Facebook explains how the content shown in your News Feed is selected

Facebook spells out how News Feed ranking and the machine learning systems work to personalize content for all users.

Facebook has unveiled new details about how the News Feed ranking system works, how content/posts are filtered based on which signals, and how the feed is customized for each user.

The News Feed ranking system is based on multiple layers of machine learning models as opposed to one single algorithm, to predict the content most relevant for the user.

Two elements that this system determines is — what posts to show in the News Feed, and in what order to show it. These elements are based on factors such as what the user has followed, liked, or engaged with recently.

Now what a user is shown, or shown higher in the News Feed depends on a single-objective optimization or multiple objectives. Which post a user is most likely to interact with, their preferred format (like more interactive with videos than photos), factors of a post (who is tagged, etc), and more of such bits are considered by the prediction models to decide.

Multiple models would have several predictions for a user, and each model ranks a post based on their prediction, these predictions are then combined into one score to decide what post would be shown in which order. The platform also surveys users to grasp what they think about the posts shown to them.

Also Read: Key Takeaways Facebook Q4 2020 Report: Ad revenues continue to grow

A step-by-step breakdown of how the system works:

• The system collects posts since the users' last login that can be ranked in the Feed by friends, Groups, and Pages into an eligible inventory

• Unseen posts are also reconsidered by applying an unread bumping logic, posts that were previously ranked in the feed but not seen, and have triggered conversations among the user's friends are also added to the inventory

• Each post is then scored upon several factors, such as post type, similarity to other items, the tendency of the user to interact with the post, and more, are all calculated through multiple models on various machines called predictors

• Then the posts go through a series of passes, such as the integrity detection pass, next, the main scoring pass where the maximum personalization occurs and each post is calculated individually.

• The final stage is the contextual pass, where attributes like content type diversity rules are applied to make sure the user has a diverse News Feed.