Post-Doctorant – Neurosymbolic Relational Graph Neural Networks

Type de contrat : CDD Contexte et atouts du poste Within the TYREX research team : tyrex.inria.fr Mission confiée The objective of this postdoctoral research programme is to design, implement, and evaluate a new class of neurosymbolic relational Graph Neural Networks (GNNs) for learning over large-scale relational databases. The project focuses specifically on the integration of symbolic knowledge into scalable GNN architectures inspired by GraphSAGE and recent relational deep learning models, with a particular emphasis on the notion of atomic routes introduced in Rel-GNN. A complementary objective of the project is to contribute to the construction of a new benchmark for neurosymbolic relational learning, designed to evaluate not only predictive performance, but also the ability of models to exploit symbolic knowledge, constraints, rules, and schema-level semantics. Principales activités The first research axis will study the symbolic structures that can be extracted from relational databases and made usable by neural models. These include table schemas, primary and foreign keys, type declarations, constraints, dependencies between tuples, and domain rules. The aim is not to build a full symbolic theorem prover over relational databases, but rather to identify fragments of symbolic knowledge that can be injected as inductive biases into scalable GNN architectures. This will lead to a formal representation of symbolic annotations over relational graphs, including typed nodes, typed edges, constrained paths, rule-compatible neighborhoods, and semantically meaningful route templates. The second research axis will focus on neurosymbolic extensions of GraphSAGE-style message passing. GraphSAGE learns inductive node representations by sampling and aggregating information from local neighborhoods. In relational databases, however, not all neighbors have the same semantic role. Some links encode entity ownership, some encode transactions, some encode temporal events, and others encode many-to-many relationships through bridge tables. The project will therefore investigate symbolic aggregation mechanisms in which messages are filtered, weighted, or transformed according to schema-level and rule-level information. For example, aggregation functions may be conditioned by relation types, constrained by integrity rules, or regularized so that learned embeddings respect known dependencies. The third and most specific research axis will build upon Rel-GNN’s notion of atomic routes. Atomic routes provide elementary relational paths that support structured information exchange between source and destination nodes, especially in the presence of bridge tables or higher-order relational patterns. This project will study how atomic routes can be made neurosymbolic. Rather than considering atomic routes only as structural units for efficient message passing, we will enrich them with symbolic semantics. A route may be typed, associated with a logical rule, constrained by a dependency, or assigned an interpretation derived from the database schema. The model will then learn not only which routes are predictive, but also how symbolic constraints should guide route selection and composition. Several technical directions will be explored. One direction is symbolic route filtering, where only routes compatible with schema constraints or expert rules are made available to the GNN. Another direction is differentiable symbolic route weighting, where rules provide soft priors over message passing paths. A third direction is logic-aware regularization, where the loss function penalizes predictions or intermediate representations that violate known constraints. A fourth direction is interpretable route attribution, where predictions can be explained in terms of the atomic routes and symbolic rules that contributed most strongly to the output. A further axis of the postdoctoral programme will be devoted to the construction of a benchmark for neurosymbolic relational learning. Current benchmarks for relational learning provide important datasets and predictive tasks, but they are not primarily designed to assess the contribution of symbolic knowledge. The project will therefore contribute to the definition of benchmark tasks where schemas, constraints, dependencies, rules, and expert knowledge are explicitly represented and can be used by learning systems. The benchmark will aim to evaluate several dimensions: predictive accuracy, data efficiency, robustness to missing or noisy data, generalization under distribution shift, constraint satisfaction, and interpretability of relational reasoning paths. The postdoctoral researcher will participate in the design of benchmark tasks, the preparation of datasets, the formalization of symbolic annotations, and the implementation of evaluation protocols. The benchmark will also serve as an experimental testbed for the neurosymbolic GNN architectures developed in the project. In particular, it wil