Tensorial and bipartite block models for link prediction in layered networks and temporal networks

Imagine a team of researchers looking for promising drug combinations to treat a specific cancer type for which current treatments are ineffective. The team has data on the effect of certain pairs of drugs on other cancer types, but the data are very sparse--only a few drug pairs have been tested on each cancer type, and each drug pair is tested in a few cancer types, at best, or has never been tested at all. The challenge is to select the most promising drug pairs for testing with the target cancer type, so as to minimize the cost associated to unsuccessful tests. We can formalize this challenge as the following inference problem: We have a partial observation of the pairwise interactions between a set of nodes (drugs) in different "network layers" (cancer types), and we need to infer which are the unobserved interactions within each layer (drug interactions in each cancer type). This challenge is relevant for the many systems that can be represented as multilayer networks [1-4], and is also formally analogous to the challenge of predicting the existence of interactions between nodes in time-resolved networks [5-11]. For instance, we would face the same situation if we had data about the daily email or phone communications between users, and wanted to infer the existence of interactions between pairs of users on a certain unobserved day; in this case each layer would be a different day. Here, we introduce new generative models that are suitable to address the challenge above. We model all layers concurrently, so that our approach takes full advantage of the information contained in all layers to make predictions for any one of them.