How can we pre-compute the movie vectors without information about the user?

Based on the video “Recommending from a large catalogue”, when dealing with huge amounts of data it is recommended to pre-compute the vectors for movies, v_m, and the Quiz for the section also said since this process does not involve knowing v_u or x_u. Also when I look at the NN diagram it makes sense since x_m gets fed into the NN and v_m comes out. However, the confusing part is this: if our NN’s parameters are tuned based on a cost function, which DOES make use of v_u, how are we able to fine tune and compute v_m without access to v_u?

Hello @Michael_Khachatrian,

This is the sequence of the idea of pre-computing movie vectors in the lecture:

  1. Train your model.
  2. Model training is done.
  3. Pre-compute v_m and/or v_u based on the trained model
  4. Use the pre-computed v_m for prediction and not for training nor fine-tuning the model finished in step 2.