Deep Learning on Graphs: A Survey – Arxiv Vanity

For CNNs, convolution is the most fundamental operation. However, standard convolution for image or text can not be directly applied to graphs because of the lack of a grid structure [6]. Bruna et al. [33] first introduce convolution for graph data from spectral domain using the graph Laplacian matrix L [54], which plays a similar role as the Fourier basis for signal processing [6]. The idea of Eq. (6) is similar to conventional convolutions: passing the input signals through a set of learnable filters to aggregate the information, followed by some non-linear transformation. By using nodes features FV as the input layer and stacking multiple convolutional layers, the overall architecture is similar to CNNs. Theoretical analysis shows that such definition of convolution operation on graphs can mimic certain geometric properties of CNNs, which we refer readers to [7] for a comprehensive survey.