This paper introduces a forensic framework for detecting AI-generated lyrical and melodic audio using MFCC features, similarity metrics, and multi-algorithm fusion. Tests on 500 samples achieved 95% accuracy in distinguishing human from synthetic audio. The system offers a transparent and efficient method for audio authenticity verification in digital forensics and AI-content detection.

The rapid democratization of superhuman chess engines has precipitated an integrity crisis in online chess. While traditional detection systems rely on statistical aggregation to flag illicit assistance, they often operate as opaque "black boxes," providing binary verdicts without accessible justification. This lack of transparency complicates the adjudication process and undermines player trust. This paper proposes a novel, hybrid Explainable AI (XAI) framework designed to detect and explain engine assistance. Utilizing a dataset of 1,000 games analyzed against Stockfish 16 at a fixed depth of 12, we engineered a feature set that captures non-linear cheating behaviors, including "Perfect Streaks" and complexity-dependent accuracy. These features trained a Gradient Boosting classifier that achieves high detection accuracy. Crucially, we integrate SHAP (SHapley Additive exPlanations) to identify global behavioral indicators and LIME (Local Interpretable Model-agnostic Explanations) to generate case-specific evidence for individual bans. Furthermore, we introduce an automated natural language generation layer that translates these complex XAI metrics into human-readable "Fair Play Reports." Our results demonstrate that this framework not only accurately identifies cheaters but also bridges the interpretability gap, providing arbiters with the transparent evidence required to validate automated decisions.

Artificial intelligence (AI) is increasingly embedded within digitalised transport and mobility systems, including journey-planning platforms, mobility-as-a-service (MaaS), ride-sharing and demand-responsive transport (DRT), dynamic demand-responsive transport (DDRT), and data-driven traffic operations. While these systems promise efficiency and service innovation, the shift toward platform-based access and algorithmic decision-making raises concerns about fairness, trustworthiness, and unequal access to mobility services, mediated in part by digital inequality, particularly for populations experiencing vulnerability. This systematic literature review synthesises research at the intersection of mobility systems, digitalisation, and vulnerability, examining how AI-mediated decision processes shape inclusion and exclusion in transport and mobility systems. Following a PRISMA-based review protocol, over 1,100 records were screened, resulting in ten studies that met the full inclusion criteria. The review examines how vulnerable groups are conceptualised and represented in digitalised and AI-mediated mobility research. It also investigates how fairness and trustworthiness are operationalised in AI-enabled mobility systems to mitigate risks for these populations. Finally, the review evaluates the extent to which existing research aligns with transport justice principles and identifies gaps in developing vulnerability-aware and trustworthy AI in mobility systems. The review highlights that many studies prioritise aggregate performance objectives, while vulnerable groups are frequently underrepresented in datasets and evaluation, and digital exclusion is often treated implicitly rather than as a central mechanism shaping access. Where fairness is operationalised, it is commonly framed through statistical metrics and bias mitigation techniques, with less attention to outcomes aligned with transport justice principles, understood here as a contested and theory dependent construct rather than a fixed standard, such as accessibility, affordability, and adaptive capacity under disruption. By integrating insights from transport justice, digital inequality, and Trustworthy AI, this paper develops an integrative vulnerability-centred conceptual framework and introduces the concept of algorithmic vulnerability to explain how optimisation logics, data dependencies, and platform governance can generate or amplify unequal mobility outcomes. The review identifies key conceptual and methodological gaps and provides a foundation for developing vulnerability aware, inclusive, and accountable AI mediated mobility systems.

IoT devices are increasingly targeted by sophisticated malware, while data privacy constraints make centralised malware detection unsuitable for distributed environments. This paper presents Fed-BLAM, a federated learning framework that integrates parameter-efficient fine-tuning of pre-trained BERT and LLaMA models using low-rank adaptation (LoRA) for IoT malware detection. Distributed clients collaboratively train a shared model without exchanging raw network traffic, supporting privacy-preserving learning across heterogeneous devices. The proposed approach is evaluated using the N-BaIoT and TON-IoT datasets under balanced, highly imbalanced, and non-IID data distributions to emulate realistic deployment scenarios. Experimental results demonstrate that the federated LLM-based models consistently outperform a multilayer perceptron baseline, with the LLaMA model achieving up to 99.9

