Non-Engagement Features
Detailing the non-engagement features that Promoted supports, and how to send them to Promoted
Promoted supports many different Feature types to make recommendations. There's rarely a concern over sending Promoted too much data — Promoted's systems do all of the processing and filtering. Here, we list the major Feature Domains, provide some examples, and describe how to send these features to Promoted. The top domains are:
Content Item features
Content item features are stable properties of the item shown. These can be attributes like color, a list of labeling tags, media like images, or free text like a title or description. Send Content Item features updates asynchronously to Promoted asap using the Content API. Promoted will process these quickly and have the data available in the ranking system. Content updates can also be sent to Promoted over Delivery API. However, larger RPC sizes will hurt latency.
User features
User features are stable properties of the user viewing the content. The user may be anonymous, in which case there are no or limited user features. User features can be user attributes like age and gender, lists of interests or behaviors that may be computationally inferred, or user-provided settings. Send User features similarly to Content Item Features but using the User method on the Content API.
Context features
Context features are properties about where the items will be shown and to whom. For example, a search query, the page address, the mobile device type, geographic location, or time. Send Context features on the Request to Delivery API.
Interaction features
Interaction features require the intersection between two or more domains. Interaction features tend to be the most powerful features in scoring (item, user, context) matches. Domain intersections include:
- (Item x Context): query match or relevance scores
- (Item x User): personalization scores, past user history with this item, collaborative filtering
- (Context x User): search filters, home feed categories
- (User x Item x Context): other recommendation models
Avoid sending features back on View, Impression and Action log records
It's a common mistake to try to send features back on these records. The Promoted Delivery system either needs to (1) already have the features pre-processed or (2) have them passed our the Delivery API call. For questions, contact the Promoted team.
Updated about 1 month ago