Regularization in logistic model for collaborative filtering

Hi all,

very quick question about collaborative filtering: For the linear models (star-based rating), Andrew added regularization terms while for the binary models (like, not like) he did not. (Week 2, video 4, minute 7:45)

Did he just leave it out to make the formula simpler for education purposes, or is there any reason why the binary model would not require it? At least I cannot think of a reason why not.

Thanks a lot in advance!

Hello @Magnus_Josef_Maichle,

I do not know Andrew’s consideration, but I think he delivers what he means to deliver and maybe that moment is just not about regularization.

However, we know that we need regularization to deal with the high variance problem (or the overfitting problem), so if you are building a model (be it binary or not) that uses a very large NN, then I agree with you that why not using regularization? Right :wink: ?