Graph classification with DFS code and LSTM

Abstract

Graphs are data structures constructed by a set of nodes connected by edges. Graph-structured data are highly prevalent in addressing real-world applications and problems within the area of mathematics, computing, molecular biology, and many other related fields. In recent times, machine learning associated with graphs is discovered to be a powerful approach to explore graph information and address various tasks. The graph classification task is one of the tasks that is to predict the class label of a given graph. In this report, we will focus on the graph classification task. The graph convolutional neural networks, including GCNN + global average pool (GCNN+GAP), GraphSage, GCNN with sort pool (DGCNN), GCNN with differentiable pool (DIFFPOOL), can extract the local and global graph information with convolution layers, and thus perform well on the graph classification tasks. At the same time, the recurrent neural networks (RNN) such as GraphRNN and GraphGen show good performance in graph generation tasks. Hence, we would like to try RNN on graph classification tasks to understand if RNN can bring benefits to this area. In this report, we are going to compare both graph convolutional neural networks (GCNN) and recurrent neural networks (RNN) on four graph classification datasets.