Skip to main content

Featurization Groups

VehicleFeaturizationComponent writes, to featurization/inventory and featurization/synthetic, a set of each vehicle's features for each top-level category of the feature hierarchy. The feature hierarchy alone defines these categories, so no OEM configures them.

Attribute encoding​

Every featurization table stores an attribute as an attribute_type, an attribute_value and an attribute_class. The type is the scheme id in upper case.

  • For a multiclass scheme, the value is the option selected and the class is MULTICLASS, such as UPHOLSTERY_MATERIAL / FINE_LEATHER / MULTICLASS.
  • For a binary scheme, the value is TRUE and the class is BINARY, such as DASHCAM / TRUE / BINARY.

featurization/features and featurization/models hold a list of these attributes per row. A binary scheme that answered false adds no attribute to the list. featurization/inventory, and featurization/synthetic where an OEM has synthetic candidates, have one row per vehicle and attribute.

By default, a vehicle has the attributes of its configurable features only. A configurable feature is one whose row in consolidated/features has is_configurable true. Setting configurable_only to false on VehicleFeaturizationComponent gives a vehicle the attributes of every feature.

Every vehicle of one model_id has the same standard equipment. An inventory record can list fewer features than the vehicle has. With configurable_only false, such a record gives the vehicle FALSE or NO_<type> for the type of a standard feature that the record leaves out. For example, an Audi RS 3 record without its four-way power lumbar (7P1) gives the vehicle NO_POWER_LUMBAR.

Candidate selection raises on a repeated row, so unique_attribute_rows is a blocking check.

unique_attribute_rows computes a 64-bit hash of each row's (inventory_id, attribute_type, attribute_value) key in the Polars streaming engine, sorts the hashes, and compares each hash with the previous one. Its memory grows with the partition's row count, but it holds no hash table of keys. A second read compares each repeated hash against the keys of its rows, so two keys whose hashes collide do not fail the check.

When a vehicle has several values of one type, such as two wheel diameters, featurization/inventory sorts the values and joins them with a comma into one value, such as 20,21. For each type that some vehicle in the partition has, a vehicle with no value of that type gets FALSE for a BINARY type and NO_<type> for any other. The take rates of each type, binary types included, therefore sum to 1. When synthetic_inventory is set, a vehicle also gets FALSE or NO_<type> for each type that a row of featurization/models or a featurized feature has (see Synthetic candidates). With configurable_only true, the default, only the configurable ones among the featurized features add their types.

Every attribute that featurization/inventory builds itself is MULTICLASS. These attributes are the sets and the crosses.

Groups​

The groups are:

  • each top-level category of the feature hierarchy, such as ADAS, which has a set and a group price;
  • each value of a multiclass scheme at the root of the attribute hierarchy, such as TECHNOLOGY or DESIGN, which is a facet group and has a group price only.

The attribute hierarchy, under schemes/attributes, classifies the attributes of the feature hierarchy rather than features. Its root holds the facet scheme, which asks whether an attribute is mainly technology or mainly design. Each materialization of featurization/inventory_measures and featurization/synthetic_measures uses the attribute hierarchy to classify the definition of each attribute that the feature hierarchy can assign:

  • a binary attribute's definition is its scheme's question, focus and true answer;
  • a multiclass value's definition is its scheme's question and the value's name, description and examples.

An attribute is in the group of each facet value that the attribute hierarchy assigns to its definition. A facet is therefore a property of the attribute. For example, adaptive cruise control is in TECHNOLOGY on every feature that has it, including a package. An attribute assigned NEITHER is in no facet group.

These classifications go through the answer cache, as a feature's classification does. An attribute is therefore classified again only when its definition or the attribute hierarchy changes. No table stores an attribute's facet.

Only configurable features count in a group, whatever the value of configurable_only. Separately, FeatureFeaturizationComponent does not classify a feature whose name is missing or matches the component's code_name_pattern in full, when its description is missing or only repeats the name. Audi's code_name_pattern matches a name that is only a code, such as TWL:1VH. A feature that is not classified has no attributes, so it belongs to no group.

One option code can have several rows on one model: the row of the standalone option, and a row for each package that contains it. Each distinct text is classified separately, so when the rows name the option differently they can get different attributes. A vehicle's attributes would then depend on whether the vehicle has the option standalone or in a package. The consolidated tier is where an OEM gives the rows one text; for Stellantis, see Taxonomy backfill.

The warning check one_attribute_set_per_option reports each option code on a model whose classified configurable rows got more than one set of attributes. Its metadata also has the share of classified configurable features that got no attributes.

A set is limited to top-level categories because a set is one categorical attribute. A set over every subcategory, or over a facet that spans every category, would have so many distinct values that each would cover few vehicles.

Sets​

<GROUP>_SET holds the binary attributes under the category and NAME=VALUE for each multiclass scheme under it, sorted and joined with commas. A vehicle with none of them gets an empty set.

featurization/inventory holds both the sets and the binary attributes for every value of a market's binary_representation. binary_representation chooses which of the two the take-rates projection and the attribute weights read.

A multiclass scheme can set exclude_from_set. The feature hierarchy excludes its color and finish schemes. It also excludes axle_ratio, because a vehicle's axle ratio varies with its engine and its towing equipment, and gvwr_class for the same reason. It excludes center_screen_size, because the screen size follows the trim and its packages, which the sets already distinguish. On Audi's 2026-10-11 partition, adding upholstery_color to INTERIOR_SET raised the median number of distinct sets per base model from 6.5 to 12.5. Adding upholstery_material and interior_trim_category changed it little, so those schemes stay in the set.

