How AI is recommending content on Facebook and Instagram

Notes on the how Meta's AI is recommending connected and unconnected content

In the movie “Don’t Look Up” the ruling government and the leading telecom provider are hand in glove. The operator is not only providing call service but also predicting how and when people will die, along with other things. For instance, the company has been monitoring you from the day you were born and is making decisions for you, making purchases actually before you could use your brain. Fascinating movie that narrates the biggest tragedy of mankind laced with humor.

The latest season of Black Mirror streaming on Netflix opens with a story where a streaming giant has mirrored the daily life of the protagonist and not only that but also all her secrets. She is not the only one, everyone subscribed to the streaming network is part of the ploy and all others are having fun and talking about it.

In both stories, AI plays a significant role as it does in the world’s largest social networking platform.

In a recent blog post Meta’s President of Global Affairs, Nick Clegg shares how AI systems are ranking content for Feed, Reels, Stories, and other surfaces. The article titled “How AI Influences What you see on Facebook and Instagram” furthermore talks about how one can control the content on these platforms.

Meta shares that the basis of prediction begins with what you share. For example, sharing a post is often an indicator that you found that post to be interesting, so predicting that you will share a post is one factor systems take into account. 

However, many other factors play a role and to give an understanding of all Meta recently released 22 system cards for Facebook and Instagram. They cover Feed, Stories, Reels, and other surfaces where people go to find content from the accounts or people they follow. You can click on the 22 system cards link, go to each product and get a significant amount of details on how recommendations are happening via AI and ways to control it too.

To understand a little more I clicked on one such product.

Facebook Video AI system

According to Meta when you view and interact with Facebook Video, one of the underlying AI systems delivers a range of video types that may match your preferences. These videos can be found in the Video tab, and they might include reels, music, gaming, shows, and other videos. This is content that you may be interested in from creators that you may not follow.

To predict the right video on the platform AI on Meta uses two recommenders. Recommender one displays the videos that appear in your Facebook Video feed, and recommender two determines the order of videos that appear after the one that you've clicked to view. 

Recommender one finds a list of questions and one of them is “How likely you are to watch a video for more than 30 seconds.” The signals that affect the predictions are:

• The length of the video

• The length of the first video you viewed in this session

• The total number of views the author's videos have had

• The total amount of time people have spent watching the video on their Facebook Feed

Once you have watched the video Recommender two will decide the order of the next videos based on a set of questions and one of them being “How likely you are to watch a video more than 60 seconds.” The signals that affect the predictions are:

• The length of the video

• How many times people have seen the video in their Facebook Feed

• How many times people have watched more than 75% of the video in their Facebook Feed

• The length of the first video you viewed in this session

Unconnected content recommendations

Subsequently, Meta has also shared how AI is recommending unconnected content on Facebook and Instagram.

Mark Zuckerberg noted on Meta’s most recent earnings call, more than 20 percent of content in a person’s Facebook and Instagram feeds is now recommended by AI from people, groups, or accounts they don’t follow.

Meta thinks that showing such unconnected content recommendations enhances the platform experience. To understand interests and also support the creator community Meta at a high level is doing these things:

Content understanding: Meta with its bunch of products is understanding the semantic meanings of content holistically across different modalities (such as image, text, audio, or videos). This then enables the systems to do more application-specific tasks, such as topic/genre classification, hashtag prediction, similarity matching, and clustering. 

Retrieval and ranking: Retrieval systems then narrow down the content pieces to a particular person’s interests. Finally, the ranking systems then select the final items based on pointwise and listwise predictions. They also adjust recommendations to deliver a balanced, engaging mix, so that people can enjoy content on a variety of interests, and a mix of popular and niche posts can appear in people’s feeds.

Personalizing the recommendations

Meta is not ignoring people’s feedback to improve the recommendations.

“After a recommendation is delivered, our AI systems respond to feedback and refine how they model each person’s preferences — if a person watched an entire video or liked a post, for example.”

Users can also personalise their feeds by controlling the recommendations. On Instagram, you can click on the “Not interested” feature if you are not happy with the content recommendation. The “Show more, Show less” feature on Facebook, is available on all posts in Feed, Video, and Reels via the three-dot menu.

Meta is even expanding its “Why Am I Seeing This?” feature in the Instagram Reels tab and Explore, and Facebook Reels in the coming weeks, after previously launching it for some Feed content and all ads on both Facebook and Instagram. “You’ll be able to click on an individual reel to see more information about how your previous activity may have informed the machine learning models that shape and deliver the reels you see.”

With regulators expressing growing concerns about AI and how the data is being used. These are Meta’s early steps to keep its sheet clean.

In the first episode of Black Mirror, Salma Hayek and the protagonist of “Joan is Awful” end the mayhem by entering the head office and destroying the AI system with a big hammer.

In reality, the AI recommendation engine will not be located in the middle of the city and one wouldn’t have Salma to fool the security.

Happy Sunday :)

P.S. I want to take a moment and thank you for reading my newsletter or blog post. It means a lot to me and pushes me to think and eventually write like I have done today. May peace be with you :)

By the way, the latest season of Black Mirror is worth a watch. Out of the six stories my favorites are - Beyond the Sea, Demon 79 and Joan is Awful.