National Technical University of Athens (NTUA)
Generation and evaluation of semantic counterfactual explanations
Abstract
dc:descriptionArtificial Intelligence (AI) has made significant strides, transitioning from research prototypes to large-scale deployments in healthcare, finance, security, and transportation. Despite their success, AI models often function as opaque black boxes, raising concerns about trust, adoption, and risk in high-stakes applications. Explainable AI (XAI) addresses these issues by developing methods to enhance human interpretability of complex models. This dissertation focuses on counterfactual explanations, which determine the minimal input modifications required to alter an AI model’s prediction. A domain-agnostic, black-box framework for counterfactual generation is introduced, applicable across multiple data modalities, including images, text, and audio. Within this framework, various algorithmic approaches are explored, including Graph Neural Networks (GNNs) for structured data and non-neural optimization techniques for counterfactual synthesis. Beyond generation, this work introduces a novel evaluation methodology for assessing counterfactual optimality, specifically leveraging a back-translation inspired approach to verify whether the applied modifications are truly minimal. This evaluation method provides deeper insights into the balance between the minimality of edits and their semantic validity, revealing properties of counterfactual generation algorithms that would otherwise remain obscured. Additionally, the proposed framework extends beyond traditional XAI applications. It is leveraged for hallucination detection in Large Vision-Language Models (LVLMs) and fine-grained evaluation of generative models in both image and story generation. Furthermore, Large Language Models (LLM) reasoning capabilities are investigated, particularly in riddle-solving, where counterfactual-based interventions enhance logical reasoning in large-scale language models. At the same time, through counterexample-based explanations, the influence of well-known cognitive biases is highlighted, especially when such models are employed as recommendation systems. Comprehensive empirical evaluations validate the framework’s effectiveness across diverse domains, demonstrating its ability to enhance interpretability, robustness, and generalizability in both explainability and downstream AI applications.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Filandrianos, Georgios
- Φιλανδριανός, Γεώργιος
Subjects
dc:subject × 16- Επεξηγήσεις μέσω Αντιπαραδειγμάτων
- Επεξηγήσιμη τεχνητή νοημοσύνη
- Επεξεργασία φυσικής γλώσσας
- Αξιολόγηση
- Τεχνητή νοημοσύνη
- Counterfactual Explanations
- Explainable AI
- NLP
- Evaluation
- Artificial intelligence
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Υπολογιστές, Υλικό (hardware) και Αρχιτεκτονική
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Computer science, Hardware and Architecture
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/59267