DLRM
DLRM stands for Deep Learning Recommendation Model and refers to a model architecture published by Facebook (now Meta) that predicts which content or products a particular user will be interested in. Such recommendation models determine what appears at the top of feeds and shops, and they are among the most economically important AI systems there are.
When you open an app like Instagram, YouTube, or Amazon, your home screen looks different from your friends'. Behind this is a program that estimates, from your past behavior, what you are likely to click on. DLRM is a specific design for exactly such programs. The abbreviation stands for Deep Learning Recommendation Model, meaning roughly “recommendation model based on learning networks.” Facebook, now Meta, disclosed this design in 2019 and made the program code freely available. Since then, DLRM has been considered a standard example of how large recommendation systems are built.
Why recommendation models make the money
Publicly, people talk about chatbots and image generators. But the revenue of large platforms is made by recommendation models. Meta, TikTok, and Amazon earn their money by matching ads and products to the right person. Even one percent more clicks means billions of dollars for these companies.
That is why such models run around the clock and in enormous numbers. Meta has reported that recommendation models accounted for the largest share of its own AI compute load. Every refresh of a feed triggers countless predictions. A language model only responds when someone asks. A recommendation model computes as soon as someone scrolls.
For the public, this is also politically relevant. What such a model shows you shapes which news and opinions you see. Debates about filter bubbles or addictive feeds are essentially about these systems. DLRM is an example through which this technology can be publicly examined.
Numbers, categories, and the art of crossing
A DLRM receives two very different kinds of input. One kind is numbers, such as how long you were in the app yesterday. The other kind is categories with no natural order, such as the specific video or your device brand. A network can process numbers directly, but not categories.
That’s why each category receives what is called an embedding: a list of numbers that the model determines itself during training. Two similar videos thereby receive similar number lists. These tables are huge, because a platform knows millions of videos and users. As a result, a single DLRM can take up several hundred gigabytes, even though the actual computation stays small. This is the big difference from language models, where the computational work dominates.
At the core of the model, these number lists are then combined with each other in pairs. Experts call this feature interaction, meaning the crossing of features. Only through this does the model recognize patterns like: users of this age group with this device like this genre of music. In the end, a small network outputs a probability, usually for a click.
Where DLRM appears in products and headlines
You never see a DLRM directly. You notice it in the order of your feed, in the ads between posts, and in suggestions like “You might also like this.” Music services and online shops also work with very similar models, though not always with exactly this architecture.
In business news, you encounter the term indirectly. DLRM is part of MLPerf, a widely used benchmark for AI hardware. When Nvidia, AMD, or Google introduce new chips, a DLRM best time is often mentioned. For investors, this is an indicator of how well a piece of hardware handles enormous lookup tables.
A common misconception is that DLRM is a finished product from Meta. In fact, it is a published blueprint along with sample code. The systems actually used in production by these companies are further-developed, secret variants of it. DLRM primarily serves as a shared reference point for research and hardware testing.
