Graph Alignment for Benchmarking Graph Neural Networks and Learning Positional Encodings
Adrien Lagesse, Marc Lelarge
TL;DR
This paper introduces a new benchmarking approach for graph neural networks based on graph alignment, demonstrating its effectiveness for model evaluation, pre-training, and achieving state-of-the-art results with fewer parameters.
Contribution
It presents a novel graph alignment benchmarking methodology, new dataset generation techniques, and shows improved GNN performance and pre-training benefits.
Findings
Anisotropic GNNs outperform standard convolutional architectures.
Graph alignment is effective for unsupervised GNN pre-training.
Achieved state-of-the-art results on PCQM4Mv2 dataset.
Abstract
We propose a novel benchmarking methodology for graph neural networks (GNNs) based on the graph alignment problem, a combinatorial optimization task that generalizes graph isomorphism by aligning two unlabeled graphs to maximize overlapping edges. We frame this problem as a self-supervised learning task and present several methods to generate graph alignment datasets using synthetic random graphs and real-world graph datasets from multiple domains. For a given graph dataset, we generate a family of graph alignment datasets with increasing difficulty, allowing us to rank the performance of various architectures. Our experiments indicate that anisotropic graph neural networks outperform standard convolutional architectures. To further demonstrate the utility of the graph alignment task, we show its effectiveness for unsupervised GNN pre-training, where the learned node embeddings…
Peer Reviews
Decision·Submitted to ICLR 2026
1. The paper combines several potentially impactful contributions in (a) extending the graph alignment augmentation from Nowak et al. (2017) to arbitrary datasets, (b) demonstrating its viability as a benchmarking task, and (c) leveraging it as an unsupervised pre-training task for learnable PEs with good empirical results. 2. The paper construction is overall quite sound; despite the density of the paper it is quite easy to follow and the individual contributions tie into each other very well —
1. I think something the authors somewhat underplay is that it is likely non-trivial to find the optimal noise level for a given dataset; one needs to be familiar with the datasets to estimate appropriate noise level or sweep over a range to determine apt values, similar to section 5. This carries over to generating GAPE as learnable positional embeddings are highly sensitive on the datasets pre-trained on (something also demonstrated by the GPSE paper [1]) as well as the noise level, as the aut
1. The use of graph alignment for benchmarking task can be useful in understanding strengths of different GNNs, going beyond what other benchmarking works in GNNs usually do. 2. Unlike prior synthetic/mathematical benchmarks (such PATTERN, CLUSTER, CSL, etc.) this approach is dataset agnostic and can adapt to diverse graph topologies without modifying the underlying structure. 3. The alignment objective can be used for self-supervised learning and can lead to meaningful learned embeddings, as sh
1. While the paper demonstrates that GAPE as a result of their study works well empirically, the benchmarking utility is not robustly demonstrated. For instance, what is the relationship between alignment accuracy and downstream task performance? For the 3 datasets in Table 1, if the proposed benchmarking gives a trend/rank of different models, do these translate to real task performance? 2. One takeaway from the formulation of graph alignment could be that higher order WLGNNs would be better th
The paper is clearly written and nicely presented with a very interesting focus on the graph alignment problem.
I feel the heart of the authors paper is the reliance on Siamese architecture developed in Nowak et al. 2017. I think it does not become clear whether there is any real contribution adding to this method or whether the authors rely entirely on the 2017 architecture. If so the main contribution made in this paper is the application to their graph alignment datasets that are created using Section 4.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Graph Neural Networks · Graph Theory and Algorithms · Machine Learning in Materials Science