Authors - Majed Rahman, Mashrur Wasek, Faisal Quader Abstract - A typical Security Operations Center (SOC) receives 500– 10,000 alerts a day, 80–95% of them false alarms, but a Tier 1 analyst can only meaningfully look at 20–30 in a shift. We present Tier 0 SOC Agent, an AI layer that triages every alert before a human sees it. For each alert, the agent runs a four-phase loop (plan, investigate, decide, explain) over a Model Context Protocol (MCP) ecosystem of 45 tools across ten data sources, and returns a structured verdict with a confidence score, MITRE ATT&CK techniques, and the evidence behind the call. A multi-factor scorer combining source reliability, evidence diversity, and severity routes each alert to one of four tiers: auto-close, T1 with a pre-written summary, T2, or T3. In a four-day production deployment across two customer tenants and 27 Wazuh agents, the agent triaged 93 alerts in a mean of 118 s each, roughly an order of magnitude faster than a human, with zero runtime failures and a 64% drop in tool calls per alert versus the previous generation. A safety scaffold of five downgrade-only policy rules, a circuit breaker, a file-based killswitch, and a read-only tool filter guarantees the agent cannot auto-close any alert beyond what its verdict and the policy jointly authorize.
Authors - HSIU-FENG WANG, PEI-LUN LEE, MAO-JEN CHEN Abstract - With the advent of new developments, social robots are increasingly adopted in educational settings. The impact of social robot gestures on children’s learning outcomes, emotional responses, and affect was explored. One hundred and five elementary students participated in the experiment. Findings revealed a significant effect of robot gestures on children’s learning emotions. The impact of social robot gestures on learning outcomes and affect in children was explored. Results indicated that robot gestures significantly influenced learning emotions and knowledge retention. Children preferred interacting with the gesturing social robot, suggesting enhanced emotional engagement during learning. However, the effect on knowledge transfer was not significant. These findings underscore the potential of social robot gestures in enhancing children’s engagement and learning effectiveness.
Authors - Mashrur Wasek, Prishita Taksalee, Faisal Quader Abstract - University rankings influence student enrollment decisions, institutional funding, and strategic planning, yet diagnostic analyses often conflate correlation with actionable drivers; for example, by including prior-year rank as a predictor, which exploits persistence rather than identifying genuine drivers of rank change. This study presents a reproducible, leakage-safe framework for diagnosing rank determinants using a 31-year panel (1996–2026) of 3,785 university-year observations across 191 U.S. institutions, constructed from College Scorecard administrative data and OpenAlex bibliometrics. The approach explicitly distinguishes diagnostic modeling—estimating expected rank from contemporaneous institutional attributes—from predictive modeling, which exploits temporal persistence. The framework integrates dual methodological lenses: a U.S. News–style observable composite and a Times Higher Education (THE)–style research/teaching proxy framework. Driver importance stability is quantified through year-by-year coefficient analysis, revealing that retention, graduation rate, and selectivity exhibit the highest magnitudes while Pell share, admission rate, and several secondary resource related variables show substantial temporal drift. Case studies of MIT, University of Maryland–College Park, and George Washington University demonstrate how residual analysis and feature decomposition translate into institution-specific improvement priorities. The framework provides a transparent, reproducible blueprint for understanding rank movements and identifying high-impact interventions.
