User and Item CMS Feature Formats

How to use the User and Item CMS systems to create automatic production features

The User and Item Content Management Systems (CMS) allow you to upload arbitrarily-formatted JSON per User and Content Item respectively for automated use in production machine learning, allocation rules, and in analytics.

Managing CMS Record State

The CMS system is a simple REST service. Typically, use BATCH or PUT to overwrite the JSON record value for any User or Item key. If you have multiple writers to the same record type, for example, a database update hook and a daily pipeline that both write to the same User ID in the User CMS, specify a different source parameter per writer. Each source is isolated. In Delivery to create features, the JSON records for all sources are unioned together. The overwrite order of the same JSON path in multiple sources is undefined. We recommend using unique keys across sources.

Example JSON CMS Records

We'll reference these example CMS JSON records in the documentation below.

Example Item CMS Record

{
  "valueScore": {
    "c": 1.8462813912476075,
    "a": 1.2222,
    "b": 2
  },
  "status_details": {
    "copyright": true,
    "author": "John Doe"
  },
  "is_enabled": false,
  "standards_label": [],
  "subscores": {},
  "categories": [
    "food",
    "drinks",
    "chicken"
  ],
  "created_at": "2022-11-10",
  "is_holiday": true,
  "seasonal_sale_pattern_usa": 0.1,
  "seasonal_sale_pattern_australia": 0.9
}

Example User CMS Record

{
  "country": "USA",
  "num_purchases": 34,
  "created_at": "2021-02-10",
  "categories": [
    "cats",
    "food",
    "chicken"
  ],
  "purchased": [
    "content_ID_1234",
    "content_ID_3482"
  ],
  "liked": [
    "content_ID_1234",
    "content_ID_3482"
  ],
  "followed_authors": [
    "John Doe",
    "Mary Smith",
    "Bill Gates"
  ]
}

How Features are Generated per Record

Numeric and Boolean Features

Keys paths with floats or integer values will be converted to numeric features. Boolean features have the value of 0 (false) or 1 (true).

Strings that can be cast to a numeric value will be converted and used like numeric values.

Categorical (string-valued) Features

Key paths with string values are converted to "Sparse ID" features are used in defining blender rules like diversity rules. For example, status_details.authorwill create a Sparse ID feature that could be used to define diversity allocation rules.

Identity Features

Integer-valued, Float-values with at most 2 digits of precision, and string-valued keypaths will create a "sparse" identity feature with a name with the pattern keypath=value and a value of 1. Identity features are used in sparse ML models. For example, valueScore.b=2= 1.0.

Maps

Each path in every map is a unique key path and generates at least one feature. For example, valueScore generates at least 3 features:

  1. valueScore.a=1.2222
  2. valueScore.b = 2.0
  3. valueScore.c=1.8462813912476075

Additionally, a feature indicating if the map is empty is set per map. For example:

subscores.empty = 1.0
valueScores.empty= 0.0

Lists

Two keypaths are generated for every item in a list: an ordered path, and an unordered path. The ordered path uses the integer position in the path, and the unordered path uses | for every position in the path. For example:

categories.0=food = 1.0
categories.1=drinks=1.0
categories.|=food=1.0
categories.|=drinks=1.0

There is a limit of at most 100 items in any list. If you need a longer list (e.g., to send a list of the most recently viewed pages), then you should be using the Measure system infrastructure to send this information to Promoted, which is dramatically more efficient and happens in real-time.

Timestamps

If a string looks like an ISO8601 timestamp string format or is a float or int within the reasonable range of an Epoch Unix timestamp in seconds or milliseconds, then Promoted will automatically represent this feature as a time-from-now feature with the same name. For example, if the value of an entry in the Item CMS is created_at="2022-11-10", then the value of the feature created_at will be the number of milliseconds at the time of Delivery when logged and when used in machine learning inference or blender rules. If 25 hours later, then the feature value will be 25 _ 60 _ 60 _ 1000 = 90,000. If 43.2 hours later, then 43.2 _ 60 _ 60 _ 1000 = 155,520.

Promoted will still create an "key=value"=1 identity feature for the exact value of the timestamp like for other features.

The numeric range is between 1999-12-31 16:00:00 as 946684800[_1000] and 2050-02-10 16:00:00 or 2528150400[_1000].

ISO8601 timestamps supported have the formats like:

  • YYYY-MM-DDTHH:MM:SSZ
  • YYYY-MM-DD

User <> Item Interaction Features

Interaction features are features that indicate a User and an Item share some relationship and are powerful for fitting high-quality personalization models. For example, you can define a feature for if a User follows the seller who sells a particular item.

