This repository contains the central pieces of codes for my mini-project for L45 Representation Learning on Graphs and Networks.
In this project, we investigate DE-MF, a method for learning on structured data proposed by Dai et al. The method provides a connection between graphical models and graph neural networks, and, interestingly, shows some similarity to the popular message passing neural network (MPNN). We first give a revision of the method, along with a simple and effective extension. Then we closely compare it to the MPNN with experiments. By identifying their key differences, we notice a new possibility of improving network’s performance with limited number of parameters.