Deep learning structure for directed graph data

Abstract

Deep learning structures have achieved outstanding success in many different domains. Existing research works have proposed and presented many state-of-the-art deep neural networks to solve different learning tasks in various research fields such as speech processing and image recognition. Graph neural networks (GNNs) are considered as a type of deep neural network and their numerical representation from the graph does improve the performance of networks. In the real-world cases, data

is not only in the form of simple graph, but also they could contain direction information in the graph resulting in the so-called directed

graph data.

This thesis will introduce and explain the first attempt in this domain to apply Singular Value Decomposition (SVD) on adjacency matrix for

graph convolutional neural networks and propose SVD-GCN. This thesis also utilizes the framelet decomposition to help better filter the graph

signals, thus to improve novel structure’s performance in node classification task and to enhance the robustness of the model when

encountering high-level noise attack. The thesis also applies the new model on link prediction tasks. All the experimental results demonstrate

SVD-GCN’s outstanding performances in both node-level and edgelevel learning tasks.