Course talk:CPSC522/Recursive Error Division Validation

I think this idea is really interesting. One thing that was mentioned briefly in the presentations is deciding how to predict ratings from the newly built dataset. I don't have any recommendations in terms of methods, but I think based off of the results of my project, it seems like matrix factorization would not work well for predicting ratings here. From my experience, matrix factorization benefits from having more hidden features, but it doesn't really matter what those hidden features are. Using matrix factorization to predict ratings with your newly built dataset might not be able to capture the nuances of the expanded domain you created. There might be some other methods that would work better for predicting ratings.

I really liked your idea; although you might already have done some steps in your project, I wanted to say your approach remind of iterative contrastive learning approach, where in each iteration the model tries to make the samples near to each other. So maybe using a simple neural network help you with your goal.