Embedding-based Retrieval with Two-Tower Models in Spotlight

Summary

The article details Snap's implementation of an embedding-based retrieval (EBR) system for its Spotlight video platform, utilizing a two-tower model architecture to enhance video recommendations based on user interests and engagement history. It describes the challenges faced in real-time content retrieval and the optimization strategies employed, including the use of dense and sparse features, ResNet-style neural networks, and advanced training techniques like cosine annealing. The system is designed to efficiently generate user and story embeddings, ensuring low-latency responses while maintaining high personalization in content delivery.

Key Learnings

• 1The two-tower model architecture allows for scalable and flexible embedding generation for user and story interactions.
• 2Combining dense and sparse features enhances the representation of user interests, leading to improved recommendation accuracy.
• 3The use of advanced optimization techniques, such as Adam optimizer and cosine annealing, facilitates faster convergence and better model performance.
• 4Implementing in-batch negative sampling helps the model learn effectively from user-story combinations, improving retrieval quality.
• 5The separation of feed processing and retrieval services enhances scalability and allows for handling multiple request types efficiently.

Who Should Read This

Senior Machine Learning Engineers developing scalable recommendation systems using embedding techniques.

Test Your Knowledge

Topics