Bayesian Personalized Recommendation

[Ferno22]

Together we solved some errors regarding the last development parts of the BRP Model:

• First the number of samples was reduced from 15000 to 100 since it kept overloading the GPU memory. We tried this for our local machines and the Google Collab TPU and it still overloaded.
• Then the functions that called the graph were changed since it was not saving the graph objects properly, returning an empty list.
• The name of the variable for the find similar artists and recommendation sections was changed from session to sess to solve an unrecognized variable issue

After uploading the file, one bug arose which was solved in the following way:

• The find similar artists and recommendation sections were moved from an independent cell to the inside of the graph since otherwise it would not load. Also the print statements for both sections were moved inside the session.
• The model was rerun for 100 samples and it got an AUC of 0.7
• Finally, hyperparameter tuning was performed to select the best number of parameters possible. We ended up with a batch of 30, 50 epochs and 1000 samples (since adding more samples causes the GPU to crash, this was tried for 2500, 5000, 75000 and 10000 samples in reverse order)

[talphaidze]

What is more, most of the information found online was compatible with TensorFlow 1.5 that is now deprecated. Therefore, we had to modify most of the functions to make it compatible with TensorFlow 2.0.