How to use testing data set to measure recommender system algorithm

Question

I am new to recommender systems and am trying to build one using item-to-time CF. Currently, I am trying to evaluate/measure results using MAE. I have one step which is unclear (after I managed to split training dataset and testing data set and computed item-to-item similarities matrix). I understand that I can now do predictions for a given item for a specific user. Let's say from the testing data set, there is User A with 10 items already rated. What should I do with that user? Randomly pick one item I and pretend I didn't know the score of User A given to item I. Then calculate the predicted score using a weighted sum of rest of the 9 items?

Or, should I actually enumerate all 10 items that User A has ratings for. Then each time take out only the target item and using rest of the 9 items to do predictions?

Answer 1

I guess the right answer might be just predict all 10 items User A has ratings for. And compute MAE using all of them.

Update: Just learnt that approach above is called "Judging Model Quality by Residuals", there are others like Cross Validation