SlenderGNN: Accurate, Robust, and Interpretable GNN, and the Reasons for its Success

Can we design a GNN that is accurate and interpretable at the same time? Could it also be robust to handle the case of homophily, heterophily, or even noisy edges without network effects? We propose SlenderGNN that has all desirable properties: (a) accurate, (b) robust, and (c) interpretable. For the reasons of its success, we had to dig deeper: The result is our GNNLin framework which highlights the fundamental differences among popular GNN models (e.g., feature combination, structural normalization, etc.) and thus reveals the reasons for the success of our SlenderGNN, as well as the reasons for occasional failures of other GNN variants. Thanks to our careful design, SlenderGNN passes all the 'sanity checks' we propose, and it achieves the highest overall accuracy on 9 real-world datasets of both homophily and heterophily graphs, when compared against 10 recent GNN models. Specifically, SlenderGNN exceeds the accuracy of linear GNNs and matches or exceeds the accuracy of nonlinear models with up to 64 times fewer parameters.