Graph-to-Table Interfaces Control Relational Generalisation

Abstract

Graph-to-table systems combine three design choices: an encoder that maps relational evidence to feature rows, a context constructor that selects labelled examples, and a tabular decoder. We investigate whether these components can be evaluated independently through matched comparisons that change one component while holding the rest of the pipeline and the evaluation constant. Across logistic regression, CATBOOST, and TABPFN-3 on CLUTRR, 57 GraphLog worlds, and 57 knowledge graphs, we show that conclusions about one component can change when the others change. On CLUTRR, absolute positional encodings maintain all distinctions among the evaluated queries and fit the short-path setting nearly perfectly; yet, they generalise worse than relation-count features because the supported prefix acts as a shortcut that breaks down on longer paths. On GraphLog, encoder preference depends on the decoder, while the relation-count representation degrades as CatBoost tree depth increases. Across different knowledge graphs, adding the frozen ULTRA score to local features is consistently useful, whereas adding local features to the score has a smaller and less stable effect. A pointwise TABPFN-3 readout using the full encoder underperforms native ULTRA on 48 of 57 datasets. These results show that encoder assessment depends jointly on context and decoder, while panel-level conclusions can also depend on evaluation and aggregation protocols.

What the experiments isolate

A graph-to-table system turns a relational query into a feature row, selects labelled examples, and asks a tabular model to predict. An end-to-end score mixes all three choices. A representation can retain the information needed for a task while exposing patterns that the decoder cannot learn from its available context.

We use matched comparisons that change one component at a time. Encoder comparisons share evaluation queries and context examples. Decoder comparisons share the same feature values before decoder-specific categorical encoding. Context interventions keep the encoder and evaluation fixed. The experiments cover short-to-long path transfer on CLUTRR and GraphLog, followed by candidate ranking across 57 knowledge graphs.

We also check what the rows can distinguish before fitting a decoder. If differently labelled queries collapse to the same row, a row-only predictor cannot separate them. When every evaluated row is unique, the ceiling is perfect even if the decoder generalises poorly. We report singleton mass alongside the ceiling and test transfer directly on CLUTRR.

How to compare graph-to-table systems

• Separate retained information from learned transfer. Report row collisions and singleton mass, then test alignment and context support. A perfect row ceiling and near-perfect context fit can accompany poor extrapolation.
• Cross the encoder with the decoder and context. GraphLog changes the preferred encoder across decoders; the knowledge-graph experiments change performance through the negative context alone.
• Keep native rankings and aggregation visible. When importing a pretrained graph score, compare against its original ranking. Show dataset and family summaries alongside query-connectivity strata.

The knowledge-graph results use one frozen ULTRA checkpoint, whose pretraining lineages overlap part of the panel. Bootstrap intervals describe sensitivity to the composition of this fixed panel. Leaving out one family at a time shows how much the aggregate depends on individual families.

Code and reproduction

The repository includes feature construction, context generation, decoder evaluation and verification of bundled reference outputs. The reproduction guide separates checking the release artifact from rerunning experiments. Full reproduction requires the external datasets, checkpoints and generated assets listed there.

Citation

@misc{cagiano2026graphtotableinterfacescontrolrelational,
      title={Graph-to-Table Interfaces Control Relational Generalisation}, 
      author={Alessandro Cagiano and Lorenzo Molfetta and Gianluca Moro and Matteo Zignani and Pasquale Minervini},
      year={2026},
      eprint={2610.25993},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2610.25993}, 
}