ABSTRACT. This article examines how union members shape the workplace governance of environmental, social and governance (ESG) practices and digitalization in two liberal market economies: Ireland and the United Kingdom. While scholarship commonly treats ESG as a firm-level reporting/governance domain and digitalization as a technological or work-design issue, this study foregrounds union members as workplace actors who mediate how sustainability and digital technologies are translated into rules, practices, and voice arrangements at work. The article adopts a comparative qualitative design based on documentary and textual analysis of policy and regulatory materials, union and confederation documents, public guidance on industrial-relations procedures, and relevant comparative scholarship. Findings show that union-member influence is conditioned by shared transition pressures, such as concerns over surveillance, insecurity, and fragmented voice, yet varies systematically with national institutional pathways. In Ireland, influence is more strongly supported by institutional mediation and forms of bargaining coordination, enabling unions to connect digital change to broader sustainability and worker-welfare agendas. In the United Kingdom, influence depends more on workplace recognition and hybrid repertoires that combine formal representation with informal organizing, legal claims, and public campaigning. Across both cases, ESG and digitalization generate a dual dynamic: they intensify workplace risks while also opening opportunities to reframe employment relations around accountability, participation, and responsible governance. The article contributes to debates on labour governance by linking ESG and digitalization to worker voice and by conceptualizing union members as mediating agents in workplace transformation.

ABSTRACT. The digital economy has fundamentally reconfigured global production networks, yet how it enhances firm-level efficiency remains theoretically ambiguous and empirically fragmented. This paper conducts one of the earliest systematic literature reviews integrating three siloed research streams—digital economy-firm efficiency, digital economy-GVCs, and GVC-firm efficiency—based on 98 core publications from 2000 to 2025, following the PRISMA 2020 protocol using Web of Science, CNKI, and Scopus with standardized search, screening, and quality assessment procedures. We construct an integrative framework taking GVCs as the core mediator and define three transmission pathways. Heterogeneity exists: private firms and domestic digital input adopters gain larger efficiency benefits due to ownership and input-source moderation. Methodologically, we identify key challenges in measurement validity, metric fragmentation, and endogeneity, and propose a tiered research agenda. This review provides a unified theoretical lens for understanding GVC-mediated digital transformation and offers evidence-based implications for policymakers and manufacturing firms.

ABSTRACT. Web path discovery is a critical reconnaissance step in web security testing. Traditional methods rely heavily on exhaustive dictionary-based fuzzing, which is resource-intensive and generates large volumes of traffic that can trigger defensive alerts on target systems. This paper presents ReasonPath, a novel approach that leverages reasoning-enhanced fine-tuning of Small Language Models (SLM) to improve the accuracy and efficiency of web path prediction for security testing while maintaining reasonable resource usage. By fine-tuning SLMs with chain-of-thought reasoning traces, the proposed method internalizes path pattern knowledge and generates more targeted candidates. Experimental results demonstrate that reasoning-enhanced fine-tuning improves F1 from near-zero (0.0002) to 0.29 for the best-performing 3B model and to 0.26 for the 0.5B model. The Qwen2.5-0.5B fine-tuned SLM achieves comparable performance to larger models with only 1.4 GiB peak VRAM, outperforming both base SLM and larger parameter baselines on efficiency-accuracy trade-offs.

ABSTRACT. Over the past decade, social media advertising has shifted significantly toward influencer marketing, with virtual influencers (AI influencers) emerging as a growing trend. Despite its rapid rise, limited research exists on this topic. This study applies social exchange theory and the principle of reciprocity to examine how AI influencer attributes —such as credibility, attractiveness, and content quality— impact consumer trust, engagement, and loyalty. Using text mining techniques, we analyzed 429,000 Instagram comments (from an initial dataset of 500,000 after preprocessing) to explore these relationships. The findings offer insights into how AI influencers shape consumer behavior and contribute to positive marketing outcomes, addressing a critical gap in the literature.

ABSTRACT. In the era of digital platforms, the rapid flow of information across communities creates both opportunities and challenges for decision-making. thus, this study investigates the phenomenon of cross-platform communication of gaming topics, World of Warcraft as a case study. A dataset of 119,871 was collected from YouTube and Reddit between 2020 and 2025. LDA, DTM, Sentiment Analysis (TF-IDF, VADER), SNA, AND ANN, were used to systematically explore the impact of platform type on topic evolution, user sentiment expression, and dissemination path. The results show that early signals from YouTube cam reliably forecast subsequent Reddit engagement, these findings highlight the potential of big data-driven predictive frameworks to serve as intelligent decision support tools for community managers, digital marketers, and platform designers. By enabling early detection of emerging topics and shifts in sentiment, organizations can proactively adapt communication strategies, manage risks, and enhance user engagement in dynamic online ecosystems.

