Can an id be encoded as a feature for a neural network?

Question

Say for example each user has their own unique model, say for example a users music taste. Sure I could encode generic features of their taste, but I want it to be truly specific and personalized per user. However, rather than making N different models, I encode N as a feature to my model, (0->N). E.g:

age, eye color, time of day, user_id

Lets say for now, I'm working with a neural network. Thoughts on whether a the user_id would be enough of a feature to produce different results for each user?

Answer 1

Definitely! This is exactly what you can use one-hot encoding for. However, this means you will have N extra inputs (where N is the amount of users).

If you have three users, the extra input to your model would look like this:

user1: 001
user2: 010
user3: 100

---

However, if you are running neural computations for a lot of users, you should probably use a network that is indifferent to the user ID (even though you said you didn't want that). Maybe look at some kind of recurrent network (LSTM).