Guide¶

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

1. Describe your data. Build a database. Register each table. Name its primary key column and the column that records when a row becomes knowable. Then link the tables with relationships. tusk takes these declarations on trust. Call validate() if you want to check them against the data.
2. Synthesize. Call deep_feature_synthesis(). It walks the relationship graph, stacks primitives up to max_depth, and returns the feature matrix and the feature definitions used to compute 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.

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