Controllable distribution will work as an added layer on the traditional click-through prediction model of the algorithm for Home Feed ranking.
Controllable distribution in the algorithm has been introduced to achieve business objectives on Pinterest since the traditional model was only designed to increase user engagement.
The special case solutions previously used in the codebase, such as inserting content, or moving the content up on the feed to manage business objectives caused practical and theoretical problems.
With the launch of the controllable distribution system, business owners can specify a global target for the percentage of impressions by content type.
For instance, if a percentage of the feed is set for the video, then controllable distribution calculates how to achieve that objective.
This layer is added after the traditional ranking layer, which helps the feed maintain the users’ content preferences and still achieve the objectives. The new system also adjusts in real-time, which helps the feed stay relevant and fresh.
To understand how this system works in detail, the engineering involved, and the equations that result in the boost or demotion of posts, read this Medium blog by Yaron Greif, Software Engineer, Home Feed Ranking.
Commenting on the impact of the new system, Pinterest mentions that it was able to achieve the original use case – allowing models to be deployed without costly delays as they were updated to achieve business constraints.
It is also being used to test the impact of different loads of various content types on the system.