Machine Learning Nodes
Choose a node below for its configuration, limits and complete examples. The
ML quick start walks through a held-out classification
flow. A standard Ciaren installation includes scikit-learn and MLflow; optional
XGBoost and LightGBM models need pip install "ciaren[ml]".
ML operations cross into pandas/scikit-learn even in Polars flows. Train nodes save fitted models to MLflow at runtime; exports fit pipelines in memory without logging or saving them. Model definitions stay unfitted for cross-validation.
What Evaluate produces
Evaluate turns scored rows into a metric / value table,
one row per computed scalar. Training-node metadata contains training scores;
use held-out predictions or cross-validation to estimate generalization.
Train / Test Split
Train/Test Split has separate train and test outputs.
Fit on train, then connect test and the trained model wire to Predict.
Feature engineering
| Node | Use |
|---|---|
| Scale Features | Scale selected numeric columns |
| Encode Categories | Encode categories into numeric features |
| Select Features | Select columns with a supported feature-selection method |
| Reduce Dimensions | Fit PCA on incoming rows and output components |
These standalone nodes fit on each incoming frame. To carry the same learned imputation/scaling/encoding from training to prediction, configure preprocessing inside a Train node instead of fitting separate transformers on the test branch.
Train nodes
| Node | Task and availability |
|---|---|
| Train Classifier | Fit a classification model and save its artifact |
| Train Regressor | Fit a regression model and save its artifact |
| Train Clustering | Fit K-Means, DBSCAN or Agglomerative Clustering |
| Train Dim. Reduction | Save a fitted PCA pipeline; does not output components |
| Train Forecaster | Scaffold; no built-in forecasting estimator |
Each implemented Train node emits one model reference containing the saved
artifact URI, not data rows. Predict requires an estimator with predict:
PCA, DBSCAN and Agglomerative Clustering do not support that consumer. Feature
Importance requires tree importance or linear coefficients.
Runtime validates row/feature limits before fitting and serialized model size after fitting, before saving. Preview does not fit or create an MLflow run. Load only trusted model artifacts; URI/path checks do not sandbox model code.
Model definition nodes
Classifier Model and Regressor Model emit unfitted definitions for Cross-Validate. They do not save final artifacts. Use a Train node afterward when you need a model for supported consumers.
Predict
Predict accepts rows and a trained model wire or explicit URI. It adds predictions and optional class probabilities, with configurable batches for prediction calls. Expected features are aligned when model/reference names are available; nameless external models need caller-selected columns.
Evaluate
Evaluate computes classification, regression or clustering metrics from scored data. Its reference explains probability requirements, undefined scores and metrics omitted from exports.
Feature Importance
Feature Importance ranks supported trained-model features. It does not train a model or establish causal importance.
Cross-Validate
Cross-Validate fits fresh classifier/regressor pipelines within folds using an unfitted definition. It returns fold scores and fit times, not a saved final model. Preprocessing is refitted inside each fold.