Tables as Graphs: A Unified GNN-Transformer Framework for In-Context Tabular Prediction

We propose a tabular prediction model that explicitly encodes rows, column names and cells as three node types in a heterogeneous graph. Previous approaches capture only part of this structure: classical methods process rows independently, Transformer-based methods encode schema implicitly through attention patterns, and graph neural network (GNN) methods cover only a subset of node types and are limited to local message passing. In contrast, we model tables as tripartite graphs over rows, column names, and cells, whose distinct semantics warrant separate representations. We train a hybrid GNN-Transformer that uses message passing to propagate schema and value signals, together with cross-row attention for context-aware prediction. Our early results on 45 datasets indicate that the method achieves the second-highest regression score and second-highest classification accuracy, with the strongest results on enterprise data.