Graph to table component choices affect relational learning outcomes
Graph-to-Table Interfaces Control Relational Generalisation
Machine Learning
Summary
Mapping complex relationships into tables to learn patterns depends a lot on how each part of the system is built. The authors show that how well the system works can change when you swap out just one part, even if the rest stays the same. For example, some methods fit short relationships well but don’t work as well on longer ones. They also found the performance of parts like encoders and decoders depends on each other and on how results are measured.
What this means in practice
- •For machine learning engineers: Choose and evaluate graph-to-table system components jointly to improve relational pattern recognition models.
- •For knowledge graph developers: Use combined local features and frozen scores for more consistent relational inference across different knowledge bases.
Authors
Alessandro Cagiano, Lorenzo Molfetta, Gianluca Moro, Matteo Zignani, Pasquale Minervini
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.