Natural counterfactual explanations with causal awareness and actionable recourse for black-box models.

The escalating complexity of artificial intelligence (AI) models, particularly black-box systems, poses substantial challenges to transparency, user trust, and actionable recourse, especially in high-stakes decision-making domains. Counterfactual (CF) explanations, which articulate the minimal input feature alterations necessary to achieve a desired model outcome, offer a promising avenue to mitigate these issues. However, prevailing CF methods often generate explanations that are linguistically unnatural, practically infeasible for end-users, or neglect underlying causal relationships within the data, thereby limiting their real-world utility and trustworthiness. This thesis systematically addresses these critical limitations by developing and validating a multi-faceted framework for generating natural language CF explanations that are simultaneously actionable, causally coherent, and user-centric. The research unfolds in three main thrusts: First, to enhance the comprehensibility and trustworthiness of CF explanations, user studies were conducted to identify effective linguistic constructs. These insights informed the development of the n-XAIT method, which uniquely combines a novel Feature Actionability Taxonomy (FAT), categorising features by their mutability and sensitivity, with templatebased natural language generation (NLG) to produce CFs that are both feasible and clearly articulated. Second, to ensure CFs are not only actionable but also reflect real-world causal mechanisms, the PICACHU approach was developed. This method integrates causal discovery (learning causal graphs and estimating Individual Treatment Effects - ITEs) with the FAT. This synergy ensures that suggested changes are causally plausible and considers the downstream consequences of interventions. This work was further extended by exploring multineighbour strategies and the Plausibility and Actionability ontology to improve solution coverage and domain compliance. Third, to foster user engagement and adaptability, an interactive agentic system was designed and implemented. This system orchestrates the NLG, FAT, and causal awareness components into a conversational workflow, enabling users to iteratively refine constraints and receive dynamically updated explanations, thereby personalizing the recourse process. The primary contributions of this thesis are: (1) an empirically-grounded methodology for generating natural language CF explanations tailored to user understanding and feature actionability; (2) a novel framework (PICACHU) for producing causally-aware and actionable CFs by robustly integrating causal knowledge with actionability constraints; and (3) a proof-of-concept interactive agentic system demonstrating enhanced user-centricity in the explanation process. Evaluations, including user studies across multiple domains (finance, healthcare, education), validate the efficacy of the proposed methods in improving the clarity, feasibility, acceptability, and trustworthiness of CF explanations. This research advances the field of explainable AI (XAI) by providing more human-aligned, actionable, and causally-aware CF explanations for interpreting and interacting with complex AI decision-making systems.

Grantor dc:publisher.institution

Robert Gordon University

Year dc:date.issued

2025

Author dc:creator

• Salimi, Pedram

Advisor dc:contributor.advisor

• N. Wiratunga, D. Corsar and A. Wijekoon

• Explainable artificial intelligence (XAI)
• Counterfactual explanations
• Black-box systems
• Causal modelling
• Natural language generation (NLG)
• User‑centred design

Language dc:language

en

Identifier

oai:rgu-repository.worktribe.com:3217640
https://doi.org/10.48526/rgu-wt-3217640

OAI identifier oai:identifier

oai:rgu-repository.worktribe.com:3217640