Abstract
While foundation models have revolutionized such fields as natural languageprocessing and computer vision, their application and potential within graphmachine learning remain largely unexplored. One of the key challenges indesigning graph foundation models (GFMs) is handling diverse node features thatcan vary across different graph datasets. Although many works on GFMs have beenfocused exclusively on text-attributed graphs, the problem of handlingarbitrary features of other types in GFMs has not been fully addressed.However, this problem is not unique to the graph domain, as it also arises inthe field of machine learning for tabular data. In this work, motivated by therecent success of tabular foundation models like TabPFNv2, we propose G2T-FM, asimple graph foundation model that employs TabPFNv2 as a backbone.Specifically, G2T-FM augments the original node features with neighborhoodfeature aggregation, adds structural embeddings, and then applies TabPFNv2 tothe constructed node representations. Even in a fully in-context regime, ourmodel achieves strong results, significantly outperforming publicly availableGFMs and performing on par with well-tuned GNNs trained from scratch. Moreover,after finetuning, G2T-FM surpasses well-tuned GNN baselines, highlighting thepotential of the proposed approach. More broadly, our paper reveals apreviously overlooked direction of utilizing tabular foundation models forgraph machine learning tasks.