1
$\begingroup$

I'm currently using a matrix factorization method to generate recommendations (for info on this, check: Matrix Factorization Techniques for Recommender Systems). At the moment, my rating estimate is given by the global mean, user and item biases, and the latent factors.

I'm interested in adding content information to this system. In this case, both my items and users can be represented in a topic (word) space. I was thinking about adding the correlation between user and item (some value in [0,1]) as an additional input source to my rating estimation.

My question is: how to best include this information? just add it to the sum? Is it not a problem if the correlation is a value between 0 and 1, and the global mean and biases are values that can be as big as 100 ?

Thanks!

$\endgroup$
1
$\begingroup$

This paper describes one approach you might consider. There is a python implementation that's easy to use.

$\endgroup$
0
$\begingroup$

You cannot directly add the correlation to the sum of ratings predicted from your sequence of models - one problem being they are not in the same units. One option is to try to fit the residuals of your original model using the correlation values between the users and items. Then for prediction, the correlation value computed for the user-item pair, multiplied by the regression coefficient will give the rating correction to be added to the current predicted rating.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.