Large Language Models (LLMs) demonstrate strong performance in reasoning, abstraction, and knowledge synthesis, yet their ability to extract and use causal structure from unstructured text remains an open challenge. We propose an automated pipeline that extracts causal variables and relations from text using self-consistency sampling, constructs Structural Causal Models (SCMs), and applies an intervention-inspired heuristic refinement stage. Evaluated on 90 real-world passages across medical, economics, and policy domains using Qwen3-14B with domain-aware extraction, the proposed method achieves 97–98% intervention accuracy with consistent performance across all three domains. A controlled ablation study using domain-agnostic extraction shows a 46% relative improvement in causal accuracy over correlation baselines, confirming that LLM-based extraction is the primary driver of performance and SCM formalisation provides additional structural regularisation. The novelty of this work lies in the closed-loop integration of LLM extraction with formal causal structure, not in any single component individually.

Student recruitment agencies assess admission likelihood across heterogeneous portfolios of institutions, but historical application data are highly imbalanced: a few universities have thousands of records while many have fewer than 50, limiting standard supervised learning for low-frequency institutions. We propose a hybrid system that combines tree-based machine learning for data-rich universities with a Takagi–Sugeno fuzzy inference system for data-sparse ones. Using a real-world dataset of approximately 8,000 applications, we train Decision Tree, Random Forest, and XGBoost classifiers, achieving 82.2% accuracy and 0.877 ROC-AUC with XGBoost. Permutation importance, SHAP, and conditional feature importance identify tuition fee band, prior academic performance, and university ranking as key predictors; these inform a fuzzy system whose rules are mined via association rule learning and whose membership functions are learned with Gaussian mixture models. A frequency-adaptive blending mechanism combines both systems based on university-specific data availability, providing calibrated probability estimates across the full frequency spectrum.

Sequential decision making with imperfect knowledge of the environment and adversary lies at the heart of defense and security. As such, contemporary reinforcement learning methods which are well adapted to these sorts of problems have the potential to provide the key tools to support decision makers in a military context. This paper presents a literature review of reinforcement learning for decision support in a defense context with a focus on recent work produced by NATO member states. To aid the reader, we provide the prerequisite background in military decision making; decision support systems in general; reinforcement learning basics; simulation; and explainability. Through a systematic review of the literature and using tools such as UMAP, we construct the research landscape and study the emergent trends and research gaps that are present. This allows us to examine the challenges for exploiting this work in reality and finally, we look to the future and how these challenges can be overcome.

This paper proposes a decision framework based on deep learning techniques that is applicable to algorithmic trading. A deep neural network is used to predict the direction of the market based on financial time-series data and technical indicators.The novelty of this research is that the decision gate, which is based on the use of explainability, is integrated into the decision-making process. In this case, the use of explainability is not limited to the interpretability of the decision but is used to validate the decision before the trade is executed. Therefore, the decision is executed only if the local feature attribution supports the decision.This decision framework is tested based on walk-forward validation under realistic constraints. The performance of the decision framework is tested, and the results showed that the decision gate reduces the unnecessary trades while improving the performance which demonstrates control over the black box behaviour of Deep Neural Network using Explainable Artificial Intelligence (XAI).This research showed that explainability is used not only to interpret decisions but also to control risk.The results demonstrate that the proposed approach reduces trading activity by approximately 60% while maintaining or improving overall portfolio performance, leading to lower transaction costs and more stable equity behaviour.

Machine learning models can be beneficial for early diagnosis in healthcare sector. However, many of these models operate as black boxes, limiting their transparency and trustworthiness. To the best of author’s knowledge, this paper proposes a first-time application of a surrogate decision trees to explain the prediction of black-box model for breast cancer classification. Two different black-box models, K-Nearest Neighbors (KNN) and Artificial Neural Networks (ANN), were used on the Wisconsin Breast Cancer (WBC) and Wisconsin Diagnostic Breast Cancer (WDBC) datasets, respectively. The proposed framework uses surrogate fidelity score, Spearman rank correlation, top-feature overlap, and stability analysis to quantify interpretability and explainability of the black-box model. Experimental results demonstrate that surrogate models can effectively approximate with a fidelity score of 0.9486 for KNN model and 0.9510 for ANN model. Also, a strong alignment between surrogate and SHAP explanations is observed for the KNN model, while moderate alignment is found for the ANN model, highlighting the challenges of explaining highly nonlinear systems.

