Introduction to Recommendation Systems: The Invisible Helper in Your Daily Life

If you’ve ever opened TikTok after a long day of school and found a compilation of orange tabby cats doing backflips waiting for you before you even type “cat videos” into the search bar, you’ve already encountered a recommendation system. For 16-year-old Lila Chen, a sophomore in Toronto who spends her free time collecting Pokémon cards, grinding Valorant ranked matches, and studying to lo-fi hip-hop playlists, that exact scenario played out last Tuesday. She’d just finished a 45-minute calculus homework set, flopped onto her bed, and tapped the TikTok app—only to be greeted first by a 15-second clip of a tabby launching itself off a couch, followed by a 10-second snippet of a new Phoebe Bridgers track she’d never heard before, and finally a Valorant clip of her main agent Jett pulling off a 1v5 clutch on the map Bind. She sat up so fast her calculus notes fell off the bed. “How does it know?” she muttered, before pulling up her Spotify app, where her weekly Release Radar playlist was already loaded with three new indie pop tracks that matched her exact taste. Later that afternoon, she searched Amazon for a new phone case, and the first “Customers who viewed this item also viewed” recommendation was the limited-edition holographic Charizard card she’d mentioned wanting for her birthday in a text to her best friend two days prior. When she logged into Roblox to play Adopt Me! with her friends, the game matched her with three players around her skill level who didn’t troll and actually wanted to trade pets fairly, instead of the random 6-year-olds who usually scammed her out of her rare pets.

Lila didn’t realize it at the time, but every one of those perfect matches was powered by a recommendation system: the invisible, behind-the-scenes software that curates content, products, and even people for you based on what it thinks you’ll like. You encounter these systems dozens of times a day, even if you never stop to think about them. When YouTube suggests the next video to watch after you finish a tutorial on how to draw anime characters, that’s a recommendation system. When Instagram shows you a post from a small artist you follow right after you like one of their sketches, that’s a recommendation system. When your school’s homework portal suggests practice problems for the math topic you struggled with on your last quiz, that’s a recommendation system too. For teens growing up in a world where almost every app and platform you use relies on these tools, understanding how they work isn’t just a cool party trick—it’s a way to take control of the content you see, avoid manipulative design, and even build your own recommendation systems for the things you care about.

Before we dive into how these systems work, let’s start with the basic building blocks that every recommendation system relies on, no matter how complex it is. These three core terms will come up over and over again in this book, so it’s worth memorizing them now: 1. Users: Anyone who uses a platform with a recommendation system. For Lila, she’s a user of TikTok, Spotify, Amazon, and Roblox. Users can be individuals, businesses, or even other systems, but for this book, we’ll focus on individual users, especially teens like you. 2. Items: Anything that can be recommended to a user. For TikTok, items are short-form videos. For Spotify, items are songs, albums, playlists, and podcasts. For Amazon, items are physical and digital products. For Roblox, items are other players to team up with, in-game cosmetics, and even game modes. Even your school’s cafeteria uses a recommendation system to suggest lunch options based on what you’ve bought before—those lunch options are items too. 3. Interactions: Any action a user takes with an item. Interactions come in two types: explicit and implicit. Explicit interactions are clear, intentional signals of what you like or dislike: rating a song 5 stars on Spotify, liking a TikTok video, writing a 1-star review for a phone case on Amazon, or reporting a player for trolling in Roblox. Implicit interactions are the small, often unconscious actions you take that the system tracks: how long you watch a TikTok video before scrolling past, whether you rewatch a clip three times, whether you pause a song halfway through, whether you add a product to your cart but don’t buy it, or whether you mute your mic during a Roblox match (which often signals you’re not enjoying your teammates). Explicit interactions are easy to interpret, but implicit interactions are far more abundant and often more honest: most people don’t take the time to leave explicit feedback, but their behavior tells the system everything it needs to know.

With these three building blocks, we can define the two most common tasks that recommendation systems perform:

• **Rating prediction**: The system guesses how much you’ll like an item, usually on a scale of 1 to 5 stars, or 1 to 10. For example, if Lila rates the new Taylor Swift album 5 stars, the system predicts she’ll rate Taylor’s new single 4.8 stars, so it places the single at the top of her Release Radar playlist.
• **Ranking**: The system sorts a list of items in order of how much it thinks you’ll like them. Your TikTok For You Page is a ranked list: the first video is the one the system thinks you’ll love the most, the second is the next most likely, and so on. Your Amazon search results are also a ranked list: the first product is the one the system thinks you’re most likely to buy, based on your past purchases and what other similar users have bought.

To understand how far recommendation systems have come, it helps to look back at their history. The first collaborative filtering algorithm, the foundation of most modern recommendation systems, was developed in 1992 by researchers at the University of Minnesota for a project called GroupLens, which recommended news articles to users based on what other users with similar reading tastes liked. It worked, but it was limited to a small group of researchers. The first commercial recommendation system launched in 1998, when Amazon added its “Customers who bought this also bought” feature to its product pages. The feature was an instant hit: Amazon’s 1998 annual report found that 35% of all sales on the platform came from recommendations, a number that has only grown in the years since.

The next big milestone came in 2006, when Netflix announced the Netflix Prize: a $1 million grand prize for any team that could improve the accuracy of its existing recommendation algorithm by 10%. Over 40,000 teams from 180 countries entered the competition, and after three years of work, a team called BellKor’s Pragmatic Chaos achieved the 10% improvement in 2009 using a combination of collaborative filtering and matrix factorization—techniques we’ll dive into in detail in Chapters 4 and 5. The competition spurred a wave of research in recommendation systems, and many of the algorithms used in today’s apps are direct descendants of the work done for the Netflix Prize.

The 2010s brought recommendation systems to the mainstream, as social media and mobile apps made personalized content accessible to billions of users. In 2015, Spotify launched Discover Weekly, a personalized playlist of 30 songs updated every Monday, based on your listening history and the listening history of users with similar tastes. By 2021, Discover Weekly had driven over 25 billion streams, and 40% of Spotify users reported discovering a new favorite artist through the playlist, according to Spotify’s 2021 Impact Report. Two years later, TikTok launched globally with its signature For You Page, a real-time personalized feed of short-form videos powered by an algorithm that tracks both explicit and implicit interactions to curate content in milliseconds. TikTok’s 2022 transparency report revealed that 90% of the content users see on the For You Page comes from the algorithm, not from accounts they follow—a statistic that hig