how to get the prediction via hop-rec

[xiangwang1223]

Dear Authors,

May I ask how to get the prediction scores (i.e., the predicted user-item interaction values), when using Hop-Rec to get the representations of users and items? Just use the inner product of user and item representations? or what is the API to get the prediction scores?

Thanks for your time and consideration.

[chihming]

The current version doesn't provide the API for computing prediction scores. To compute it, just get the inner product of user and item representations.

[xiangwang1223]

Thanks for your reply. I have two more questions.

1. What does the “sample_times” mean? I find that it refers to “number of training samples *millions”, is it different from the real number of user-item interactions?

2. How to decide the “sample_times” for a fair comparison to other baselines? Say, if the number of user-item interactions is 2 million, when negative sampling ratio is 1:1, MF will use 4 million points to train the model, so is the “sample_times” set to 4 in Hop-Rec for a fair comparison to MF?

Thanks for your time and consideration!

[chihming]

In this tool, one update contains 1 positive sample and 1*N negative samples. Suppose sample_times is set to 10, then it runs 10 million updates. For Hop-Rec, one update contains k-step positive samples and k*N negative samples. If you consider the k-step as 1 update, you can set the same number to both. If you consider k-step as k updates, you can set 4/k to MF.

As to other common MF implementations, they might treat "onee epoch" as one unit. In this case, I think 1-epoch is equalivant to |Edges|/1-million updates in pronet in general.