Improving stability of GNNExplainer inlarge citation network datasets

Identifier:  TFM:1426

Handle: https://hdl.handle.net/20.500.11797/TFM1426

Authors:  Cabezas Rodriguez, José Joaquin

Abstract:
Graph Neural Networks (GNNs) is a Machine Learning framework that brings neural networks to graph and relational data. It is of special relevance for areas like social network analysis, biological sciences, chemistry, smart transportation systems and many others, where data can be thought of as a network. Explaining why a GNN made a decision is a challenge, due to the black-box nature of neural networks, but it is crucial when applying it to decision-making processes that affects the life of many. In this work we review the current state of the art and analyze the most well-known method for explaining GNNs, GNNExplainer. We find that its application to academic citations datasets present issues due to the variability of the explanations and we propose a modification for improving stability of the results and interpretability of the graphical explanation. In particular, we propose the use of an adjusted coefficient computed beforehand for every explanation instead of a fixed parameter. We find that our proposal improves the stability by more than 10 \% in experiments using Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) with two citation networks datasets (Cora and Pubmed).