Authors - Jude Osakwe, Josephina Muntuumo, Daphine Gondo Abstract - Digital ecosystems have altered how organisations approach marketing decision-making, creating demand for advanced data management methods. This paper analyses 144 peer-reviewed articles published between 2013 and 2025, examining customer segmentation algorithms, real-time data processing systems, and integrated marketing analytics platforms. Organisations implementing these techniques report marketing return on investment improvements of 15 to 25 percent and conversion rate gains of 10 to 30 percent. A structured comparison of four technique categories, namely big data analytics, machine learning, AI and natural language processing, and predictive analytics, reveals distinct performance profiles and deployment trade-offs. Big data analytics delivers the broadest gains but demands the highest infrastructure investment, while predictive analytics offers a lower-cost entry point with shorter payback periods. Persistent challenges include data quality deficiencies, skills shortages, integration difficulties, privacy compliance obligations, and organisational resistance to change. The paper contributes a systematic framework and benchmarking evidence, and maps implementation constraints that explain the gap between reported performance benchmarks and operational outcomes
Authors - Firangiza Komilovna Fozilova, Bakhtiyor Vasievich Nasimov, Feruza Shodibek qizi Hamrayeva, Shoh-Jakhon Khamdаmov, Fazilat Akhmedova, Saida Khabibullaeva, Doniyor Niyozmetov Abstract - Environmental, Social, and Governance (ESG) reporting has become a strategic priority for chemical enterprises due to increasing regulatory pressure, stakeholder expectations, and sustainability risks. However, traditional ESG reporting practices in the chemical industry are often fragmented, manually intensive, and limited in real-time transparency. The rapid advancement of digital technologies provides new opportunities to transform ESG disclosure into a datadriven and integrated reporting system. This study proposes a digital transformation framework for ESG reporting in the chemical industry, emphasizing the integration of enterprise resource planning systems, blockchain-based traceability, big data analytics, and automated sustainability dashboards. The research develops a conceptual model linking digital infrastructure capability, ESG data quality, reporting transparency, and stakeholder trust. Using a quantitative approach, the study employs Structural Equation Modeling (SEM) to test the relationships among digitalization level, ESG reporting efficiency, and organizational performance indicators. Empirical findings indicate that digital integration significantly improves ESG data accuracy, reduces reporting costs, and enhances transparency. Moreover, improved digital ESG reporting positively influences stakeholder confidence and corporate sustainability performance. The proposed framework contributes to the literature by bridging digital transformation theory and ESG reporting research within high-impact industrial sectors. The study provides practical recommendations for chemical enterprises seeking to enhance sustainability governance through digital technologies and offers policy implications for promoting standardized digital ESG ecosystems.
Authors - Chann Nyein Soe, SOKOUT Hamidullah Abstract - Developing countries determine the housing market based on how far from the city center. However, in Myanmar, submarkets are stronger than the Central Business District (CBD). The differences in the data analysis become results of the collected housing data for 45 days in Yangon, Mandalay, and Pyin Oo Lwin. Price per square foot is the basic unit to compare different properties. The manually collected dataset consists of 710 total listings, in which apartment in Yangon, landed houses, and vacant land in all three cities taken into consideration. Descriptive diagnostic methods, hedonic regressions, and segmentation were used. The result indicates that Myanmar housing prices are not determined by a smooth general rule, but rather distinct submarkets and different pricing behaviors exist. Townships are strongly substantial, whereas apartments show the strongest evidence of segmentation. Although CBD is statistically meaningful, the effect differs across property types. Meanwhile, housing markets exhibit significant differences for each property type. On the other hand, apartments are naturally based on the utility area and are more stable. Housing markets, price estimation models, and property valuation in Myanmar need to change from a city-level single average model to individual property and each township model, relying on median instead of means, because of the outliers in elite properties. Despite the scarcity of estate data in Myanmar, the research supports the Myanmar housing market insights for developing data analysis.
Authors - Farangiz Furkatzoda Majidova, Guljahon Madrahimova, Fotimabonu Doniyorova, Dinora Alisherovna Baratova, Shoh-Jakhon Khamdаmov, Samandar Pulatovich Kurbonov, Dekhkanova Nargiza Sharifovna Abstract - The digitаlizаtion of heаlthcаre governаnce hаs become а centrаl instrument for improving trаnspаrency, regulаtory compliаnce, аnd public heаlth outcomes. In 2024, Uzbekistаn introduced а mаndаtory electronic prescription (e-prescription) system thаt prohibits phаrmаcies from dispensing prescription drugs without ver-ified digitаl аuthorizаtion. Phаrmаcies were grаnted reаl-time online аccess to physiciаns’ prescriptions through аn integrаted nаtionаl plаtform, mаrking а sig-nificаnt regulаtory shift in phаrmаceuticаl mаrket control. This study exаmines the impаct of digitаl prescription systems on phаrmаceuticаl governаnce аnd dis-pensing prаctices in Uzbekistаn. The reseаrch develops а conceptuаl frаmework linking digitаl heаlth infrаstructure, regulаtory compliаnce, dispensing trаnspаrency, аnd mаrket аccountаbility. Using а quаntitаtive аpproаch, the study аnаlyzes survey dаtа from phаrmаcy mаnаgers аnd heаlthcаre professionаls, complemented by secondаry regulаtory dаtа. Structurаl Equаtion Modeling (SEM) is employed to аssess the relаtionships аmong digitаl system integrаtion, compliаnce behаvior, аnd governаnce outcomes. The findings indicаte thаt the introduction of the e-prescription plаtform significаntly reduced unаuthorized drug sаles, improved trаceаbility of phаrmаceuticаl trаnsаctions, аnd strength-ened regulаtory oversight. Digitаl verificаtion mechаnisms enhаnced trаnspаrency аnd reduced opportunities for informаl dispensing prаctices. Fur-thermore, the reform contributed to improved stаkeholder trust in the phаrmаceu-ticаl distribution system. The study contributes to digitаl governаnce аnd heаlth policy literаture by providing empiricаl evidence from аn emerging economy un-dergoing rаpid heаlthcаre digitаlizаtion. The results offer prаcticаl implicаtions for policymаkers seeking to strengthen phаrmаceuticаl regulаtion through in-tegrаted digitаl heаlth infrаstructures.
