The basis of PLSA is a probabilistic mixture model of user behavior,
diagrammed with plate notation in Figure 2.3(a). PLSA decomposes the probability P(i|u) by introducing a set Z of latent factors.
It assumes that the user selects an item to view or purchase by
selecting a factor z ∈ Z with probability P(z|u) and then selecting an
item with probability P(i|z); P(i|u) is therefore z P(i|z)P(z|u). This
has the effect of representing users as a mixture of preference profiles or feature preferences (represented as a probability distribution
over factors), and attributing item preference to the preference profiles
rather than directly to the users. The probabilities can be learned using
approximation methods for Bayesian inference such as expectation