Training a recommendation model with dynamic embeddings

Modern recommenders heavily leverage embeddings to create vector representations of each user and candidate item. These embedding can then be used to calculate the similarity between users and items, so that users are recommended candidate items that are more interesting and relevant. But when working with data at scale, particularly in an online machine learning setting, embedding tables can grow in size dramatically, accumulating millions (and sometimes …

data dynamic embedding embeddings google machine machine learning nvidia recommendation recommendation model recommenders scale tables tensorflow core tensorflow recommenders training vector