Diffusion convolution recurrent neural network – a comprehensive survey

Graphs find its place in many applications like social network analysis, computer vision and bioinformatics. It has the ability to capture the structural relationship among the data, thus provides more insight. Graph Neural Network (GNN) has a deep learning way of analyzing the graph. The target nodes representation is obtained by iterative propagation of neighbour information until the stability is reached. Representational learning is widely used for capturing the insight of graph representation model. The complex structure of graph is hidden by representational learning results in shallow learning mechanism. Convolutional Neural Network (CNN) exploits the stationary properties and hierarchical pattern of the data which are in Euclidean space. Non-Euclidean characteristics of the graph can be captured precisely using graph convolutional neural network. In graph convolution, vertex domain is represented as aggregation of neighbour node’sinformation. In order to encompass the dynamics of graph, diffusion process is used, in which spatial dependency and temporal dependency are considered simultaneously. In Diffusion Convolution Recurrent Neural Network (DCRNN) uses diffusion convolution to capture spatial dependency and Gated Recurrent Unit (GRU) to capture temporal dependency. DCRNN is capable of handling long-term dependencies. In this survey, we conduct comprehensive survey on diffusion convolutional operations on graph, which is one of the most prominent deep learning models for forecasting in time series domain. First, we categorize the variants of graph convolutional models and its convolution operations on graph. Then based on application, graph convolutional models are categorized with their applications. Finally, open challenges in the area of graph convolutional network and future directions for research are discussed.