This paper presents an intelligent reservoir characterization framework through a progressive tiered AI approach using well log data. Four tiers of increasing complexity are implemented and compared: Tier 1 employs a baseline Fuzzy Inference System (FIS) with equal weights; Tier 2 utilizes a dual-objective optimized FIS; Tier 3 applies XGBoost machine learning; and Tier 4 incorporates Fourier-denoised XGBoost. Ground truth is established through petrophysical calculations with winsorized thresholding to mitigate outlier bias. Results demonstrate progressive improvement across tiers, with Tier 2 achieving simultaneous gains in both correlation (R2: 0.443→0.469) and classification accuracy (44.2%→51.3%). Tier 3 achieves near-perfect performance (R2=0.995, accuracy=99.8%, with only 3 errors out of 1,634 samples), while Tier 4 confirms that machine learning has already extracted all available signal. STOIIP volumetric analysis validates the economic impact, with Tier 3 reducing estimation error from +44.0% (Tier 1) to -0.8%. The proposed intelligent framework provides a rigorous, interpretable methodology for progressively evaluating AI complexity in reservoir characterization.

This study applies data analytics and machine learning techniques to identify key factors influencing mathematics achievement among secondary school students and to predict their academic performance. We analysed 549 student records from a secondary school between 2018 and 2023. Various factors were examined, including grades, gender, ethnicity, and Special Educational Needs status. After comprehensive data preprocessing and feature engineering, we developed multiple predictive models to forecast mathematics results in secondary education. These models were optimised using GridSearchCV, achieving predictive accuracy with R-squared values exceeding 0.89 across all models, with the Decision Tree model performing best (R2 = 0.912840, Mean Squared Error = 0.341011). Our findings emphasise the importance of additional support for students with special educational needs. We also observed consistent progress from mock exams to final results. Furthermore, this research demonstrated the effectiveness of machine learning in identifying key determinants of mathematical success while providing actionable insights for targeted educational interventions.

Nowadays, in various industries, particularly those related to information technology and telecom, AI-driven projects have increased. The implementation of these projects poses fundamental challenges, and the presence of competent project managers can enhance the likelihood of their success. Therefore, the primary goal of this article is to develop a model for evaluating the competencies of AI-driven project managers. In this context, after reviewing conventional models in this field and conducting surveys with experts based on the IPMA model, evaluation indicators for AI-driven project managers were identified. Subsequently, using the decision tree algorithm and optimizing class weights with the Ant Colony Optimization algorithm, a decision support system was developed for continuous and real-time evaluation of AI-driven project managers. The article's findings indicate that indicators such as Change and Transformation, Resourcefulness, and Time are among the most important competencies for project managers. Moreover, the decision tree model optimized with the Ant Colony Optimization algorithm can accurately identify the competencies of project managers with a 94% accuracy rate, outperforming other algorithms. Additionally, a sensitivity analysis of various model features, conducted using the Shapley Additive exPlanations Algorithm, reveals the impact of each feature on the competencies of project managers.

Student engagement prediction plays an essential role in improving the learning experience for better educational outcomes. The study investigates the utilization of machine learning models, such as Gradient Boosting Regressor, in predicting students' engagement levels along with Random Forest Regressor. To establish a performance baseline, a Random Forest Regressor was employed as a benchmark model. As a result of this, the Gradient Boosting model beat the benchmark with a lower Mean Absolute Error (MAE) of 2.49 after cross-validation and hyperparameter tuning. To improve model explainability, the author employed SHAP (Shapley Additive Explanations) and LIME (Local Interpretable ModelAgnostic Explanations). Moreover, fairness analysis was performed to assess potential biases in the model's predictions. The findings demonstrated the necessity for bias mitigation strategies by showing that skill gaps and emotion score did not induce bias, while frustration level showed slight variations in engagement predictions. The results suggest that explainable AI can provide educators with actionable insights to identify at-risk students and tailor learning interventions. Future research should focus on bias reduction techniques, adaptive learning strategies, and model generalization to ensure fair and effective student engagement prediction in AI-driven educational platforms.

