International Trade Modelling with Temporal Relational Graph Neural Networks

Abstract

Graph Neural Networks (GNN) are a powerful technique to model data on graph domains with neural networks. They are mainly used on static networks where nodes and edges do not change over time and only one type of edge is present between two nodes. To overcome this issue new models extended the GNN model to incorporate temporal data and relational data. The resulting model is defined as a Temporal Relational Graph Neural Networks (TRGNN). We use this novel technique to model the trade evolution of the International Trade Network (ITN) for different products. The nodes in the network represent the countries, encoded as a feature vector. The products are encoded as feature vectors and the edges represent the trade relations for each product between two countries, encoded as a relational feature vector. We then analyze the topological, statistical properties and predictive capabilities of the developed model. We then analyze and visualize the evolution of the feature embeddings of countries and relations between countries. We then evaluate the predictive performance on link prediction and reconstruction capabilities.