Post-Doctoral Research Visit F/M Neurosymbolic Relational Graph Neural Networks

Context

Within the TYREX research team : tyrex.inria.fr

Assignment

Relational learning has recently gained renewed momentum through the development of end-to-end deep learning approaches over relational databases. In this setting, relational tables are typically transformed into heterogeneous graphs, where rows become nodes and primary–foreign key relationships induce edges. GNNs can then propagate information across the relational structure and learn predictive representations without relying exclusively on manual feature engineering. Benchmarks such as RelBench have made this research direction more concrete by providing realistic relational datasets and prediction tasks, while models such as Rel-GNN have shown that relational database structure can be exploited more effectively through carefully designed message passing mechanisms.

Despite this progress, current relational GNNs remain largely statistical. They exploit graph connectivity and learned node or edge representations, but they make limited use of the symbolic and semantic information that is naturally available in relational databases. Such information includes schemas, entity and attribute types, integrity constraints, functional dependencies, temporal constraints, business rules, and domain-specific expert knowledge. As a result, existing models may require large amounts of labelled data, may generalize poorly outside the training distribution, and often provide limited interpretability regarding the relational paths used to produce a prediction.

The central hypothesis of this postdoctoral project is that symbolic knowledge can be introduced into relational GNNs in a scalable and effective way by acting on the message passing structure itself. Instead of treating symbolic knowledge as a separate reasoning layer added after learning, the project will investigate how schemas, constraints, and rules can guide the construction, selection, weighting, and interpretation of relational message passing routes. This is particularly relevant for GraphSAGE-style architectures, which rely on local neighborhood aggregation, and for Rel-GNN-style architectures, which introduce atomic routes as elementary paths enabling more direct and structured information exchange between heterogeneous relational entities.

Main activities

• First, the project will provide a principled characterization of how symbolic knowledge can be represented and injected into relational GNNs.
• Second, it will introduce neurosymbolic extensions of GraphSAGE-style aggregation for relational databases.
• Third, it will develop symbolic and differentiable variants of atomic-route-based message passing inspired by Rel-GNN.
• Fourth, it will contribute to the construction of a benchmark for neurosymbolic relational learning, including datasets, symbolic annotations, tasks, and evaluation protocols.
• Fifth, it will deliver empirical evidence showing when and how symbolic knowledge improves relational learning in terms of accuracy, robustness, data efficiency, constraint satisfaction, and interpretability. Results will be published in first rank venues in AI, databases.