How the TikTok algorithm works - Explained by the CEO

Introduction

The recommendation algorithm is essentially just math. For example, if you liked videos one, two, three, and four, and I liked videos one, two, three, and five, maybe someone else liked videos one, two, three, and six. Because we all liked videos one, two, and three, we will be shown videos four, five, and six, along with each other’s other preferences. This pattern repeats at scale in real time across more than a billion people.

AI and machine learning have enabled this process to be executed at an immense scale. The system quickly learns the interest signals exhibited by users and shows content that is highly relevant to them almost instantly.

Importantly, the platform does not ask users 20 questions about their content preferences or interests. Instead, it incorporates this learning process into the app experience organically. Users "vote" with their actions—by watching a video, swiping, liking, or sharing it. Through these actions, users are essentially giving interest signals.

Mathematically, these signals are aggregated into a formula, which is then analyzed through pattern recognition. This is the basic idea behind the algorithm.

Keywords

Recommendation algorithm
AI
Machine learning
Interest signals
Pattern recognition
User actions
Scale

FAQ

Q: How does the recommendation algorithm work?
A: The algorithm uses math and pattern recognition to recommend videos based on the user's actions and interests exhibited by other similar users.

Q: Does TikTok ask for user preferences directly?
A: No, TikTok does not ask direct questions about content preferences. Instead, it learns from user actions like watching, swiping, liking, or sharing videos.

Q: How does TikTok handle such a large scale of users?
A: AI and machine learning technologies enable TikTok to handle and analyze interest signals from over a billion users in real-time.

Q: What actions are considered interest signals?
A: Actions such as watching a video, swiping, liking, or sharing are considered interest signals.

Q: How quickly does the algorithm learn user preferences?
A: The algorithm learns user preferences very quickly, often showing relevant content almost instantly.