Recommendation Systems
You finish a video and the site instantly suggests five more. You buy headphones online and an hour later see ads for headphone cases. You rate a book three stars and the app immediately updates its entire list of suggestions for you. Behind each of these moments is a recommendation system — an AI designed to predict what you want next.
What a Recommendation System Is
A recommendation system is a type of AI that predicts which items from a large catalog a particular user is most likely to enjoy or engage with. The catalog might be movies, songs, products, videos, articles, or social posts. The user is you — identified by your past behavior, your profile data, or both. Recommendation systems are one of the most commercially important AI applications ever built. Netflix has estimated that its recommendation engine saves the company over a billion dollars a year by keeping subscribers watching rather than canceling. Spotify, Amazon, YouTube, and TikTok are all built around the premise that a sufficiently accurate recommendation engine will keep users coming back indefinitely.
A recommendation system is an AI that predicts which items a specific user is most likely to enjoy, based on patterns in that user's history and the behavior of similar users. It matches people to content at scale.
Collaborative Filtering: The Power of Similarity
The most widely used technique in recommendation systems is called collaborative filtering. The core idea is elegantly simple: people who agreed with you in the past will probably agree with you again. Imagine a giant table. Rows are users. Columns are items — movies, songs, products. Each cell contains a rating or a signal of engagement. Most cells are empty because no one has seen everything. Collaborative filtering fills in the missing cells by finding users with similar patterns to you and using their ratings to predict your preferences. If User A and User B both rated the same ten movies similarly, and User B loved a movie User A has not seen, the system predicts User A will love it too. This happens across millions of users simultaneously, finding subtle similarities that no human analyst would ever spot.
Collaborative filtering predicts what you will like by finding users whose tastes resemble yours, then recommending items those similar users enjoyed that you have not yet seen. It is collaborative because it uses the crowd's collective behavior.
Content-Based Filtering: What the Item Is Like
A second major approach is content-based filtering. Instead of looking at other users, it looks at the properties of items themselves. If you watched three action movies with fast car chases, a content-based system might describe those films by their attributes — genre: action, theme: chase, pace: fast — and then recommend other films sharing those attributes. Content-based filtering works even for new users with no history, because it does not need crowd data. But it can get narrow: if you only ever liked one genre, it might never step outside that genre to discover something surprising.
Match each recommendation technique to its defining characteristic.
Terms
Definitions
Drag terms onto their definitions, or click a term then click a definition to match.
How TikTok's Approach Changed the Game
Older recommendation systems leaned heavily on explicit feedback: star ratings, thumbs up or down, written reviews. The problem is that most people never rate anything. TikTok popularized an approach that relies almost entirely on implicit feedback — especially watch time. If you watch 95% of a short video, that is a strong signal you enjoyed it. If you watch two seconds and swipe away, that is an equally strong signal you did not. By tracking these micro-behaviors at massive scale, TikTok's AI can build a highly accurate interest model for a new user in as little as a few videos — far faster than systems that wait for explicit ratings. This approach also means the AI is constantly updating. Your interest model today is not the same as it was a week ago. The system is perpetually re-optimizing for whatever keeps you watching.
Engagement and satisfaction are not the same thing. An AI optimized for watch time will recommend content that triggers strong emotional reactions — surprise, outrage, fascination — because those emotions reliably keep people watching. Content that makes you feel calm and informed may be less engaging by that measure.
What is the central idea behind collaborative filtering?
Why does a recommendation system optimized for watch time not necessarily show you the content that is best for you?
Flashcards — click each card to reveal the answer
Design a Tiny Recommender
- Step 1: Imagine you are building a book recommendation system for your school library.
- Step 2: List five signals you could collect from students without making them fill out forms (hint: think about checkout history, time spent reading, books abandoned halfway, genres of past checkouts).
- Step 3: Describe, in plain language, how you would use those signals to recommend a book to a student who has never used the system before (cold start).
- Step 4: Identify one way your system could give a bad recommendation even if it perfectly analyzed the data — think about what the data might be missing.
- Step 5: Share your design with a classmate and compare your approaches.