Lessons Learned at Instagram Stories and Feed Machine Learning

Instagram machine learning has grown a lot since we announced Feed ranking back in 2016. Our recommender system serves over 1 billion users on a regular basis. We also now use machine learning for more than just ranking Feed and Stories: we source and recommend posts from Hashtags you follow, blend in different types of content together, and power intelligent app prefetching. All of the different ways Instagram uses machine learning deserves its own post, but we want to discuss a few lessons we've learned along the way of building our ML pipeline. We made a few decisions for how we do modeling that have been beneficial to us either by improving our models' predictive power and providing top line improvements or by maintaining the accuracy and lowering our memory consumption.