Benchmarking novel graph neural networks

• Main Author: Bhagwat, Abhishek
• Other Authors: Xavier Bresson
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/147982
• Institution: Nanyang Technological University

The human brain’s reasoning is postulated to be done by the creation of graphs from the experiences that are encountered in daily life. Therefore, modelling entities and interactions as a graph can be extremely powerful. We can make use of Graph Neural Networks (GNNs) to analyse physical systems which can be modelled as graphs. As the field grows rapidly, we have seen a steady increase in the number of GNNs of various architectures. Due to this increase, we are finding it even more difficult every day to evaluate the performance of such models. In order to overcome this problem, a benchmark is created which is a toolkit able to evaluate the performance of GNNs fairly with standard datasets. This enables us to compare various models for particular tasks such as graph classification, graph regression as well as edge or link prediction tasks. In this paper, we try to introduce two novel GNNs and methods to experiment and evaluate whether these new methods are able to achieve SOTA results on existing tasks and datasets. A challenge here is to be able to reimplement generalizable versions of these architectures within the benchmark as we move towards larger datasets.