TikTok Algorithm MOC
TikTok's recommendation engine is arguably the most sophisticated real-time personalization system ever built. Unlike traditional social platforms that rely on who you follow, TikTok's For You Page is powered by a continuous feedback loop that learns your preferences within a single session. This MOC links together a deep technical breakdown of how it works.
Architecture
- TikTok Recommendation Engine Overview - The big picture: interest graph vs social graph
- Monolith Framework - ByteDance's real-time training system
- Collisionless Embedding Tables - Cuckoo hashing and why hash collisions matter at billion-scale
- Two-Tower Neural Network - User tower + video tower, ANN search for candidate retrieval
- Multi-Stage Ranking Pipeline - The funnel: retrieve 500, rank 100, serve 10
- Real-Time Feedback Loop - Event ingestion, parallel processing, pre-ranking during playback
What the Algorithm Optimizes For
- TikTok Signal Hierarchy - What signals matter and in what order
- Video Completion Rate as Currency - Why completion rate dominates and the 70% threshold
Distribution Mechanics
- Cold Start Problem - How new users and videos are handled
- Progressive Amplification Waves - The viral funnel: 200 -> 50K -> 500K+
Synthesis
- Why TikTok's Algorithm Feels So Good - Tight feedback loops, session-level adaptation
- TikTok as a Search Engine - The 2026 shift toward search intent and TikTok SEO
- TikTok vs YouTube vs Instagram - Comparative analysis of recommendation approaches
Related Notes in the Vault
- How Social Media Is Taking Us To A Dark Path - The attention economy and algorithmic manipulation
- X Like Recommendation Engine For Misskey - A contrast: closed-world ranking vs interest-graph recommendation
- Drifter bot experiment - Echo chambers and algorithmic amplification
- Vector Database - ANN search infrastructure that powers candidate retrieval
- There Is No Benefit To Scrolling - The behavioral cost of infinite-scroll feeds