For the question 17, I don't see how we can only use SGD on the matrix entries corresponding to the new item instead of retraining the model from scratch. As we are adding a new item, the size of the matrix changes, so I am wondering if all the training would not change as well? :/
In theory, adding a new item could change the representation of the users that have rated this item, and these changes would ripple through to the other items that those users have rated, etc. The question here assumes that this effect is minimal. If the user representations are considered 'good enough / stable', we just have to come up with a representation for thew new item, and that's it.
Thank you. And if we are given a new user instead of a new item, and we want to predict his rating on some already existing items. In this case, do we just come up with a representation for the new user and thanks to the similarity with other users, we guess the ratings, or should we retrain?
Pb 17 exam 2020
Hello,
For the question 17, I don't see how we can only use SGD on the matrix entries corresponding to the new item instead of retraining the model from scratch. As we are adding a new item, the size of the matrix changes, so I am wondering if all the training would not change as well? :/
In theory, adding a new item could change the representation of the users that have rated this item, and these changes would ripple through to the other items that those users have rated, etc. The question here assumes that this effect is minimal. If the user representations are considered 'good enough / stable', we just have to come up with a representation for thew new item, and that's it.
1
Thank you. And if we are given a new user instead of a new item, and we want to predict his rating on some already existing items. In this case, do we just come up with a representation for the new user and thanks to the similarity with other users, we guess the ratings, or should we retrain?
Add comment