Causal discovery for trustworthy artificial intelligence
File(s)
Author(s)
Russo, Fabrizio
Type
Thesis
Abstract
Artificial Intelligence (AI) is increasingly deployed in high-stakes domains such as healthcare, finance, and public policy, where decisions have far-reaching societal consequences. Ensuring trust in these systems requires moving beyond opaque, pattern-driven predictions to methods that are both effective and interpretable, while aligning with expert knowledge. Causal discovery, which uncovers cause-and-effect relationships from data and represents them as graphs, plays a central role in achieving these goals as a foundation for trustworthy AI.
This thesis introduces three novel methodologies to advance causal discovery and enhance the robustness, transparency, and accountability of AI systems. First, we propose Contestable Neural Networks (NNs), a framework that empowers domain experts to contest NNs, by validating and refining their mechanisms, using causal graphs. By exposing the causal relationships learned by the NN, this approach unlocks transparency and aligns model outputs with expert knowledge, enhancing robustness and accountability.
However, as systems grow complex, placing the entire burden on experts becomes impractical. To address this, we develop the Shapley-PC algorithm, which enhances constraint-based causal discovery by improving the identification of v-structures. By integrating Shapley values, a game-theoretic measure of variable contribution, the algorithm mitigates errors caused by noisy or limited data, ensuring robust causal graph recovery while maintaining theoretical guarantees.
Building on this foundation, we introduce Causal ABA, an Assumption-Based Argumentation framework that resolves conflicts between statistical evidence and expert knowledge. While Shapley-PC improves robustness in finite-sample settings, Causal ABA ensures stronger consistency guarantees between causal discovery outputs and input data. By combining structured reasoning with transparent conflict resolution, it unifies diverse information sources to produce trustworthy causal graphs.
Together, these contributions firmly position causal discovery as a cornerstone for trustworthy AI. By aligning data-driven insights with expert input, improving reliability, and ensuring transparency, this work lays a solid foundation for AI decision-making in high-stakes domains.
This thesis introduces three novel methodologies to advance causal discovery and enhance the robustness, transparency, and accountability of AI systems. First, we propose Contestable Neural Networks (NNs), a framework that empowers domain experts to contest NNs, by validating and refining their mechanisms, using causal graphs. By exposing the causal relationships learned by the NN, this approach unlocks transparency and aligns model outputs with expert knowledge, enhancing robustness and accountability.
However, as systems grow complex, placing the entire burden on experts becomes impractical. To address this, we develop the Shapley-PC algorithm, which enhances constraint-based causal discovery by improving the identification of v-structures. By integrating Shapley values, a game-theoretic measure of variable contribution, the algorithm mitigates errors caused by noisy or limited data, ensuring robust causal graph recovery while maintaining theoretical guarantees.
Building on this foundation, we introduce Causal ABA, an Assumption-Based Argumentation framework that resolves conflicts between statistical evidence and expert knowledge. While Shapley-PC improves robustness in finite-sample settings, Causal ABA ensures stronger consistency guarantees between causal discovery outputs and input data. By combining structured reasoning with transparent conflict resolution, it unifies diverse information sources to produce trustworthy causal graphs.
Together, these contributions firmly position causal discovery as a cornerstone for trustworthy AI. By aligning data-driven insights with expert input, improving reliability, and ensuring transparency, this work lays a solid foundation for AI decision-making in high-stakes domains.
Version
Open Access
Date Issued
2024-12-26
Date Awarded
01/03/2025
License URL
Advisor
Toni, Francesca
Sponsor
UK Research and Innovation
Grant Number
EP/S023356/1
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)