Title: Fixed Aggregation Features Can Rival GNNs

Abstract:

The prevailing consensus holds that Graph Neural Networks (GNNs) achieve superior node representation learning through learnable neighborhood aggregation mechanisms. We contest this assumption with the introduction of Fixed Aggregation Features (FAFs), a method that eliminates the need for training by converting graph learning challenges into standard tabular problems. This paradigm shift allows researchers to leverage robust, established tabular techniques, thereby enhancing interpretability and enabling the integration of a wide variety of classifiers. Our evaluation across 14 benchmark datasets demonstrates that carefully tuned multilayer perceptrons utilizing FAFs match or surpass the performance of state-of-the-art GNNs and graph transformers on 12 of those tasks. Notably, this success is often achieved with simple mean aggregation. The two exceptions, the Roman Empire and Minesweeper datasets, appear to necessitate the depth characteristic of specialized GNN architectures. To provide a theoretical foundation for the efficacy of non-trainable aggregations, we link our observations to Kolmogorov-Arnold representations and analyze the conditions under which mean aggregation proves adequate. Ultimately, these findings suggest three critical directions: the development of benchmarks that better exploit diverse neighborhood aggregation learning, the adoption of strong tabular baselines as a standard practice, and the continued application and advancement of tabular models to graph data to foster fresh insights into related domains.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC