airmosaic-ai

Use the mlr3 framework

English 中文

This is framework guidance. mlr3_workflow_reference is catalogued but not executable through run-tool. For supported numeric regression, use the regression skill and ml_regression_workflow.

Prepare and configure

Use R with mlr3 and the selected learner backend. The synthetic examples require rpart; learned preprocessing uses mlr3pipelines, and tuning requires suitable mlr3tuning/paradox versions and a declared budget.

Write down the target and sample unit, predictors and types, prediction-time availability, split design, metric/direction and resource budget. For classification, specify class levels, positive class and probability/threshold decisions.

Check available APIs in the task environment:

stopifnot(requireNamespace("mlr3", quietly = TRUE))
print(packageVersion("mlr3"))
print(mlr3::mlr_learners$keys())
learner = mlr3::lrn("regr.rpart")
print(learner$packages)
print(learner$param_set$ids())

Use the code

The synthetic regression/classification examples provide complete R code for Task creation, three-fold development CV, fitting and new-data prediction. They use in-memory synthetic data and do not write files. They are direct R examples, not registered tools.

  1. Create the appropriate Task and learner; keep declared sample roles and split identity.
  2. Put learned imputation, encoding, scaling and selection inside the resampling graph.
  3. Compare a baseline and bounded candidates. Separate inner tuning from outer performance evaluation.
  4. Freeze choices, score the final holdout and use the fitted preprocessing/learner for new inputs.

Use operations and validation for graph, tuning and final-fit details. Rerunning CV after repeated selection does not create an independent test.

The automatic-tuning route adds a concrete direct R helper and synthetic script: supplied grouped folds, explicit search ranges, evaluation/fit budgets and final-test separation. Only base-R contract checks passed in the inspected default environment; the full tuning example requires the missing dependencies. Existing registered tools still use their original fixed-candidate contracts.

Check and save

Check failed folds, finite predictions, matching sample IDs and target units; for probabilities check class order and row sums. Save the fitted learner/graph, settings, split keys, predictions, metrics and dependency versions in a fresh external project result directory. Read back the model and compare predictions on the same prepared inputs.

General framework code does not inherit run-tool path or receipt handling. Use the asset contract for study outputs. Method selection explains when tabular learning is appropriate. Upstream tracking records the external reference version; it is not an automatic sync service.

This site is open source. Improve this page.