What motivated your choice of a two-stage recommender system architecture?
Two-Stage recommender systems are a common practice in industry applications. We had an catalog of 1.5M products. It is computational complex to score all items for each session. A typical approach is to use a simple model (or multiple models) in the retriever stage to generate high quality candidates. The 2nd stage can be a more complex model to rerank the high quality candidates. Alternative, it can be viewed as stacking. Stacking is a popular approach in Kaggle competitions. The 2nd stage is a model, which used the input of many 1st stage models.