I am going through the documentation on recommenderlab package in R. I have few questions on how the user/item based collaborative filtering can be considered as a part of model based approaches. There is no math model being built its only finding the similar items and (loosely) predicting ratings for unknown items basis similarities and ratings of other users. Hence, is it necessary to split the data in train and test and then apply multiple algorithms like user/item/popular/random to find out the one with the least error?
Consider the following 2 scenarios: 1. With data split - best model = IBCF, 2. Without data split - best model = UBCF. Can this actually happen?