Graph representation learning with deep learning

Abstract

Graphs, a natural and generic data structure, can be seen as the backbone of numerous systems because of its wide usage in domains such as social network analysis, computational biology, and chemoinformatics. With the fast development of deep learning techniques and the great results they achieve in fields ranging from images to natural language processing, there is a surging interest in performing deep learning on graph-structured data to do analytics tasks like classifications and predictions. More recently, new approaches of graph representation learning whose idea is to encode structural information of graphs to embedding space attract public attention. The representations generated from this approach can then be feed as feature input into machine learning modules to do the analysis. However, how to capture adaptive and structural representations of graphs is the key challenge at this stage. In this project, an in-depth study was first conducted to gain an understanding of various representation learning approaches and methods they used to generate node or whole graph representations. Then by applying Graph Convolutional Network (GCN) model together with DIFFPOOL layer, a pooling strategy which can remain the hierarchical structure of graphs, we conducted experiments to test the performance of graph classification using five benchmark datasets. This report states the methodology and implementation details used in the experiments, followed by discussions and analysis of the obtained results.