Authors - Merja Drake, Mariitta Rauhala Abstract - Artificial intelligence (AI) is increasingly embedded in knowledge work, reshaping how tasks are performed and how work is experienced. While prior research has emphasized efficiency and productivity gains, less is known about how everyday human–AI interaction influences employees’ well-being and sense of meaningful work. This study examines AI-enabled work through the dual lenses of hedonic enjoyment and eudaimonic well-being, focusing on how pleasure-oriented and growth-oriented experiences emerge in organizational contexts. Drawing on a mixed-methods design, the study combines a survey of 474 Finnish knowledge workers with 44 in-depth interviews conducted in spring 2024. The analysis explores how generative artificial intelligence use and development re-late to psychological resources such as self-efficacy, proactivity, autonomy, work engagement, and social participation. The findings show that generative artificial intelligence often functions as a hedonic facilitator by streamlining tasks and reducing effort. However, more sustained well-being is associated with eudaimonic experiences, particularly opportunities for learning, participation, and meaningful contribution to AI development. Importantly, the results highlight that access to generative AI tools alone is in-sufficient. Organizational practices such as transparent communication, inclusive development processes, and shared learning environments play a central role in shaping positive human–AI interaction. The study contributes to Human–Computer Interaction research by offering a human-centered perspective on generative AI as a sociotechnical system that can support both enjoyment and flourishing at work.
Authors - Dilobаr Isomjonovnа Ruzieva, Gulasal Madrakhimova, Shirin Akmalovna Zakirxodjayeva, Gavkhar Jumanova, Dilshoda Akramova, Akromjon Rustamovich Gulamov, Ogabek Sadullayev Abstract - This research evaluates the impact of the 2024 green bond issuance by Agrobank ($455 million total) on agricultural productivity and climate resilience in Uzbekistan. Utilizing a quantitative approach, the study analyzes the 2025 Green Bonds Allocation and Impact Report and performance metrics from over 19,000 sub-loans distributed across the Republic's agricultural regions. Findings indicate that the $357.8 million allocated to climate-smart technologies as of late 2025 resulted in annual water savings of 1.003 billion m^3 and an energy reduction of 349 GWh, directly correlating with a 20–30% increase in net farmer income. The data demonstrates that green-financed drip irrigation systems significantly mitigate the risks of water scarcity while lowering operational costs through reduced fertilizer and electricity usage. The study concludes that green bonds are an effective mechanism for modernizing Uzbekistan's agricultural sector and recommends further expansion of green credit lines to achieve the "Green Economy 2030" targets.