Group prices​

A group's price is the summed price of the vehicle's configurable features in the group, rounded to the cent. A feature with no price counts as $0. The prices are measures; no attribute holds them. A scheme that sets exclude_from_set still counts toward its category's price.

Packages​

A package often holds features of several groups. Its price is therefore split, and is not counted whole in each group that one of the package's own attributes belongs to.

A package's contents on a model are the option codes of the model's consolidated features whose parent_package_code is the package. On a vehicle that has the package, a feature with one of those codes is a content feature of the package. This holds even when the vehicle's feature_ids name the code's standalone row, the consolidated row with no parent_package_code.

The package's price is split equally among:

  • each content feature of the package that the vehicle has and that has an attribute in some group;
  • each attribute of the package's own feature that is not in the PACKAGE category and that none of those content features has.

The second kind covers a content that the package's description names but that has no consolidated row whose parent_package_code is the package, such as hands-free highway driving in Audi's Q7 Premium Plus.

A content feature's share counts once in each of its groups, its facet groups included. An attribute's share counts in each of that attribute's groups, its facet group included. A content feature's own price is not counted. A content the vehicle does not have gets no share. For example, a package can offer two wheels, and a vehicle has only one of them.

The package's own feature adds its whole price to PACKAGE through its attributes in the PACKAGE category, such as SPORT_PACKAGE. It adds nothing to any other group.

When a package has no content feature or attribute to split its price among, its own feature adds the whole price to each group that one of its attributes belongs to.

Few content features have a price of their own in the consolidated features table, so the split is equal and not weighted by price. On Audi's 2026-10-13 partition, 34 of the 2,381 contents with a standalone row had a price on that row.

All features price​

A vehicle's all_features_price is the summed price of its configurable features. VehicleFeaturizationComponent computes it from feature_ids and the consolidated features table only, and does not read the vehicle's msrp. An inventory vehicle and a synthetic candidate with the same features therefore get the same price. A vehicle's msrp comes from its inventory source. On most OEMs a synthetic candidate has no msrp, so a price read from msrp would differ between an inventory vehicle and a synthetic candidate with the same features.

A feature counts once per option code, and only when its consolidated row has no parent_package_code. A package counts at its own price. A content of a package the vehicle has adds nothing, even when the vehicle also has the content's standalone row. A feature with no price adds $0. The price leaves out the model's base price and destination charge. Both are the same for every vehicle of a model.

Measures​

An attribute is a category that take rates, weights and crosses group vehicles by. A measure is a numeric value of one vehicle that the code reading the table computes with.

VehicleFeaturizationComponent writes the measures of the inventory vehicles to featurization/inventory_measures and those of the synthetic candidates to featurization/synthetic_measures. Each table has one row per vehicle, with inventory_id, model_id, dealer_id, is_inventory and these measures:

  • all_features_price, a Float64 (see All features price).
  • group_prices, a list of structs with the fields feature_group and price. The list has one entry for each group, sorted by feature_group, with the vehicle's group price. A group that none of the vehicle's features contributes to has a price of $0.

The blocking checks one_row_per_vehicle and finite_measures fail a partition with a repeated inventory_id, or with a null or non-finite all_features_price or group price. Candidate selection's PriceDistributionScorer reads them.

Synthetic candidates​

When synthetic_inventory names a synthetic candidates asset, VehicleFeaturizationComponent also writes each synthetic candidate's attributes to its table, with is_inventory false. A synthetic candidate gets the sets of its groups from its feature_ids.

VehicleFeaturizationComponent writes featurization/inventory, and when synthetic_inventory is set, also featurization/synthetic, each from an asset of its own:

  • featurization/inventory holds the inventory vehicles. Its asset does not read the synthetic candidates, so demand, which reads it, does not depend on them.
  • featurization/synthetic holds the synthetic candidates. Its asset does not read consolidated/inventory or featurization/inventory.

Each inventory vehicle is featurized and stored once. A reader of inventory vehicles alone, such as attribute_stats, attribute_weights and order_set_evaluations, reads featurization/inventory. A reader that needs every candidate concatenates featurization/inventory and featurization/synthetic. candidates_output and its candidate variety checks do so when candidates_output has synthetic_inventory set. Candidate selection does so through the extra_namespaces of its candidate_attributes table.

In featurization/inventory and featurization/synthetic, a vehicle gets FALSE or NO_<type> for each type that a row of featurization/models or a featurized feature has (see Attribute encoding), so a candidate and an inventory vehicle are filled with the same types. With configurable_only true, only the configurable ones among the featurized features add their types. Without that fill, a candidate would get a NO_<type> row for a type that only another synthetic candidate has, and the inventory vehicles would lack that row.

Crosses​

CROSSES in ai_dagster.components.featurization.vehicles lists the pairs of attribute types crossed into one attribute, such as PAINT_COLOR_X_UPHOLSTERY_COLOR with the value WHITE|BLACK. A side the vehicle has no value for takes NO_<type>. The pairs are chosen by hand. Each side of a pair is an attribute type of the vehicle's features or of the vehicle's model. A cross cannot read a <GROUP>_SET attribute.