Link Prediction in Stagewise Graphs

A stagewise graph has distinct edge types that represent the different stages. Two nodes can be connected by an edge of the current stage only if an edge of the preceding one is already connecting them. Stagewise graphs can represent many kinds of interactions. For example, a jobseeker-vacancy interaction can be labeled by the subsequent edge types click, apply, job interview, and, finally, hired. Also, biological and medical interactions, such as infection processes or administration of drugs, often occur in stages. In this work, we formalize link prediction problems on such graphs as ‘stagewise link prediction’. Though relevant and rapidly gaining attention, these types of problems are to date highly underexplored. We identify and address arising difficulties, such as competition in stagewise networks. We explore an activation function for stagewise modelling and an evaluation strategy that satisfies the stagewise constraints. We confirm our insights through a set of experiments on both well-chosen simulated data sets and real data related to job recommendation and synthetic biology.