Authors - Vijak Sethaput, Supachate Innet Abstract - The rise of digital currencies—encompassing cryptocurrencies, stablecoins, tokenized deposits, and central bank digital currencies (CBDCs)—has outpaced frameworks for understanding their coexistence as financial infrastructure. This paper proposes the Digital Finance Infrastructure Framework (DFIF): a four-layer architecture—Settlement, Commercial, Retail, and Open— that maps five digital currency categories to complementary infrastructure tiers and explicitly links each tier to smart-systems enablers, sustainability dimensions, and security considerations. This paper makes four contributions: (1) a seven-dimensional comparative taxonomy of five digital currency categories; (2) the DFIF framework; (3) a four-dimensional interoperability model; and (4) a six-category systemic risk taxonomy with evidence-based smart-system mitigations. We situate the DFIF against four established frameworks (IMF Money Flower, BIS Singleness of Money, Carstens Digital Money Galaxy, Brunnermeier Digital Currency Areas) and validate it against deployment evidence, including BIS mBridge (reaches MVP stage and BIS exit), Project Agorá (seven central banks), MiCA full applicability, Nigeria's eNaira (
Authors - Banu Yergesh, Tilekbergen Mukhamet, Manas Yergesh, Aisha Zhumagulova Abstract - Emotion recognition for Kazakh is constrained by limited labeled data and by indirect affective expression through idioms and culturally grounded lexical cues. This paper quantifies the contribution of emotion-annotated phraseology and a semantic knowledge base to seven-way (single-label) emotion classification (Ekman’s six basic emotions plus the Kazakh-specific shamerelated class, uiat). The aim of this study is to assess the contribution of emotionally annotated idioms and semantically labeled lexical units to emotion recognition in Kazakh texts. We hypothesize that both types of resources improve classification performance, while their combined use yields the strongest effect, especially for emotions with culturally and pragmatically marked meanings. We curate KazEmoPhras with 3379 emotion-bearing idioms in Kazakh, and two semantically tagged lexical resources (1,200 media units; 1,800 classical literature units). Fine-tuning XLM-RoBERTa on the augmented data improves accuracy from 68.3% to 74.9% and weighted F1 from 0.26 to 0.37. We additionally provide an impact-focused ablation protocol to isolate the effects of idiom and semantic features and discuss practical requirements for morphologyaware preprocessing in Kazakh.
Authors - Usman G. Abdullahi, Kashifu I. Abdullahi Abstract - Weak governance of public-sector digital investments continues to undermine the effectiveness of digital transformation initiatives in many developing countries, resulting in inefficiencies, project failures, and erosion of public trust. While existing literature has largely focused on digital service delivery and user adoption, less attention has been given to upstream governance mechanisms that shape how digital investments are assessed and approved. This paper examines stakeholders' perceptions of the effectiveness of the IT project clearance process, implemented by Nigeria's National Information Technology Development Agency (NITDA) as a risk-based, stage-gated centralised digital investment governance mechanism. The study adopted a descriptive survey research design approach. The design was suitable for this study because the research focused on evaluating the perceived impact of the clearance framework on compliance across Ministries, Departments, and Agencies (MDAs), local content promotion, reduction in duplication and waste, standardisation and interoperability, transparency in IT projects, and, more importantly, trust in public investments. The study employed a quantitative research approach, which facilitated the objective measurement of respondents' views using numerical data and statistical analysis. The findings from this study indicate a very strong positive perception of the effectiveness of NITDA's IT Projects Clearance process among stakeholders. Using a 5-point Likert scale, all evaluated attributes recorded mean scores above 4.40, suggesting widespread agreement that the clearance framework contributes significantly to accountability, coordination, transparency, trust and value optimisation in government IT investments. More importantly, the study contributes to digital governance literature by reframing IT project assurance as a foundational element of ethical and trust-based governance systems, offering insights relevant to governments seeking to strengthen oversight of digital investments. The findings are also expected to provide useful insights for policymakers, government institutions, ICT regulators, researchers, and development practitioners seeking to improve digital governance frameworks and optimise public sector technology investments.
Authors - Jose R. Rosas-Bustos, Jesse Van Griensven The, Roydon Andrew Fraser, Sebastian Ratto Valderrama, Nadeem Said, Andy Thanos Abstract - CHSH tests are increasingly used as integrity probes for quantum devices and services: when a verifier observes a value S > 2, it is tempting to treat this as a compact certificate of nonclassical correlations. In operational deployments, however, the meaning of a small violation depends on how accurately measurement settings are implemented and how uncertainty is budgeted. We develop a resolution matched comparison between (i) a finite-precision quantum description based on smeared POVMs and (ii) a coarse-grained Bell-local baseline built from deterministic response functions. Using total-variation distance and witness-level tolerances, we define convergence vicinities where the two descriptions become indistinguishable for a realistic verifier. We then interpret weak violations through relaxed Bell bounds under measurement dependence, highlighting that controller leakage can compromise CHSH-only acceptance rules even when the device itself is locally causal. Angle-scan simulations map where these vicinities arise, and IBM Quantum demonstrations illustrate that hardware noise further narrows the practical margin above S = 2. We conclude with controller-aware guidance for integrity pipelines and DIQKD-oriented deployments, emphasizing diversified tests and explicit uncertainty budgets.
