University of Ioannina
Advanced methods for explainable artificial intelligence based on fuzzy cognitive maps
Abstract
dc:descriptionAs Artificial Intelligence (AI) systems become increasingly integrated into high-stakes domains such as healthcare, industrial automation, and finance, the demand for transparency, interpretability, and trustworthiness has become paramount. While modern machine learning (ML) and deep learning (DL) models offer exceptional predictive capabilities, their opaque, black-box nature poses significant challenges for human understanding, auditing, and accountability. This has led to a surge of interest in eXplainable Artificial Intelligence (XAI), a research field devoted to creating methods that elucidate the reasoning behind AI-based decisions in a human-comprehensible manner. Despite considerable progress in this regard, existing XAI approaches, particularly post-hoc ones, often suffer from issues of fidelity, inconsistency, and vulnerability to manipulation, which casts doubt on the reliability of the provided explanations. More fundamentally, both interpretable and black-box models typically rely on correlation-based learning, making them susceptible to spurious correlations that compromise their robustness and generalizability. Furthermore, most models fail to incorporate expert domain knowledge, a gap that significantly weakens their utility in data-scarce or rapidly changing environments, where up-to-date human insight is crucial but not captured in existing datasets. In response to the aforementioned limitations, this dissertation investigates the use of Fuzzy Cognitive Maps (FCMs) for developing robust, interpretable, and causally grounded AI models. FCMs resemble knowledge-based recurrent neural networks and are represented as directed causal graphs. They have demonstrated high effectiveness in modeling and simulating dynamic complex systems and are widely applied to predictive tasks such as pattern classification, multi-output regression, and time-series forecasting. Their unique ability to combine expert knowledge with data-driven learning, while providing transparent and dynamic reasoning, makes them particularly suitable for the goals of XAI. Nevertheless, traditional FCMs still face significant challenges, including distinguishing causation from mere correlations in data, managing the computational complexity of analyzing and interpreting large-scale models, and the lack of methods for delivering actionable explanations. This dissertation presents three novel methodologies, each discussed in a dedicated chapter, that collectively aim to overcome these barriers and elevate the role of FCMs in the XAI landscape. The first contribution introduces Information Flow-Based Fuzzy Cognitive Maps (IF-FCMs), which incorporate Liang-Kleeman Information Flow (L-K IF) analysis into the FCM learning process. L-K IF analysis provides a mathematically rigorous framework for causal inference, enabling IF-FCMs to capture authentic causal relationships from multivariate time-series data. Unlike traditional data-driven FCM construction methods that rely on black-box or correlation-based learning algorithms, IF-FCMs embed these inferred relationships as structural constraints during training. This approach eliminates spurious associations, resulting in more accurate and interpretable models. Empirical validation using both synthetic and real-world industrial datasets confirms that IF-FCMs outperform classic FCMs and recent advanced variants. The second contribution addresses the scalability challenge of analyzing large and complex FCMs, where traditional explanation methods, such as causal effect analysis, become computationally infeasible owing to the combinatorial explosion of causal paths as the network size and edge density increase. To overcome this limitation and support interpretability in high-dimensional settings, the dissertation introduces the Total Causal Effect Calculation for FCMs (TCEC-FCM), an efficient algorithm designed to compute the total causal effect of one concept on another across all causal paths between them. TCEC-FCM combines binary search with graph traversal techniques, significantly reducing runtime complexity. Numerical simulations on synthetic networks of varying sizes and densities confirm its superior performance over baseline methods, thereby enhancing the global interpretability of large-scale FCMs. The third contribution introduces CF-BP (Counterfactuals via BackPropagation), a novel method for generating counterfactual explanations in quasi-nonlinear FCMs, a recent extension of traditional FCMs designed to overcome long-standing convergence issues. Counterfactual reasoning is vital for decision support because it helps users understand how input variables must change to achieve a specific outcome. CF-BP frames this as an optimization problem, identifying the minimal input perturbations required to achieve a desired output. This enables intuitive "what-if" analyses and reveals potential model biases, thereby promoting fairness and transparency. Leveraging the similarity between the recurrent reasoning processes of q-FCMs and forward propagation in artificial neural networks (ANNs), the proposed algorithm applies backpropagation to compute the required input changes efficiently. Empirical evaluations against eight model-agnostic counterfactual methods demonstrate CF-BP’s superior performance across key criteria, including proximity, sparsity, validity, stability, and computational efficiency. Beyond their theoretical value, the proposed methodologies were extensively tested in various high-impact domains, demonstrating their practical effectiveness. In industry, they were applied to predictive maintenance, fault detection in industrial and robotic systems, as well as overload diagnosis in electrical power grids, achieving significant improvements in diagnostic accuracy and interpretability. In environmental science, they were employed for the sustainable management of Mediterranean river basins through participatory modelling and simulations under different scenarios, revealing the growing pressure on water resources due to increasing demand and highlighting the importance of socio-economic parameters. Lastly, in medicine, our research showcased the potential of Causal eXplainable AI (CXAI) methods in enhancing transparency, robustness, and trust in clinical decision support systems for cancer diagnosis. Overall, these contributions advance the development of both XAI methodologies and FCMs, providing tools that make AI models more transparent, causally grounded, and reliable, which are key prerequisites for their safe deployment in critical areas.
Degree
thesis:*- Grantor dc:publisher
- University of Ioannina
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Tyrovolas, Marios
- Τυροβολάς, Μάριος
Subjects
dc:subject × 24- Ερμηνεύσιμη τεχνητή νοημοσύνη
- Ασαφή γνωστικά δίκτυα
- Αιτιώδης ερμηνεύσιμη τεχνητή νοημοσύνη
- Αναδρομικά νευρωνικά δίκτυα
- Συτήματα βασισμένα στη γνώση
- Ροή πληροφορίας
- Ανάλυση αιτιακής επίδρασης
- Επεξηγήσεις μέσω Αντιπαραδειγμάτων
- Βιομηχανία 4.0
- Explainable artificial intelligence
- Fuzzy cognitive maps
- Causal explainable artificial intelligence
- Recurrent neural networks
- Knowledge-based systems
- Information flow
- Causal effect analysis
- Counterfactual Explanations
- Industry 4.0
- 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/59828