Further refinements to the offline evaluation model take advantage
of the temporal aspects of timestamped data sets to provide more realistic
offline simulations of user interaction with the service. A simple
temporal refinement is to use time rather than random sampling to
determine which ratings to hold out from a test user’s profile [53];
this captures any information latent in the order in which the user
provided ratings. Further realism can be obtained by restricting the
training phase as well, so that in predicting a rating or making a recommendation
at time t, the recommendation algorithm is only allowed
to consider those ratings which happened prior to t [25, 53, 85]. This
comes at additional computational expense, as any applicable model
must be continually updated or re-trained as the evaluation works its
way through the data set, but allows greater insight into how the algorithm
performs over time