Question about adding negative instances to the training set

[nvduc]

Thank you very much for your efforts in creating this great library. I have a question regarding the generation of the training set as follows. According to my understating, a number of negative instances is added to the training set, i.e., _get_pointwise_all_data(dataset, num_negatives) function in util/DataGenerator.py
However, the goal of the item recommendation problem is to recommend N items to a user given the set of items the user has interacted with. Thus, adding negative instances to the training set seems violate the assumption of the item recommendation problem.

Could you please explain the rational behind adding negative instances to the training set and why is it reasonable?

Duc Nguyen