Build Features

Description

The Build Features step engineers a table of derived feature columns from a source table — arithmetic and interaction combinations, date-part extraction, binning, scaling, and one-hot encoding — in one pass, without hand-writing an expression per column. It runs as an ordinary transform against the warehouse.

Configuration

• Source / Output tables — the input table and the table the derived columns are written to.
• Passthrough Columns (passthrough_columns) — source columns carried through unchanged.
• Features (features) — an ordered list of feature specs, at least one required. Each spec has a type and produces one or more output columns. Output column names — passthrough columns plus every non-one-hot feature’s name — must all be distinct; saving a duplicate name, or a duplicate one-hot (column, prefix) pair, is rejected.

Feature Types

Type	Fields	Produces
Expression (expression)	name, expression, dtype (default text)	One column: an arbitrary expression over source columns.
Interaction (interaction)	name, columns (exactly 2), op (multiply or add, default multiply), dtype (default numeric)	One column: the pairwise arithmetic combination of two numeric columns.
Date Part (datepart)	name, column, part (year, month, day, dow, or hour)	One column: the extracted date/time component.
Bin (bin)	name, column, thresholds (ascending, at least 1), labels (exactly one more than thresholds)	One column: the label for the bucket column’s value falls into. Thresholds are fixed values you supply — see Limits and Caveats.
Z-Score (zscore)	name, column, dtype (default numeric)	One column: the column standardized to zero mean, unit variance.
Min-Max (minmax)	name, column, out_min (default 0), out_max (default 1), dtype (default numeric)	One column: the column rescaled to the [out_min, out_max] range.
One-Hot (onehot)	column, prefix	One column per distinct value of column, each named <prefix><value>.

Limits and Caveats

• Binning needs fixed thresholds up front. bin buckets a column against thresholds and labels you supply — there is no data-driven binning (equal-width or equal-frequency bins computed from the data at run time). Compute thresholds yourself (for example from a prior aggregate step) before configuring a bin feature.
• One frame at a time. Build Features reads and writes a single source table; it does not aggregate values across related tables the way a cross-table feature-synthesis tool does. Model any cross-table relationship as a join upstream of the step instead.
• No built-in dimensionality reduction or target-aware encoding. There is no PCA-style component reduction and no target encoding (a categorical column re-expressed as a statistic of the label, such as its mean). Use onehot for categorical features, and a separate ML: Train Model step’s own feature handling if a training pipeline needs a different encoding.

Related

• ML: Train Model — train a model on a table, including one built by this step.
• Alteryx Conversion Matrix — how the Alteryx Build Features tool’s primitives map onto these feature types, and which of its primitives (cross-table aggregation, PCA, target encoding, auto-binning) have no equivalent here.