How Facebook Uses Deep Learning to Predict What You Might Like

A method for training AI models to recommend new content by comparing a user's past interactions with unseen items in a social network.

What it covers

This patent describes a way to teach a computer model to better understand user preferences. It takes a list of things a user has already interacted with, like posts or pages, and turns them into mathematical lists called vectors. By temporarily hiding one of those items and using the rest to build a profile of the user, the system can test if it correctly predicts the hidden item. It then compares this user profile against items the user has never seen, updating the model's math so that relevant items move closer to the user in a virtual space. This helps the system learn to suggest content that is more likely to interest the user.

What it doesn't cover

• —Does not cover non-deep-learning recommendation methods like simple keyword matching.
• —Does not cover systems that do not use vector-based embedding spaces for entity comparison.
• —Does not cover methods that do not involve removing a target entity from the interaction set to perform the training feedback loop.

The clever bit

The system treats the user's own history as a training ground by hiding a known interaction to see if the model can 're-discover' it, effectively creating a self-supervised feedback loop to tune the model's accuracy.

Why it matters

This technology is foundational to the modern social media experience. By refining how AI models learn from user behavior, platforms can keep users engaged by showing them content they are statistically likely to enjoy. This specific approach of using 'negative sampling' or 'target entity removal' is a standard technique in modern recommendation engines.

Real-world examples

• 1.Facebook News Feed content ranking
• 2.Instagram Explore tab suggestions
• 3.Targeted advertising systems based on user interest graphs