Too Long; Didn't Read
Embeddings are dense numerical representations of real-world objects and relationships, expressed as a vector. They can be used to accurately represent sparse data like clickstreams, text, and e-commerce purchases as features to downstream models. One-hot encoding was a common method for representing categorical variables, but it creates an unmanageable number of dimensions. Embedding vectors that are close to each other are considered similar. An example, YouTube’s recommender team realized that using the “predict the next video is going to click a user to click to click, and that representation is an art, and dramatically affects the behavior of the embeddings.