Hi Ethan, I came across this article of yours on BPR and WARP with LightFM. In the BPR part you mentioned the negative sampling rule is “Randomly select an item j which the user has clicked on fewer times than item i” but in the BPR paper it says “For items that have both been seen by a user, we cannot infer any preference”. I also saw in your code that when creating the sparse matrix, each entry was initialized as 1 even if the user interacted with that item more than once. Could you please help me understand this part more? Does BPR and WARP even handle the case of confidence vs preference described in the Collaborative Filtering for Implicit Feedback Datasets paper? Thank you!