Question about Wide and Deep Model for Movie Recommendation

Description

in this example there are some itemID(movie_id) in test data that they don't exists in train data, so during training the model can't build or it is better to say "can't update" embedding for them. in normal situation it's not a problem but "ranking_pool" that was used for evaluation also contains these itemIDs pair with some userIDs therefore they have negative affects on evaluation metric like NDCG because we multiplied some random vectors with learned user vectors. Is there any reason for doing so?