Publication
FLODO: A Protocol for Collaborative Causal Discovery
Matteo Gregorini; André Meyer-Vitali
In: Maryam Alimardani; Tom Lenaerts; André Meyer-Vitali; Ann Nowé; Joost Vennekens; Shenghui Wang (Hrsg.). Proceedings of the 5th International Conference on Hybrid Human-Artificial Intelligence. International Conference on Hybrid Human-Artificial Intelligence (HHAI-2026), July 6-10, Brussels, Belgium, Pages 359-368, IOS Press, 7/2026.
Abstract
The rising application of Causal Theory in machine learning provides enhancements to both performance and explainability, yet Causal Discovery is mostly studied under the lens of a single agent exploring an environment. This paper presents a theoretical framework for Collaborative Causal Discovery, in which a decentralised collective of agents cooperates to reconstruct a global causal graph of an environment or data space. Decentralised multi-agent systems realistically model partial observability, enhance data privacy, and exploit concurrency, while enabling the possibility of a human-in-the-loop, providing critical domain expertise to resolve causal ambiguities. We formalise a system in which distinct agents (potentially including Human in the Loop (HITL) participants) are limited to observing and intervening upon partial subsections of the global state space. We rigorously define the protocols for information exchange and examine the potential trade-offs associated with communication costs versus the fidelity of the reconstructed causal model. Furthermore, we explore the spectrum of underlying hypotheses regarding the environment’s structure and agent behaviour. Finally, we discuss how explicit modelling of causal relationships facilitates enhanced interpretability and propose a specific role for human agents as coordinators, leveraging their understanding to guide the discovery process and resolve ambiguities in the aggregated causal graph.