Abstract
Graph Neural Networks (GNNs) have recently become increasingly popular due totheir ability to learn complex systems of relations or interactions arising ina broad spectrum of problems ranging from biology and particle physics tosocial networks and recommendation systems. Despite the plethora of differentmodels for deep learning on graphs, few approaches have been proposed thus farfor dealing with graphs that present some sort of dynamic nature (e.g. evolvingfeatures or connectivity over time). In this paper, we present Temporal GraphNetworks (TGNs), a generic, efficient framework for deep learning on dynamicgraphs represented as sequences of timed events. Thanks to a novel combinationof memory modules and graph-based operators, TGNs are able to significantlyoutperform previous approaches being at the same time more computationallyefficient. We furthermore show that several previous models for learning ondynamic graphs can be cast as specific instances of our framework. We perform adetailed ablation study of different components of our framework and devise thebest configuration that achieves state-of-the-art performance on severaltransductive and inductive prediction tasks for dynamic graphs.