The input recordings dataset should be normalized on a per user basis. This will help in reducing the effect of a single user can have on the overall dataset. Moreover, this will help to establish a relative scale for the recordings of a user.

For example, user A listens to recording C 100 times and recording D 50 times. User B listens to recording C 20 times and recording D 25 times. It would be reasonable to deduce that B likes D more than A. However, currently the system would make the opposite conclusion.

This is the main objective of normalizing the input dataset.