Authors - Waqas Ahmed, Muhammad Khalid, Adil Khan, Gulraiz Khan Abstract - Robust traffic sign recognition being important component in integrated Autonomous Driving Assistance Systems (ADAS), driving safety systems; suffers from many challenges including Occlusions and obstructions in road signs. Occlusions may occur due to natural or man made objects; weather conditions, lightening effects, vegetation, paint sprays, parked vehicles may result in traffic sign misclassifications adding more to endangerment of human life and infrastructure. This research contributed by addressing the occlusion problem by devising an occlusion-aware masked autoencoder Vit-OCC for reconstruction of obstructed traffic sign while utilizing German Traffic Sign Recognition Benchmark (GTSRB). Study encompassed four Vit models; a standard classifier (ViT-CLS), a region-aware Cutout-augmented model (ViT-Cutout), a model initialized via random-mask Masked Auto encoding (ViT-MAE), and a proposed occlusion-aware Masked Autoencoder (ViT-OCC). All the models are evaluated using variable and increasing occlusion levels, comprising of 9 different levels from 0% being lowest occlusion level to highly obstructed 80% threshold. Results indicate different patterns of degradation among the four models, although ViT-MAE achieved highest baseline accuracy of 87.23%, ViT-Cutout exhibited stable performance with the introduction of high occlusion levels. ViT-Cutout ViT-OCC exhibited sharp drop in performance under high occlusion thresholds. This study contributed in a systematic exploration of Vision Transformer robustness across variants under controlled and increasing occlusion thresholds while concluding unique patterns in degradation and stability.
Authors - Almas Temirbekov, Bekdaulet Khudaibergen, Nurlan Temirbekov, Aigul Sagatbekkyzy Abstract - The problem to be solved is to determine how the bend angle of an Lshaped channel affects steady two-dimensional incompressible viscous flow and the associated hydraulic losses. Two channel geometries are considered and compared: a channel with a 45° bend and a channel with a 90° bend. Numerical computations are performed for Reynolds numbers Re = 500, 1000, and 2000. The flow is described by the stationary Navier–Stokes equations and solved using the finite element method. Newton’s method is applied to solve the resulting nonlinear algebraic system. The numerical results are analyzed using velocity and pressure fields, stream-function distributions, pressure drop, loss coefficient, and Euler number. The simulations show that the 90° bend produces a stronger rearrangement of the flow downstream of the corner than the 45° bend. Due to the sharper change in flow direction, the velocity field becomes more distorted, pressure gradients increase, and recirculation zones become more pronounced as the Reynolds number grows. In contrast, the 45° bend provides a smoother flow transition and more regular velocity and pressure distributions. The integral characteristics confirm these observations. For all considered Reynolds numbers, the 90° bend leads to a higher pressure drop than the 45° bend. Although the loss coefficient and Euler number vary with Re, the 90° geometry remains less favorable. These results demonstrate that the bend angle significantly affects both local flow structure and pressure-loss characteristics.
Authors - Moncef Zarrouk, Abdelmajid Bessate, Faissal El Bouanani Abstract - Edge artificial intelligence systems require inference mechanisms that jointly account for predictive accuracy and resource consumption. Compact edge models provide low-cost and low-latency inference, but they may be less reliable on difficult inputs, whereas larger cloud models are more accurate but expensive to invoke for every sample. In this paper, we propose a constrained optimization framework for calibration-aware adaptive edge–cloud inference. The edge model first produces a local prediction and a confidence score. This confidence is post-calibrated by scalar temperature scaling, and a threshold gate then either accepts the local prediction or defers the input to a cloud model. The threshold is selected by solving an empirical constrained optimization problem that maximizes accuracy under an operator-specified of- load budget. We show that temperature scaling preserves the edge-model decision, that the empirical offload rate is monotone in the threshold, and that the finite empirical threshold-selection problem can be solved exactly by enumerating the gate partitions induced by the observed confidence scores, correctly handling the strict deferral boundary. Experiments on Fashion-MNIST show that constrained deferral recovers most of the edge–cloud accuracy gap while offloading only a controlled fraction of inputs. Moreover, an additive-noise stress test shows that calibration substantially reduces confidence miscalibration when the edge model be- comes unreliable, confirming that calibration mainly contributes reliabil- ity, while selective deferral drives the accuracy gain.