Connected and autonomous vehicles, along with the expanding Internet of Vehicles (IoV), are increasingly exposed to complex and evolving cyberattacks. Consequently, Intrusion Detection Systems (IDS) have become a vital component of modern vehicular cybersecurity. Federated Learning (FL) enables multiple vehicles to collaboratively train detection models while keeping their local data private, providing a decentralized alternative to traditional centralized learning. Despite these advantages, FL-based IDS frameworks remain vulnerable to attacks. To address this vulnerability, we propose an explainable federated intrusion detection framework that enhances both the security and interpretability of IDS in connected vehicles. The framework employs a Deep Neural Network (DNN) within a federated setting and integrates explainability through the Shapley Additive Explanations (SHAP) method. This Explainable Artificial Intelligence (XAI) component identifies the most influential network features contributing to detection decisions and assists in recognizing anomalies arising from malicious or corrupted clients. Experimental validation on the CICEVSE2024 and CICIoV2024 vehicular datasets demonstrates that the proposed system achieves high detection accuracy. Moreover, the XAI module improves transparency and enables analysts to verify and understand the model's decision-making process. Compared with both centralized IDS models and conventional federated approaches without explainability, the proposed system delivers comparable performance, stronger resilience to attacks, and significantly enhanced interpretability. Overall, this work demonstrates that integrating FL with XAI provides a privacy-preserving and trustworthy approach for intrusion detection in connected vehicular networks.

One of the fundamental challenges in managing software and information technology projects is monitoring and predicting project status at the end of each sprint, release or project. Agile project management has emerged over the past two decades, significantly impacting project success. However, no comprehensive approach based on the features of this approach has been found in studies to monitor and predict the status of a sprint, release or project. This study aims to develop a data-driven approach for predicting the status of software projects based on agility features. For this purpose, 22 agility features were first identified to evaluate and predict the status of projects in four aspects: Endurance, Effectiveness, Efficiency, and Complexity. The findings indicate that the aspects of Effectiveness and Efficiency have the greatest impact on project success. Additionally, the results show that features related to team work, team capacity, experience and project objectives have the most significant impact on project success. An artificial neural network algorithm was then used, and a model was developed to predict project status, which was optimized using the Neural Architecture Search algorithm with a 93 percent accuracy rate. The neural network model was interpreted using the SHapley Additive exPlanations (SHAP) algorithm, and sensitivity analysis was performed on the important components. Finally, the behavior of the projects in each category was analyzed and evaluated using the Apriori algorithm.

We explore the current state and future directions of reasoning in Large Language Models (LLMs). Key approaches for enhancing machine reasoning capabilities are reviewed, such as Chain-of-Thought prompting, ReAct, self-reflection, and memory-augmented architectures. We highlight how attention mechanisms and memory modules form the foundation for information integration and context preservation, essential for any reasoning process. Further, we emphasize the computational trade-offs involved in achieving human-like reasoning within LLMs. Through analytical estimates and comparative evaluation, we show that systems aspiring to approximate the depth, coherence, and abstraction of human reasoning require exponentially greater memory, multi-step internal reflection loops, and more energy-efficient architectures. We conclude with a vision for next-generation models that balance reasoning power with computational sustainability, including quantum-inspired architectures and adaptive attention systems.

Computer-aided transport modelling is essential for testing different control strategies for traffic lights. One approach to modelling traffic control is by heuristically defining fuzzy rules for the control of traffic light systems and applying them to a network of hierarchically dependent crossroads. In this paper, such a network is investigated through modelling the geometry of the network in the simulation environment Aimsun. This environment is based on real-world traffic data and is used in this paper with the MATLAB R2019a-Fuzzy toolbox. It focuses on the development of a network of intersections, as well as four fuzzy models and the behaviour of these models on the investigated intersections. The transport network consists of four intersections. The novelty of the proposed approach is in the application of heuristic fuzzy rules to the modelling and control of traffic flow through these intersections. The motivation behind the use of this approach is to address inherent uncertainties using a fuzzy method and analyse its main findings in relation to a classical deterministic approach.

As Connected Vehicles (CVs) increasingly depend on deep learning-based Intrusion Detection Systems (IDS), the need for models that are both accurate and interpretable has become essential. This paper explores the use of Explainable Artificial Intelligence (XAI) techniques to improve the transparency of a Convolutional Neural Network (CNN)-based IDS trained on the CICIoV2024 dataset. We evaluate four widely adopted XAI methods—SHAP, LIME, Integrated Gradients, and Grad-CAM—by examining their ability to explain predictions across various cyberattack scenarios, including spoofing and denial-of-service (DoS) attacks on CAN bus traffic. Our results show that SHAP and Integrated Gradients effectively highlight key features, with SHAP assigning up to 0.16 contribution to specific class decisions. LIME provided near-perfect agreement with the model’s predictions in local explanations, while Grad-CAM offered visual insights aligned with convolutional activations. The CNN model achieved 98.3