In Deep Learning we try to learn the signal found in the data. Some of the features can be the IDs of real world objects (word / item / category). But what happens when we get in inference time new objects never seen before? How can we prepare ourselves in advance so we can still make sense out of the input?

In this talk you'll understand how to handle OOV (Out Of Vocabulary), specifically in the world of recommendation systems where new items arrive on a daily basis.