Scaling Instagram Explore Recommendations with Two Towers Neural Networks

The text explains the scaling of the Instagram Explore recommendations system, emphasizing the use of machine learning models such as Two Towers neural networks to make the system more scalable and flexible. It details the multi-stage approach to ranking, leveraging caching and pre-computation, and the use of retrieval sources based on heuristics and ML approaches. The text also delves into the Two Tower neural network model and its role in retrieval, as well as the ranking stages and parameters tuning techniques. Overall, the text highlights the evolution and complexity of the system and outlines plans for further improvements.

• Instagram's Explore recommendations system is one of the largest on the platform, leveraging machine learning to ensure users see relevant content.
• The system uses a multi-stage approach to ranking, incorporating retrieval, first-stage ranking, second-stage ranking, and final reranking.
• Two Towers neural networks are employed in retrieval, allowing for the efficient caching of user and item embeddings.
• The ranking process involves a two-stage approach, utilizing lightweight and heavy models to maintain quality recommendations.
• Parameters tuning is crucial for achieving good online results, involving techniques such as Bayesian optimization and offline tuning.
• The system's growing complexity is addressed through plans for continuous improvement and adoption of new ranking models and retrieval sources.