25.5.3.2 Detecting attacks using Profile Clustering
In [18] the observation is made that attacks consist of multiple profiles which are
highly correlated with each other, as well as having high similarity with a large
number of authentic profiles. This insight motivates the development of a clustering
approach to attack detection, using Probabilistic Latent Semantic Analysis (PLSA)
and Principal Component Analysis (PCA).
In the PLSA model [11], an unobserved factor variable Z = {z1, . . . zk} is associated with each observation. In the context of collaborative recommendation, an
observation corresponds to a rating for some user-item pair and ratings are predicted