Enhancing Graph Neural Networks with Random Graph Ensembles

Jan von Pichowski, Vincenzo Perri, Ingo Scholtes

Keywords: graph neural network, random graph ensembles, configuration model, data cleaning

Abstract:

Graph Neural Networks (GNNs) have shown remarkable performance in various network analysis tasks.
However, their results depend on the reliability of the network structure, making them sensitive to inherent variability in real-world data.
This study investigates the use of graph ensembles to improve GNN performance, focusing on node classification tasks.
We use random graph ensembles to define edge scores, quantifying the deviation of observed edge frequencies from those expected based on node activity.
This approach allows us to distinguish between statistically significant connections and those potentially arising from random fluctuations in the network structure.
We use this information to refine the message-passing procedure, aiming to enhance node representations and increase performance in downstream tasks.
In our experiments, we propose and evaluate two
ensemble-based strategies.
Our results show that these strategies lead to better GNN performance in four out of five datasets.
Our work lays a foundation for future research, opening new avenues for either applying other random graph ensembles to GNNs, or considering other graph-based tasks.

Submission Type: Extended abstract (max 4 main pages).

Submission Number: 104