On the modelling and impact of negative edges in graph convolutional networks for node classification

Signed graphs are important data structures to simultaneously express positive and negative relationships. Their application ranges from structural health monitoring to financial models, where the meaning and properties of negative relationships can play a significant role. In this paper, we provide a comprehensive examination of existing approaches for the integration of signed edges into the Graph Convolutional Network (GCN) framework for node classification. Here we use a combination of theoretical and empirical analysis to gain a deeper understanding of the strengths and limitations of different mechanisms and to identify areas for possible improvement. We compare six different approaches to the integration of negative link information within the framework of the simple GCN. In particular, we analyze sensitivity towards feature noise, negative edge noise and positive edge noise, as well as robustness towards feature scaling and translation, explaining the results obtained on the basis of individual model assumptions and biases. Our findings highlight the importance of capturing the meaning of negative links in a given domain context, and appropriately reflecting it in the choice of GCN model.