ABSTRACT. As artificial intelligence transforms e-business practices, educational institutions face mounting pressure to equip students with AI competencies for the digital economy. This paper presents the RFM (Reasonable Challenge, Fun, Motivation) education design model—a theoretically grounded framework for designing AI-enhanced educational e-business information systems (EBIS). Through a design science approach, we integrate generative AI tools (ChatGPT, DeepAI, PixVerse), a chatbot platform (Rochat), and a collaborative system (Padlet) into a postgraduate marketing course and evaluate the resulting AI‑augmented learning outcomes. Evaluation data from students, drawn from quantitative surveys and qualitative assignment analyses, demonstrate that the RFM-guided EBIS design effectively fostered AI competencies, perceived usefulness and student confidence in applying generative AI to marketing tasks. The study contributes both a novel motivational design framework for AI-rich business education and practical insights for institutions seeking to evolve their educational information systems to meet AI-era skill demands. Findings reveal critical design considerations, including tool accessibility, prompt engineering scaffolding, peer learning dynamics, and ethical AI discourse integration.

ABSTRACT. E-Government short for electronic government, also known as e-gov, digital govern- ment consists of the digital interactions between a government and citizens, government and businesses/Commerce, government and employees, and also between government and governments /agencies. E-government is no longer just an option but a necessity for coun- tries aiming for better governance. The e-government performance assessment is important for the development of e-government. As conventional evaluation models for e-government performance are generally too subjective and inaccurate, this study proposed belief Rule- base Inference Methodology using the Evidential Reasoning approach (RIMER) method to evaluating the e-government performance.The aim of the research is to evaluate optimal performance of E-government. The method will generate result an accurate system for E-government assessment.In the research, we employed belief Rule-base Inference Method- ology using the Evidential Reasoning approach (RIMER) for design and development of belief rule-based expert system. In the system, belief rule base (BRB) is used to model uncertain E-Government domain knowledge, the evidential reasoning (ER) approach is employed to build inference engine, a BRB training module is developed for learning the BRB through accumulated a lot of cases. We use a set of simulated data and to vali- date the developed belief rule-based expert system prototype. The results show that the prototype can provide reliable recommendations and the performance of the system can be improved significantly after training BRB using accumulated different cases.The expert system is designed and developed by using Visual Basic, MS SQL Server and Mat lab. It can show that the developed Expert System can generate appropriate results taking uncer- tain information.The achievement of the system is that it has assessed optimal performance of E-government. The developed system handles uncertainty in an efficient manner and generates real system output.

ABSTRACT. Edge detection is a fundamental operation in computer vision and image analysis, but repeated spatial-domain filtering over sliding windows can become computationally expensive when densely applied to an image. This paper presents a quantum Fourier transform (QFT)-based edge filtering framework for efficient edge detection. Instead of applying a two-dimensional filter directly in the spatial domain, the proposed method exploits the separability of Sobel and Prewitt operators and reformulates filtering as a sequence of one-dimensional row-wise and column-wise operations. Before quantum encoding, each row or column signal and each corresponding one-dimensional kernel are transformed into a form compatible with quantum-state preparation through zero extension to a common length, resizing to the nearest power-of-two dimension, and amplitude normalization. The normalized vectors are then encoded into quantum circuits, transformed into the spectral domain by QFT, and converted to statevectors. Spectral filtering is performed by element-wise multiplication between the QFT-domain image and kernel statevectors. Because the resulting product vector is generally not normalized, it is re-normalized before being loaded into a new quantum circuit, after which inverse QFT (IQFT) reconstructs the filtered output. This procedure is applied with a specified stride to MNIST images. To obtain stable quantitative estimates, the evaluation is performed on 100 randomly selected samples per digit class, and mean squared error (MSE) and peak signal-to-noise ratio (PSNR) are computed between classical separable filtering and the proposed quantum QFT-based filtering. The results show that the proposed method preserves the dominant edge structure of the original filters while providing a circuit-based spectral implementation of separable edge detection.

ABSTRACT. The rapid growth of online video platforms has increased the need for effective mechanisms to protect young audiences from exposure to inappropriate content. Traditional manual moderation approaches are often labor-intensive, costly, and difficult to scale to the large volumes of multimedia content uploaded daily. This study proposes an AI-based automated video screening system for youth-appropriate media consumption using a hybrid deep learning framework that combines a pretrained ResNet-50 Convolutional Neural Network (CNN) and a Long Short-Term Memory (LSTM) network. The proposed architecture leverages ResNet-50 to extract high-level spatial features from video frames, while the LSTM captures temporal dependencies across sequential frames to improve contextual understanding of video content. The system is integrated into a browser extension that performs near real-time content moderation through server-side inference and local caching of classification results using IndexedDB. Experimental evaluation was conducted on a combined dataset consisting of YouTube-8M, NSFW video samples, and a custom youth-oriented dataset containing safe and unsafe content. Results demonstrate that the proposed ResNet50–LSTM framework outperforms conventional keyword-based and metadata-based filtering approaches, achieving 93.2% precision, 91.6% recall, 92.4% F1-score, and 92.3% accuracy. The findings highlight the effectiveness of combining spatial and temporal visual analysis for automated content moderation. This work contributes to the development of scalable, reliable, and practically deployable AI-driven moderation systems that support safer digital media environments for youth audiences.

