25.3.4.1 Popular Attack
Let us assume that the recommender system uses the widely studied user-based
algorithm proposed in [27], where similarities between users are calculated using
Pearson correlation3. In a similar manner to the bandwagon attack, attack profiles
are constructed using popular (i.e. frequently rated) items from the domain under
attack.
A high degree of overlap does not, however, guarantee high similarities between
attack and authentic profiles. The bandwagon attack used random filler items to generate variation among ratings with the aim of producing at least some profiles that
correlate correctly with any given user. The Popular Attack makes use of average
rating data and rates the filler items either rmin + 1 and rmin, according to whether
the average rating for the item is higher or lower. Linking the rating value to the
average rating is likely to result in positive correlations between attack and authentic profiles and furthermore also maximises the prediction shift (see Section 25.4)
of attack profiles as computed by the algorithm under consideration (see [25] for
details).4