Relation-Dependent Sampling for Multi-Relational Link Prediction

Multi-relational graphs specifically allow for representing different types of relations and are an effective way to model various types of data including social networks (Balakrishnan & T. V., 2020), knowledge graphs (Vrandečić & Krötzsch, 2014), and biomedical networks, such as interactions between molecules (Krogan et al., 2006). Link prediction (Zhang & Chen, 2018) in these graphs is an actively studied problem that is critical for many applications. For example, drug-drug interaction (DDI) prediction is a task important for drug development and repurposing: on the one hand, some diseases are best treated by combinations of drugs (e.g., antiviral drugs are typically administered as cocktails), on the other hand, one has to know about critical side effects between new molecules and existing drugs (in a stricter sense, side effects are consequences of DDIs).