Authors - Ussen Marassulov, Orken Mamyrbaev, Gulnur Kazbekova, Aigerim Yerimbetova, Madina Sambetbayeva, Duman Telman Abstract - This paper evaluates fake news detection in Kazakh and Russian media by comparing sparse TF-IDF baselines with multilingual transformer models. The task is formulated as binary text classification with Real (0) and Fake (1) labels. The protocol combines duplicate-aware cleaning, split-leakage verification, label-consistency checks, non-neural baselines, transformer finetuning, and bidirectional cross-lingual testing. After exact title + text deduplication, the corpus contained 38,013 Kazakh records, 37,181 Russian records, and 75,194 bilingual records. Five regimes were evaluated: KZ_only, RU_only, MIX_only, KZ->RU, and RU->KZ. TF-IDF remained highly competitive in in-domain testing, reaching Macro-F1 = 0.9979 on Kazakh, 0.9984 on Russian, and 0.9989 on mixed data. In cross-lingual testing, multilingual transformers showed clearer advantages: mBERT reached Macro-F1 = 0.9833 in KZ->RU, while XLM-R achieved 0.9947 in RU->KZ. The results indicate that lexical baselines are strong when train and test data share the same language setting, whereas multilingual transformers are more reliable when the target language changes.
Authors - Guillermo Francia III, Eman El-Sheikh, Md Abdur Rahman Abstract - Global Navigation Satellite Systems (GNSS) are essential components of modern unmanned aerial vehicle (UAV) operations, providing positioning, navigation, and timing (PNT) services that enable autonomous flight, waypoint navigation, and coordinated mission execution. However, the increasing reliance of UAVs on radio frequency (RF) communications and GNSS signals has exposed these systems to a growing range of cybersecurity threats, including spoofing, jamming, malware infections, distributed denial-of-service (DDoS) attacks, and anomalous network behaviors. This paper investigates the application of deep learning techniques for enhancing RF-based security in GNSS-enabled drone communication networks. A comprehensive drone communication dataset containing 52,585 records and four traffic classes—normal traffic, malware infections, DDoS attacks, and anomalous behavior—was utilized to develop and evaluate a Deep Learning Radio Frequency Security (DL-RFS) model. To address severe class imbalance, a balanced TensorFlow data pipeline incorporating stratified sampling, class-wise dataset generation, and equal-probability sampling was designed. The proposed neural network architecture employs fully connected layers with Rectified Linear Unit (ReLU) activation functions and a softmax output layer optimized using the Adam optimizer. Experimental evaluation conducted on an NVIDIA A100 GPU demonstrated exceptional classification performance, achieving an AUC of 0.999, accuracy of 99.3%, precision of 99.9%, recall of 99.9%, and an F1-score of 1.000. Comparative analysis shows that the proposed DL-RFS model outperforms several state-of-the-art machine learning and deep learning approaches for drone network intrusion detection. The results demonstrate the effectiveness of balanced deep learning pipelines for RF signal security analysis and establish a foundation for future research in GNSS security, RF fingerprinting, and AI-driven cyber defense mechanisms for autonomous systems.
Authors - Danial Hooshyar, Yeongwook Yang, Raija Hamalainen, Tommi Karkkainen Abstract - Recent advances in general intelligence paradigms, particularly large language models (LLMs), have accelerated the adoption of artificial intelligence (AI) in education. However, recent studies show that LLMs often exhibit shallow adaptivity and struggle to reliably model learners’ evolving knowledge over time. Similar to other deep neural networks, their opaque nature and susceptibility to bias and spurious correlations can limit alignment with pedagogical principles, raising concerns regarding transparency, fairness, and trustworthiness, particularly in educational settings classified as high-risk under the EU AI Act. While responsible AI use has received increasing attention, responsible AI development—an essential prerequisite for responsible use—remains comparatively overlooked. This paper presents Estonia’s five-year experience applying hybrid neural-symbolic AI (NSAI) methods in educational contexts to operationalize responsible AI development. By integrating symbolic knowledge directly into neural learning processes, the presented approaches aim to support human-centered, interpretable, and pedagogically grounded AI systems. In contrast to recent trends emphasizing loosely coupled symbolic components around closed large-scale models, the Estonian approach focuses on tighter integration between neural and symbolic representations to improve transparency, learner modeling, and educational alignment. Through five illustrative case studies developed across educational applications ranging from student performance prediction to deep knowledge tracing, the paper demonstrates how hybrid neural-symbolic methods can support explainability, pedagogical robustness, and responsible AI objectives in education. Finally, the paper discusses key lessons learned, remaining challenges, and future opportunities for applying neural-symbolic AI in education.
