Link prediction is a common problem that allow us to know missing links in a complex networks, unknown preferences of a user who wants that someone recommends him a movie, what a politician will vote… In our work we focus on mix-membership stochastic block models to predict links in bipartite networks [1] but adding node’s metadata like an user’s age, a movie’s genre, a politician’s state... In our approach we consider that metadata are connected with our nodes of the considered bipartite networks forming multipartite networks, and also adding an hyperparameter that tells to the model how is the importance of this metadata when we make predictions. Using our approach in synthetic networks, but with different membership-metadata correlation, we observe that for low number of observed links and some correlation, our model helps to make better predictions. But if we add observed links, we observe that metadata can’t help unless threre is enough correlation. Also we found an optimal value of our hyperparameter for the prediction power of our model.
