The Multiple Multiplicative Factor Model For Collaborative Filtering
Benjamin Marlin and Richard Zemel
University of Toronto
ABSTRACT:
We present a discrete latent variable model called the Multiple Multiplicative
Factor (MMF) model. The MMF model has natural generative semantics for data
where multiple factors may influence each element of a data vector. A data
vector is represented in the latent space as a vector of factors that have
discrete, non-negative expression levels. The distribution over values for a
data element is a product of each factor's prediction for that element, taking
into account the degree to which the factor is expressed. The latent, discrete
factor vectors, and multiplicative generative semantics of the MMF model make
it distinct from other generative latent variable models such as factor
analysis, latent Dirichlet allocation, and the mixture of multinomials model.
We present empirical results from the collaborative filtering domain showing
that a binary/multinomial MMF model outperforms a wide range of other rating
prediction methods on two data sets.