Have you ever wondered how those random posts appear on your Explore page on Instagram? Well, they’re not random at all. The platform uses AI to find the stuff you might like. In a recent blog post, Instagram’s engineers explain more about how the app decides what gets to be in your Explore tab.
There are some useful applications of AI on Instagram, such as fighting comment trolls and online bullying or restricting harmful posts. But Instagram uses an algorithm to tailor your main feed, which probably isn’t anyone’s favorite feature of the app. I mean, just look at all those tricks for beating it. And other than the main feed, the app also relies on AI for tailoring the recommended content for you in the Explore tab.
In a rather technical blog post, Ivan Medvedev, Haotian Wu, and Taylor Gordon explain how the system works. It starts with a domain-specific language “optimized for retrieving candidates in recommender systems.”
The engineers note that it’s challenging to recommend content on Explore considering the vast variety of topics people are interested in. The topics vary “from Arabic calligraphy to model trains to slime,” the post reads. Therefore, the AI-system doesn’t rely on individual posts users would like to see, but rather on accounts they might find interesting.
Next, how does Instagram recognize accounts that are similar to each other? Well, the system uses so-called “word embedding.” It studies the order in which words most commonly appear to determine how related they are. It starts with billions of accounts. But after filtering spam, misinformation, or content that may violate the terms, it comes down to 500 eligible accounts which are then filtered further.
The system ranks the accounts based on how relevant they may be and how likely you are to interact with them. The first pass selects 150 out of 500 accounts. They come down to 50 in the second pass, picked by a lightweight neural network. Finally, a deep neural network cuts this number in half, leaving 25 accounts for the first page of your Explore grid.
As I mentioned, Instagram already uses AI for a variety of purposes, and it seems it will be developed further. There is certainly room for improvement, considering that some misinformation persistently remains on the platform (such as anti-vax posts). But as Medvedev told The Verge, Instagram plans to work on it. “We are also training AI models to proactively detect vaccine misinformation and take automatic action.” So, while AI isn’t a perfect solution, it can be improved to make the platform safer and offer more relevant, interesting, and informative content.