ABSTRACT. Lip reading aims to decode spoken words from visual lip movements, but it remains challenged by the high computational requirements and the difficulty of accurate discriminating lip motion patterns. Although recent spatio-temporal deep learning and transformer models have improved recognition performance, many rely on complex network architectures that limits practical deployment. To address these limitations, this study proposes an efficient and effective spatio-temporal lip reading framework that combines lightweight feature extraction with discriminative temporal aggregation. The proposed architecture combines a compact depthwise separable 3D Convolution with inspired residual blocks enhanced with Squeeze-and-Excitation (SE) modules to strengthen spatial representations with minimal overhead and a hierarchical Bidirectional GRU encoder to capture long-range temporal dependencies. To further improve robustness against visual ambiguity, NetVLAD-based cluster-guided aggregation is introduced to pool temporal descriptors according to learned visual word clusters, enabling the network to capture the complex lip-motion distributions. Evaluations on the Lip Reading in the Wild (LRW) benchmark demonstrate that the proposed model achieves a precision of 0.87, recall of 0.85, and F-score of 0.86. Additional evaluations on two specialized LRW subsets show strong robustness under challenging conditions, achieving F-scores of 0.74 on the HOMOPHONES dataset for visually confusable words and 0.84 on the INTER-S dataset for inter-speaker variability. Furthermore, the model requires only 21.05×10⁹ FLOPs and 19.61×10⁶ parameters, offering substantial computational savings compared to state-of-the-art baselines.

ABSTRACT. Choppy pulse (CP) is one of the pulse patterns with important diagnostic value in traditional Chinese medicine pulse diagnosis. However, its assessment still depends heavily on the clinical experience of physicians, and therefore objective and standardized identification methods are still needed. In recent years, deep learning has been widely applied to medical image classification tasks. However, studies on CP waveform images have rarely investigated the complementary value of different field-of-view scales in a systematic manner. Therefore, this study proposes a CP image classification method based on multi-scale feature fusion with a Light ConvNeXt-style architecture. The purpose of this study is to compare the classification performance of single-scale models and multi-scale fusion models in the CP classification task, and to further examine the role and complementary effect of waveform information from different scales. The dataset was constructed from 6-second pulse wave images measured by a pulse diagnosis instrument in 56 CP cases confirmed by traditional Chinese medicine physicians. A single pulse waveform was first extracted from each 6-second recording. Then, multi-size images were cropped with the highest waveform peak as the center to establish a binary dataset of CP and Non-Choppy Pulse (NCP). The data were divided by patient-wise split into training, validation, and test sets in order to reduce the risk of data leakage. This study focused on three representative scales, namely 33×33, 61×61, and 105×105, which correspond to local detail, main contour, and global context, respectively. Six fixed experiments were designed, including three single-scale models, two dual-scale fusion models, and one triple-scale fusion model. The results show that multi-scale fusion models can integrate waveform features from different field-of-view levels and can improve the classification performance of CP images. Overall, this study suggests that CP images contain multi-level information, and that multi-scale feature fusion can provide a more complete basis for classification. These findings may serve as a useful reference for the development of intelligent assistance systems for traditional Chinese medicine pulse diagnosis and may further improve their completeness and practical value.

ABSTRACT. Electronic Health Records (EHRs) represent one of the most sensitive and valuable assets in modern healthcare. Despite widespread digitalisation, EHR systems remain fragmented across clinical institutions, impeding collaborative diagnostics and population-level analytics while raising significant privacy and security concerns. Existing federated learning (FL) approaches address data siloing but remain vulnerable to gradient inversion attacks and lack transparent auditability. Blockchain-based solutions provide immutability and decentralisation yet introduce latency and scalability bottlenecks when used for direct data handling. The present work introduces MedFedChain, an innovative framework that integrates: (i) the adaptive differential privacy (ADP) process that calibrates noise injection based on per-layer gradient sensitivity; (ii) an InterPlanetary File System (IPFS)-based encrypted model repository for efficient off-chain storage of FL model updates; and (iii) Ethereum smart contracts for decentralised, tamper-proof aggregation and incentivisation across federated hospital nodes. Extensive experiments on two recent, large-scale public healthcare datasets---the Kaggle Diabetes Prediction (2023, 100{,}000 records) Dataset and the CDC BRFSS Diabetes Health Indicators Dataset (2021, 253{,}680 records, CDC/UCI)---demonstrate that MedFedChain achieves a classification accuracy of 91.8\% with a privacy budget of $\varepsilon = 1.0$, outperforming FedAvg (85.4\%), FedProx (87.1\%), DP-FedAvg (88.3\%), and EPP-BCFL (89.6\%) whilst reducing communication overhead by 34.6\%. Security analysis confirms resistance to membership inference and model poisoning attacks. These results establish MedFedChain as a scalable, privacy-compliant, and practically deployable solution for next-generation EHR analytics.