Abstract
While foundation models have revolutionized such fields as natural languageprocessing and computer vision, their potential in graph machine learningremains largely unexplored. One of the key challenges in designing graphfoundation models (GFMs) is handling diverse node features that can vary acrossdifferent graph datasets. While many works on GFMs have focused exclusively ontext-attributed graphs, the problem of handling arbitrary features of othertypes in GFMs has not been fully addressed. However, this problem is not uniqueto the graph domain, as it also arises in the field of machine learning fortabular data. In this work, motivated by the recent success of tabularfoundation models (TFMs) like TabPFNv2 or LimiX, we propose G2T-FM, a simpleframework for turning tabular foundation models into graph foundation models.Specifically, G2T-FM augments the original node features with neighborhoodfeature aggregation, adds structural embeddings, and then applies a TFM to theconstructed node representations. Even in a fully in-context regime, our modelachieves strong results, significantly outperforming publicly available GFMsand performing competitively with, and often better than, well-tuned GNNstrained from scratch. Moreover, after finetuning, G2T-FM surpasses well-tunedGNN baselines. In particular, when combined with LimiX, G2T-FM oftenoutperforms the best GNN by a significant margin. In summary, our paper revealsthe potential of a previously overlooked direction of utilizing tabularfoundation models for graph machine learning tasks.