Supervised Learning with Informed Graph Convolution

Graph Convolutional Neural Networks (GCNs) have been undergoing improvements in recent years as well as receiving attention due to their usefulness. However the base idea involves computationally expensive mathematics [1]. Some recent advances have shown that modifications can be made that improve the computational efficiency of the concept without any major loss of accuracy. Here we present an improvement to the simplification of GCNs shown by Simple Graph Convolution (SGC) that increased their overall accuracy while maintaining the algorithmic simplicity [8]. We also show that GCNs may be changed from message passing architectures to information gathering to improve their accuracy. This is achieved using the calculations presented in SGC to create new feature vectors using concatenation rather than simply modifying the original feature vectors.