where ri is the mean over all users of the ratings for item i, ru is the mean over all
items of the ratings for user u and r is the mean over users and items.
A Hv score is assigned to all users in the database and users are sorted according
to this score. The top r = 10 users with highest score are identified as potential
attackers and are examined to identify a target item. The target is identified as that
which deviates most from the mean user rating. Next, a sliding window of r users
is passed along the sorted user list, shifting the window by one user each iteration.
The sum of the rating deviation for the target item is calculated over the window
and a stopping point is reached when this sum reaches zero. The users traversed
during this process become candidate attack profiles, which are then further filtered
by removing any that have not rated the item or whose rating deviation is in the
opposite direction to the attack. Precision results for this method on an average
attack are reproduced in Figure 25.10, compared with the PCA clustering strategy.
In general, the authors report that this method performs well particularly for midsize attacks, in which other methods show a dip in performance.