Skip to content

Latest commit

 

History

History
59 lines (51 loc) · 3.73 KB

feature-view.md

File metadata and controls

59 lines (51 loc) · 3.73 KB

FeatureView - A Table of Features

A FeatureView provides metadata to derive a table of feature values from other tables. FeatHub currently supports the following types of FeatureViews.

  • DerivedFeatureView derives features by applying the given transformations on an existing table. It supports per-row transformation, over window transformation and table join. It does not support sliding window transformation.
  • SlidingFeatureView derives features by applying the given transformations on an existing table. It supports per-row transformation and sliding window transformation. It does not support join or over window transformation.
  • OnDemandFeatureView derives features by joining online request with features from tables in online feature stores. It supports per-row transformation and join with tables in online stores. It does not support over window transformation or sliding window transformation.
  • SqlFeatureView derives features by evaluating a given SQL statement. Currently, its semantics depends on the processor used during deployment. We plan to make it processor-agnostic in the future to ensure consistent semantics regardless of processor choice.

FeatureView provides APIs to specify and access Features. Each Feature is defined by the following metadata:

  • name: a string that uniquely identifies this feature in the parent table.
  • dtype: the data type of this feature's values.
  • transform: A declarative definition of how to derive this feature's values.
  • keys: an optional list of strings, corresponding to the names of fields in the parent table necessary to interpret this feature's values. If it is specified, it is used as the join key when FeatHub joins this feature onto another table.

Transformation - Declarative Definition of Feature Computation

A Transformation defines how to derive a new feature from existing features. FeatHub currently supports the following types of Transformations.

  • ExpressionTransform derives feature values by applying FeatHub expression on one row of the parent table at a time. The FeatHub expression language is a declarative language, sharing a syntax and grammar reminiscent of the SQL SELECT clause. See here for a comprehensive list of built-in data types, functions and operators.
  • OverWindowTransform derives feature values by applying FeatHub expression and aggregation function on multiple rows of a table at a time.
  • SlidingWindowTransform derives feature values by applying FeatHub expression and aggregation function on multiple rows in a sliding window.
  • JoinTransform derives feature values by joining parent table with a feature from another table.
  • PythonUdfTransform derives feature values by applying a Python UDF on one row of the parent table at a time.