Automatically-generated Interaction Features

When the exact same keypath appears in both the User and Item CMS, a third feature unique is created with the feature name pattern [keypath].interact with the value set as the maximum value of the two Item and User features. Interaction features will not be created if the resulting value is 0.

For example, in the examples above, the categories is shared and will generate interaction features.

  • categories.interact = 2.0 The value is the number of times items in these two keypaths have the same value. There are two matches in between categories sets in Item and User ("food" and "chicken").
  • categories.|=food.interact= 1.0 One of the matches in categories.
  • categories.|=chicken.interact= 1.0One of the matches in categories.

Colliding keypaths used in interaction features typically have a string value and include at least one list of strings. If you have two lists of strings (as in this example), then a set-intersection algorithm is used. If at least one of the keypaths is a string, then the maximum value of the interaction feature is 1.

It's possible to set a numeric-value interaction feature if the keypaths in the Item and User CMS records are equal. This is not a typical use case, and its use requires some clever use of the CMS record formats. Interaction features including lists of complex types (maps and nested lists) are not supported.

Interaction Features with Content ID and User ID

The keypath content_id is always available in Delivery for use in interaction features in the Item CMS. Its value is always set to the content_id in Delivery. Warning: if you set the content_id keypath in the Item CMS, then its value will be overwritten in Delivery. It's not forbidden to set the content_id keypath in the Item CMS, but it's probably a mistake that will lead to unexpected behavior.

Likewise, the keypath user_id is always available in Delivery for use in interaction features in the User CMS. The same collision for user_id warning applies for the User CMS as it does for the content_id in the Item CMS: it's not forbidden, but it's probably a mistake to set it.

Manually Mapping Interactions

You can define keypath aliases to generate interaction features. This is useful when it may be infeasible or cumbersome to use the exact same keypath between User and Item CMSs to create interaction features. Or, you may want a one-to-many or many-to-many mapping between keypaths. You can define any mapping using the Interaction Feature endpoint. To use this endpoint, please contact Promoted.

Example User<>Item Interaction Configuration

Note that "categories" is not defined because it's already automatically generated due to the same keypaths between the Item and User CMS.

{
  "user": {
    "status_details.author": "followed_authors",
    "liked.": "content_id"
  },
  "item": {
    "content_id": [
      "purchased",
      "purchased_qty"
    ]
  }
}

Context-Interaction Features (alpha)

Features set on the Request.Properties can also be used to generate interaction features with the Item or User CMS system. The functionality is the same as with the User and Item CMS. Contact Promoted if you want to enable this feature with your proposed use case.

One application for Context<>Item Interaction features are for modeling seasonal patterns. Promoted automatically and continuously updates, but sometimes learning can take time, and Promoted's training data may not extend to over one year or more.

Modeling Seasonality

Rather than using Context<>Item interaction features, consider some feature engineering on the Item CMS and regularly updating Item CMS records every month or, if necessary, every day. For example, consider the following features in the Item CMS example, which could be set by your daily analytics pipeline:

Example Item CMS properties to help model seasonality:

"is_holiday": true,
"seasonal_sale_pattern_usa": 0.1,
"seasonal_sale_pattern_south_africa": 0.9

"Seasonal sales pattern" type features can be set in a monthly pipeline that normalizes relative monthly sales volume for this category of product for the current month. For example, an item could be a summer-only item, or it could be only popular around a major holiday, like Christmas.

Avoid Collisions Unless Interaction is Intended

Promoted currently only supports User <> Item interaction features from the CMS. Features set on Insertion.Properties are assumed to overwrite or append to features from the Item CMS for the purpose of setting interaction features. This can be used in testing or for prototyping changes.

If you do not want automatically-generated interaction features generated between User and Item keypaths, then use unique keypaths in the User and Item CMSs and in all sources, including Request.Properties and Request.Insertion.Properties. Only one keypath will be used to set the corresponding feature (not including the interaction feature) when there are multiple synonymous keypaths from multiple sources. The order of priority for setting the value, from lowest to highest, is:

  1. Promoted built-in features
  2. Item CMS
  3. User CMS
  4. Request.Properties
  5. Request.Insertion.Properties
  6. [coming soon] A/B testing parameters

Example Problematic Collision

For example, both the User and Item records use the keypath created_at. This is probably an error because there isn't any obvious interaction relationship between Item and User for an attribute like this in this interaction feature system. An interaction feature created_at.interaction will still be created, but it's unlikely to be useful in machine learning. Further, you will lose the distinct ML features for Item.created_at and User.created_at. To correct this, rename one or both created_at feature name to something unique, like item_created_at.