Data Science Stack Exchange is a question and answer site for Data science professionals, Machine Learning specialists, and those interested in learning more about the field. It only takes a minute to sign up.
Sign up to join this community
I am going to build a recommender system using TensorFlow recommender and the two-tower-model. I have wondered, how to choose the size of the embedding dimension. Are there any papers on this for large scale recommender systems? For the example, Google chose a size of 32 dimensions for the movie recommender. My vocabulary contains around 30,000 different items.
Help is highly appreciated!
Generally, the size of the embedding layer is not an important hyperparameter.
Research has found embeddings dimension lower than ~19 are not performant and there is asymptotic improvement after ~200 dimensions.
Multiples of 32 are chosen for hardware efficiency.
