Build TikTok's Personalized Real-Time Recommendation System in Python with Hopsworks

TikTok’s recommender system uses a two-tower model architecture with separate encoders for user queries and videos to create embeddings in the same vector space

System is composed of three main pipelines:

• Feature pipeline: processes user interactions and video metadata
• Training pipeline: creates and updates recommendation models
• Inference pipeline: handles real-time predictions and serving

Key features are stored in a feature store (Hopsworks) which maintains:

• User features (age, country, gender)
• Video features (category, views, likes, length)
• Interaction data between users and videos

Recommendation process has two main phases:

• Retrieval: Uses vector similarity search to find hundreds of candidate videos
• Ranking: Personalizes and orders candidates based on user preferences

System maintains fast feedback loop by:

• Quickly logging user interactions (views, likes, watch time)
• Updating feature values within seconds
• Using fresh features for next predictions

Models are implemented using:

• TensorFlow for embedding models
• CatBoost for ranking model
• Vector index for similarity search

Infrastructure uses:

• Kafka for event streaming
• K-serve for model serving
• Feature store for data management
• Vector database for embeddings

System achieves personalization through:

• Recent user activity history
• User demographic features
• Video metadata and engagement metrics
• Interaction patterns

Model training happens on regular schedules:

• Embedding models updated periodically
• Ranking models retrained frequently
• Vector index updated with new videos

Data validation rules ensure data quality:

• Value range checks
• Data type validation
• Business logic constraints