An embedding-based method for inferring novel interlayers in multilayer networks

In biology, networks are applied for modelling heterogeneous entities (e.g., gene, disease, drugs) and their own interactions. In this context, the multilayer networks allow modelling multiple types of interactions on independent layers, which are in turn interconnected by interlayer edges. Link prediction is a crucial task, e.g., which allows discovering of novel relationships between biological entities (e.g., proteins and genes). The state-of-the-art reports several methods focused on link prediction. However, no one is specifically designed for inferring entire interlayers between the unconnected layers of a multilayer network. In this paper, we presented an in-house method for the inference of entire interlayers from pairs of unconnected layers of interest. The proposed method exploits two main approaches: the first constructs a set of primitive links between unconnected layers of interest, based on properties intrinsic to graph network models; the second refines these based on more complex features denoted from node embeddings to infer the candidate interlayer edges, which ultimately constitute the resulting interlayer. In our experimentation, the proposed method has exhibited an effective capability in inferring novel interlayers, even when the number of nodes within the layers of interest increase. Performance was evaluated by using several well-known Key Performance Indicators. Briefly, results showed an improvement by +15.73% and +116.38% in terms of F1-Score and MCC, respectively. Furthermore, the accuracy improved on average by +46.30%, as can also be seen from ROC-AUC and PR-AUC, which showed +44.48% and +38.45%, respectively.