Guide¶

tusk turns a set of related tables into a single wide feature matrix. The workflow is always the same three steps:

1. Describe your data. Build a database: register each table, say which column is its primary key and which column records when a row became knowable, then link the tables with relationships. Those are declarations tusk takes on trust — call validate() when you want them checked against the data.
2. Synthesize. Call deep_feature_synthesis(). It walks the relationship graph, stacking primitives up to max_depth, and returns both the feature matrix and the feature definitions that produced it.
3. Re-apply. Feed those definitions back to apply_features() to compute the same columns on new data.

Everything tusk builds is a narwhals expression, so the whole pipeline is one query plan on the backend you already use. Nothing is materialized until you ask for it.

• Databases
• Running DFS
• Primitives
• Custom primitives
• scikit-learn pipelines
• Differences from featuretools