ML

Stage: Build · What it does: classic machine learning inside your pipelines — Train Model, Forecast, Cluster, and anomaly detection — trained, evaluated, and explained in the same flow as your data, no neural network required.

Why it matters

Wide model family. Random Forest, XGBoost, Logistic/Linear Regression, GBT, SVM, MLP — configure, train, and save from one step.
Honest evaluation. Cross-validation, hold-out splits, confusion matrix, ROC/AUC, and SHAP/LIME explanations before anything ships.
Reusable models. A saved model feeds Classify / ML Infer for batch scoring in any later pipeline.

How to use

Start from clean features: Profile, Impute, Encode, Scale, PCA, Auto Features, Feature Selection, and Lag Features.
Add Train Model (pick the family and target; enable saveModel), then an Evaluator (hold-out or cross-validation) and Explain (SHAP/LIME/permutation) for governance.
Use the trained model downstream with Classify or ML Infer; for time-series use Forecast, for segments use Cluster, and to flag outliers use Anomaly Detection.
Quick Report / Reporter publish results as HTML, JSON, or PDF; MetadataCollector keeps the governance trail.

CLI users: every ML step maps to a CLI command — the console is a visual front-end for the same jobs.

Real-world examples

🏦 Financial services — credit risk with an auditable trail

A bank trains an XGBoost model on application features, evaluates with cross-validation, runs Explain (SHAP), and logs everything through MetadataCollector — the model card and feature attributions go to risk review before the model scores live applications via ML Infer.

🛒 Retail — weekly demand forecast

A retailer aggregates daily sales, adds Lag Features, and runs Forecast (Prophet) for the next 30 days. The forecast lands in a catalog table the replenishment team reads every Monday.

🏭 Manufacturing — sensor anomaly alerts

An IIoT team runs Anomaly Detection (IQR) over hourly sensor telemetry, flags outliers, and exports flagged readings to the alerting bucket — no ML team required.

🏥 Healthcare — PHI-safe research features

A health system runs PHI Detect (redact) before Text Embeddings and Cluster, so topic models are built on de-identified notes.

Gotchas

Fit on train only. Impute/encode/scale statistics must come from the training slice — use PreProcessor for automatic train-only fitting and avoid leakage.
Models train on the configured compute and are metered — check the cost estimate before big runs.
saveModel must be enabled for a model to be reusable by Classify / ML Infer.

ML