into year, month, day, dayofweek, and dayofyear

Artifacts

Depending on the configuration and available data, the operator can write:

  • training_predictions.csv

  • test_predictions.csv

  • predictions.csv

  • training_metrics.csv

  • test_metrics.csv

  • global_explanations.csv

  • local_explanations.csv

  • report.html

  • model.pkl

  • deployment_info.json

global_explanations.csv and local_explanations.csv are written when generate_explanations: true and explainability output is successfully produced. Global explanations summarize importance across the model, while local explanations contain per-row feature contributions.