Abstract
Understanding why events occur—the search for causality—has long been central to human cognition, scientific inquiry, and decision-making. Causal attribution refers to the process of inferring the causes of events or behaviors, and it plays a foundational role across disciplines ranging from statistics and system engineering to epidemiology, economics, and machine learning. Whether diagnosing a system's fault, determining maintenance strategy, or identifying the important components, causal reasoning underpins critical interpretations and decisions. However, effectively and explainable finding causes of effects still remains challenging in both methodology and applications due to the rapidly increasing complexity of the modern system.The thesis is dedicated to introducing a causal-based attribution framework, aiming to provide reliable and white-box solutions for both methodology and practice.
It comprises a series of three studies that address different research challenges in the attribution problem of complex systems. By considering the causes of effects, the root causes of effects, and the sensitivity of variables, we proposed novel posterior causal effects, probability of root causes, and causal-based global sensitivity measures respectively.
The first study proposes multivariate posterior total, intervention, and direct causal effects conditional on the observed evidence. We describe the assumptions of no confounding and monotonicity, under which we prove the identifiability of the multivariate posterior causal effects and provide their identification equations.
The proposed approach can be applied to causal attributions, medical diagnosis, blame, and responsibility in various studies with multiple effect or outcome variables. We also discuss the relationship between posterior causal effects and the classical importance measure in the reliability area.
The second study introduces causality at the individual level to discuss the influence of causal mechanisms on attribution problems, and define the root cause at the individual level. Then we propose a novel causal-based framework for root cause attribution, termed the ‘Probability of Root Cause’ based on the definition. The identifiability and identification formulas of the probability of root cause are provided under monotonicity and no-confounding assumptions. The proposed approach can be applied for root cause analysis, mechanical fault diagnosis, and various areas.
The third study introduces causal effects into global sensitivity analysis problems and proposes a novel causal-based global sensitivity analysis approach. We show that the causal sensitivity indices are well-defined and satisfy a continuity property for information refinements. A double robust given-data estimation strategy is provided, and the consistency is proved. The new importance measures contribute to meeting the increasing demand for methods that make black-box models (e.g. neural networks, machine learning methods) more explainable to analysts and decision-makers.
With the aforementioned studies addressing methodological difficulties and practical challenges across different sides of attribution problems, this thesis aims to contribute to the field of causal inference, as well as insights into decision-making and maintenance in systems engineering. The effectiveness of the methods is validated with some numerical experiments.
| Date of Award | 20 Aug 2025 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Min XIE (Supervisor) |