Static representation of temporal graphs

Graph Neural Networks (GNNs) have proven to be a powerful tool for machine learning tasks on static graphs, such as node classification, link prediction, and graph classification. However, many real-world networks are dynamic in nature, with edges and node attributes changing over time. For this reason Temporal Graph Neural Networks have recently emerged as a promising research area for machine learning tasks over this kind of temporal networks. Nevertheless, the added complexity introduced by the temporal dimension exacerbates well-known issues of GNN-based algorithms, such as over-squashing and over-smoothing, while further limiting the scalability of these approaches. In particular, this work focuses on dynamic supervised link prediction. To address this task our proposal is a general procedure that maps a temporal network onto an equivalent static representation through the supra-adjacency method. This static representation allows us to perform an adapted prediction task using simpler GNN architectures. We present two dynamic graph mapping strategies and several architectural variants. Building upon a HeteroSAGE backbone, we introduce targeted refinements to better address the key challanges related to node embeddings scarcisity and cold-start problems and message passing. Each model is described and evaluated step by step, and iteratively improved in response to the identified issues and available solutions. Finally, we discuss the overall results and the potential applicability of our approach to other datasets and tasks.