A Top-N Recommender Model with Partially Predefined Structure

Recommender systems can retrieve appropriate results based on users behavioral patterns and preferences. They may be built based on multi-label learning approaches, as each customer transaction may be labeled with several results that interest him/her. It is therefore useful to model the correlations between labels while controlling complexity of the learning algorithm. This paper presents a generative probabilistic model for online resources (products/URLs) recommendation, by capturing the complex local correspondence between the user´s queries and the resources he/she has actually viewed. The structure of our model is partially defined and it is completed according to the observed data. Consequently, several links between observed and/or latent random variables are induced from the training dataset before starting the estimation of parameters. Experiments conducted on real data show the effectiveness of our approach.