“Symbolic conditioning of Graph Generative Models (SymboliG)” bando a cascata “FAIR - Future Artificial Intelligence Research” cod progetto PE0000013
Type
Bandi a cascata PNRR - Piano Nazionale di Ripresa e Resilienza
CUP
C63C22000770006
Principal Investigator
Nicolò Navarin (Università degli Studi di Padova - Dipartimento di Matematica "Tullio Levi Civita")
Duration
22/04/2024 - 21/10/2025
Description
The SymboliG project has a foundational character. It aims to bridge the gap between graph generative models and the neuro-symbolic integration of logical constraints.
The main contribution of this research will be the definition of different ways to incorporate domain knowledge in state-of-the-art graph generative models, potentially enabling their applications in new settings (e.g. cases in which the training data is scarce).
To do so, we will extend the current state-of-the-art generative models for graphs, mainly focusing on diffusion and transformer models, allowing the incorporation of background knowledge expressed as weighted first-order formulas. We will exploit several alternative approaches from the literature, including regularization-based methods and counterexample-guided inductive synthesis.
Moreover, we will explore how to obtain logical formulas in user-friendly ways such as exploiting knowledge graphs or ontologies.
We will consider the incremental inclusion of constraints, obtaining an interactive generation process driven by domain experts, where the constraints to include will be driven by the observation of the previously generated graphs.
Moreover, we will study cases in which the input and the output are composed of sequences of graphs, e.g. for modelling the evolution of a world model, allowing us to exploit the proposed models for perception and action in the environment.
Our case studies will show the potentiality of the developed approaches in real-world applications and will have an impact per-se in the respective fields.
From a basic research perspective, we expect the results of the project to be used by other research centres, possibly also in other disciplines.
Moreover, we expect this project to initiate a fruitful research collaboration with the partners of FAIR Spoke 2.
The results of the project will make the inclusion of background knowledge in graph generative models simpler, possibly having an impact on novel scientific discoveries.
Activities
4 postdocs hired with 1-year contracts, several publications in international conferences and journals
Related publications
-Matteo Zavatteri, Davide Bresolin and Nicolò Navarin. "Automated Synthesis of Certified Neural Networks", ECAI 2024.
-Zavatteri, M., Bresolin, D., & Navarin, N. (2024). Automated Synthesis of Certified Neural Networks: Initial Results and Open Research Lines. In Artificial Intelligence and Formal Verification, Logic, Automata, and Synthesis 2024.
-Gonzalo Jaimovitch-López, Luca Bergamin, Fabio Aiolli and Roberto Confalonieri (2024) Integrating 𝐿0 Regularization into Multi-layer Logical Perceptron for Interpretable Classification. In Artificial Intelligence and Formal Verification, Logic, Automata, and Synthesis 2024.
Stefano Fioravanti, Francesco Giannini, Pietro Barbiero, Paolo Frazzetto, Roberto Confalonieri, Fabio Zanasi and Nicolò Navarin. Categorical Explaining Functors: Ensuring Coherence in Logical Explanations. In proceedings of KR 2025:
