What is it about?
This paper is about increasing the robustness of Graph Neural Networks against structural poisoning attacks, and also measuring the severity of the perturbation on the graph. For a fair comparison, we showed results on both homophilic and heterophilic graphs.
Featured Image
Photo by Conny Schneider on Unsplash
Why is it important?
Data cleaning is expensive for large-scale graphs. Hence, increasing the robustness on the model layer, and having a metric that tracks the severity of the perturbation can reduce this data cleaning overhead.
Perspectives
I worked on a very intuitive approach where I assembled a very simple MLP network with a variety of GNN layers to tackle structure perturbations. This approach, although being naive, showed impressive results. But in my opinion, an interesting fact about the paper is that we showed that, in some cases, we can even assess the severity of perturbation on the graph.
Haci Ismail Aslan
Technische Universitat Berlin
Read the Original
This page is a summary of: β-GNN: A Robust Ensemble Approach Against Graph Structure Perturbation, March 2025, ACM (Association for Computing Machinery),
DOI: 10.1145/3721146.3721949.
You can read the full text:
Contributors
The following have contributed to this page
