Component availability

Use this catalogue to find an existing component and its configuration parameters. The linked API categories describe the shipped implementations. Registration makes a method discoverable; its suitable check still determines whether it can run on the task and data at its position.

See Build and adapt a pipeline to select training components and Configure a complete study to select analytical components. Add a method missing from the catalogue with Extending IAML.

Availability at a glance

Capability	Status	Scope
Training and analytical families below	Implemented	Included in IAML; individual methods require compatible data and prediction outputs.
LightGBMClassifier and LightGBMRegressor	Backend dependent	Available when lightgbm imports successfully. The adapters are excluded by suitability checks otherwise.
WeibullAFT	Backend dependent	Uses a compatible scikit-survival WeibullAFT implementation or falls back to lifelines.WeibullAFTFitter. Requires one of these backends to be available.
Advanced domain types, a constraints DSL and partial replacement of complete subpipelines	Unsupported	Outside the current recipe API. See Supported scope and limitations for the supported boundaries.

iaml.steps exposes readable names for IAML’s training adapters, including those above. The historical Act... imports refer to the same classes.

Prepare data

API category	What it provides
Initial data preparation	Standardize formats before modeling, such as trimming whitespace, parsing dates, converting numeric strings and recognizing missing-value codes.
Cleaning and encoding	Handle missing values and convert categorical, textual or date columns into model inputs. Includes imputation, category encoders and text representations.
Feature selection	Retain informative variables and reduce redundant features, using variance, correlation, statistical tests or model-based selection.
Scaling and normalization	Adjust numerical feature scales for models sensitive to units or magnitudes. Includes standard, robust and min-max scaling.
Class balancing	Change the balance of classes in training data through over- or undersampling, including SMOTE. Resampling is skipped during prediction.
Feature transformations	Explore alternative representations through dimensionality reduction, projections, feature aggregation and distribution transformations. Includes PCA and quantile transformations.

Predict outcomes

API category	What it provides
Classification	Predict categorical outcomes, such as disease status or a clinical event. Individual estimators determine which target types and probability outputs they support.
Regression	Predict numerical outcomes, such as a measurement or length of stay, using linear models, trees, boosting and other estimator families.
Survival analysis	Model time-to-event outcomes with censoring. Predictors include Cox models, survival forests and survival boosting.

Evaluate and optimize

API category	What it provides
Metrics	Score predictions for classification, regression or survival. The main metric defines the objective used to rank candidate pipelines.
Validation splitters	Define training and validation partitions for candidate evaluation, using cross-validation or holdout splits. Group information supports separation of patients or sites where required by the study.
Search optimizers	Propose further candidates from previous evaluations. Genetic search can change steps and their parameters, while Bayesian and random optimizers tune parameters within existing pipeline structures.

Inspect data and results

API category	What it provides
Descriptive statistics	Summarize the dataset through missingness, distributions, variability and associations. These summaries describe the data rather than prediction performance.
Plots	Visualize data, model performance and SHAP contributions. Individual plot classes specify the task, inputs and prediction outputs they need.
Explanation adapters	Request KernelSHAP through a configured explanation collection. Calculations run on demand on a fitted classification or regression pipeline.

See How it works for how these components work together.