Example: Facebook newsfeed ranking. This is typically done by relevance scoring each post using a regression model based on a number of features.
Typical features might include information from the profile of the user, e.g. numerical features such as age, signup timestamp as well as categorial features such as their interests, friends, pages followed, etc. How would I best encode these categorical features for training a neural network?
One-hot encoding the friends list would result in a 2 billion dimensional vector(one for each facebook user). Not only is this way too large but the number of users changes constantly so I would have to constantly retrain the model whenever a new user signs up.
So, how do they do it?