5 Ways Facebook Uses Machine Learning

Above: Andrew Blumenfeld, co-founder of Telepath, and Lily Adelstein, creative project manager, discuss Facebook and 5 ways the company uses ML.

On Monday morning, many logged into their computers, grabbed their coffee, typed Facebook (FB) into their browser, and were met with an error message. Normally 1.9 billion users are active on FB daily but even if you weren’t attempting to log in yesterday, it was hard to miss the impact of such a colossal player fumbling the ball. 

In order to make this kind of impact, FB collects troves of data and uses that data to create, for many, irresistible user experiences. This week FB is making headlines for this outage and a whistleblower scandal but in order to understand why FB even matters in the first place, it is helpful to remember that there is powerful technology, like machine learning, being used under the hood that helps the company maintain its premiere position. It is important to heed the mistakes of FB, but I doubt the future solution will be no FB all together. Instead, I imagine a more ethically minded newcomer challenging the company. But in order to challenge and learn from Goliath, it is important to understand what technology is at play. 

In a 2018 paper written by FB researchers, it states, “Machine learning sits at the core of many essential products and services at FB.” With 1.9 billion daily users, FB has access to tons of data. Having access to data is a prerequisite for machine learning, although you don’t need to have as much as FB does to get started. Machine learning is the process by which a machine learns (identifies relationships and improves with more access to information) from data and produces outputs based on those learnings. 

FB uses machine learning in various user facing ways including for news feeds, ads, search, face detection, language translation, and speech recognition among many others. Understanding and making this technology more accessible to smaller companies and entrepreneurs is critical to leveling the competitive playing field. Below I’ve highlighted a few ways FB uses machine learning. 

News feeds: FB uses machine learning to determine what you see on your news feed. FB researchers state that, “ranking algorithms help people see the stories that matter most to them first, every time they visit FB.” When understanding machine learning, it can be helpful to think about our own human intelligence. For example, if you read the paper everyday, you’ve probably been exposed to a lot of different stories. But your friend isn't going to be interested in all of them. So you take into account information you know about your friend. Maybe that she is visiting a new country soon, she is X age, X gender, interested in X things, friends with X people, you also have historical data on the types of stories she’s told you she’s liked in the past. Using all of this information, you can make a prediction for what article she might be interested in. Machine learning is similar. Some basic information FB collects might include user data that you’ve provided when you set up an account, and a history of articles you’ve clicked on before. Given this data, a learning model can start to make connections between your user information and the types of articles you like. With this information a model can predict with various levels of certainty what the next article is that you might like. 

Ads: FB uses machine learning algorithms to determine what ads to display to users. Machine learning can use data like user context, previous interactions and characteristics of an advertisement to determine what type of person is most likely to click on a certain ad. Then they choose to display that ad for that individual. 

Search: FB uses machine learning to predict the category of search you are looking for. For example, if you search cats, it uses machine learning to predict with a certain likelihood that you are looking for a photo, video, event or people. Given that there is so much multimedia content on FB, it is a much more efficient process to predict the likely category before searching all categories. Since FB has access to so much real time data, these models can also receive feedback to determine their accuracy. For example, if you searched for X, and were given Y results and you immediately re-searched, this could be valuable information for FB that the search result wasn’t satisfactory. 

Sigma: This is a framework used by FB to classify content and detect anomalies. Machine learning is used here because with so much content, it would be impossible to establish a series of logic and rules to categorize each piece of content. A machine learning model can be trained on content that is already classified, and then that model can help classify other content and potentially future content. This is helpful for anomaly detection and content moderation because a piece of content may fall into the category of violent or graphic and then be taken down if it violates platform policy. 

The last use case I’ll mention at FB is Facer which is their face detection framework. Face detection - a task that is easy for humans (even if you can never remember that person’s name) — has long been difficult for computers. Now machine learning models can train on vast amounts of photographs of people and learn to identify a human face. FB is then able to use machine learning to predict whether the identified face is one of your friends, in which case it would recommend you tag them. 

This is a brief summary of 5 ways FB uses machine learning but it only scratches the surface. Machine learning enables companies, like FB, to learn and make smarter decisions using vast amounts of data. Making this technology more accessible to smaller companies and entrepreneurs is critical to leveling the competitive playing field.