Are graph neural networks better than standard classifiers?

Abstract

Graph neural networks (GNNs) are becoming very popular these days due to their ability to perform classification and prediction depending on node connections. Since features of samples belonging to the same class can be related, graph-based models may perform classification better than other classifiers. The big challenge for this evaluation is to know if there is a sufficiently adequate relationship between the connections of the nodes to justify the use of these models, since connections to unrelated classes can reduce the capacity of these models. This paper proposes a thorough comparative evaluation between graph models and other well-established classifiers to assess the extent to which GNNs may be superior. This will be done by evaluating the relationships between the probability of connections between nodes and changing database features. The evaluation is performed using synthetic data, which is a task that can be evaluated in future work. The results show that when there are connections of classes different from the node under evaluation, the GNNs lose their advantages over other classifiers.