User-based Collaborative Filtering: Sparsity and Performance

It is generally assumed that all users in a dataset are equally adversely affected by data sparsity and hence addressing this problem should result in improved performance. However, although all users may be members of a sparse dataset, they do not all suffer equally from the data sparsity problem. This indicates that there is some ambiguity as to which users should be identified as suffering from data sparsity, referred to as sparse users throughout this paper, and targeted with new recommendation improvement strategies. This paper defines sparsity in terms of number of item ratings and average similarity with nearest neighbours and then goes on to look at the impact of sparsity so defined on performance. Counterintuitively, it is found that in top-N recommendations sparse users actually perform better than some other categories of users when a standard approach is used. These results are explained, and empirically verified, in terms of a bias towards users with a low number of ratings. The link between sparsity and performance is also considered in the case of predictions rather than top-N recommendations. This work provides the motivation for targeting improvement approaches towards distinct groups of users as opposed to the entire dataset.