Towards efficient learning of GNN on high-dimensional multi-layered representations of tabular data

For prediction tasks using tabular data, it is possible to extract additional information about the target variable by examining the relationships between the objects. Specifically, if it is possible to receive a graph in which the objects are represented as vertices and the relationships are expressed as edges, then it is likely that the graph structure will contain valuable information. Recent research has indicated that jointly training graph neural networks and gradient boostings on this type of data can increase the accuracy of predictions. This article proposes new methods for learning on tabular data that incorporates a graph structure, in an attempt to combine modern multilayer techniques for processing tabular data and graph neural networks. In addition, we discuss ways to mitigate the computational complexity of the proposed models, and we conduct experiments in both inductive and transductive settings. Our findings demonstrate that the proposed approaches provide comparable quality to modern methods.