Talk: Democratizing End-to-End Recommendation Systems
5 - 5:30pm
Modern recommender systems require diverse data processing and feature engineering at tremendous scale and usually employ heavy and complex deep learning models that require expensive GPU clusters to shorten the training time. Techniques for developing an effective, performant recommendation system remains a challenge for most data scientists and engineers.
In this session, we first survey the recommendation system landscape. Then, we walk through challenges endemic to building a recommendation system with a focus on the cost prohibitive nature of training a new system from scratch. We then propose an end-to-end solution that makes it much more cost, resource and time efficient to develop a recommendation system. This solution will include practical ways to optimize parallel data processing based on Spark and hyperparameter optimization and model selection with SigOpt. We will then apply this method on commodity CPU clusters to demonstrate how this combination of tooling boosts the pipeline efficiency for typical end-to-end recommendation systems like DLRM and DIEN. Finally, we’ll conclude the talk with a discussion around the future of this process and where further research would be valuable.