Curvature based pooling in Graph Convolutional Networks Archived

Project information

This project is the implementation for my masterthesis "Krümmungsbasiertes Pooling innerhalb von Graph Convolutional Networks". It contains implementations of different pooling-methods for Graph Convolutional Networks and aims to compare them based on achieved training accuracy and a couple of other statistics on a bunch of different datasets.

CurvPool is my new contribution to the field based on the Balanced Forman Curvature (https://arxiv.org/pdf/2111.14522.pdf) a metric for curvature in graphs, while DiffPool (https://arxiv.org/pdf/1806.08804.pdf) and CliquePool (https://arxiv.org/pdf/1904.00374.pdf) are established methods used to evaluate and compare my new method.