Authors - Saltanat Adilzhanova, Gulshat Amirkanova, Bauyrzhan Amirkhanov, Dana Sybanova, Anas Salem Abstract - The increasing adoption of large language models (LLMs) by adversarial actors has introduced a critical threat to web application security: AI-mutated malicious payloads - injection attacks automatically rewritten by LLMs to preserve malicious functionality while evading signaturebased detection. Existing intrusion detection approaches, including classical machine learning classifiers and deep learning architectures trained on static historical corpora, do not address this threat class and degrade severely when confronted with LLM-obfuscated variants of known attacks. This paper presents a three-component framework for detecting and explaining AImutated malicious API payloads. First, a novel three-class labelled dataset is constructed by applying a controlled GPT-4o-mini mutation pipeline, spanning five structurally distinct obfuscation strategies, to payloads drawn from three established attack corpora, with validation against a sandboxed DVWA instance. Second, a CodeBERT encoder is fine-tuned for three-class classification distinguishing benign traffic, classically malicious payloads, and AI-mutated payloads, achieving a macro-F1 of 0.9472 and an AUC-ROC of 0.9990 on the held-out test set, with a class-specific F1 of 0.8627 on AI-mutated samples - an improvement of 27.18 points over the strongest baseline. Third, a dual-layer explainability module combining attention visualisation and SHAP-based token attribution is evaluated through a faithfulness deletion test, confirming that both methods identify decision-relevant tokens and revealing distinct detection strategies for classical versus AI-mutated payloads. A controlled ablation study demonstrates that AI-mutated training data is a necessary condition for detecting this class, with class-specific F1 collapsing to zero when such data is withheld. The dataset and model are released publicly to support reproducible research.
Authors - Okose Divine Ofure, Azunka N. Ukala Abstract - Low-speed campus mobility presents a practical use case for compact EVs where high-voltage platforms are costly to procure, operate and maintain. Across campuses, short trips rely on walking, motorcycles, petrol vehicles or informal transport, despite predictable low-speed use. This creates demand for EVs that can be locally assembled and maintained with available parts. Yet much EV literature focuses on high-voltage platforms, leaving limited guidance for micro-EVs. This paper presents a safety-aware 48 V architecture for a single-seater campus micro-EV, integrating lithium-ion storage, BLDC traction, fused distribution, pre-charge sequencing, contactor switching and 12 V auxiliary supply. An analytical model estimates tractive force, power demand, battery current, energy use and state-of-charge under stop-start use. Results show 40 Wh/km, with ranges of 18 km, 21 km and 24 km for 15 Ah, 17.5 Ah and 20 Ah packs. The 20 Ah option lowers peak C-rate to 1.34 C, although 40 km needs 2.0–2.5 kWh or mass and loss reduction.
Authors - Noah Holmdin, Martin Gilje Jaatun Abstract - The increased reliance on IT has resulted in a need for decision support for IT system usage scenarios. This is often done through risk assessment, where events are evaluated based on their likelihood of consequences and the severity of consequences. However, this may be difficult in scenarios when there is limited, uncertain, conflicting, or changing information about the IT system’s behaviour. A complementary approach for this is evaluating the system based on a Decision Maker’s (DM) trust in it. This paper introduces a subjective trust assessment framework for IT usage decisions.
Authors - Kyounghee Cho Abstract - South Korea is widely recognised as one of the world’s leading digital societies and a global model of e-government development. While previous studies have primarily examined Korea’s ICT development and e-government achievements from technological and administrative perspectives, relatively limited attention has been paid to how these developments collectively contributed to broader governance transformation. This paper examines how South Korea’s ICT and e-government development contributed to governance transformation beyond technological modernisation from the 1990s onward. Adopting a qualitative case study approach, the study analyses governmental policy documents, e-government development plans, institutional reports, and relevant academic literature. Particular attention is given to the Korean Information Infrastructure (KII), major e-government initiatives, and the role of state leadership and institutional coordination in shaping digital transformation. The findings suggest that South Korea’s digital transformation was not merely a technological process but a broader governance transformation. ICT development and e-government expansion facilitated administrative integration, strengthened institutional coordination, enhanced governance capacity, and reshaped interactions between government institutions and citizens. The study argues that these outcomes were driven not only by technological innovation but also by long-term strategic planning, state leadership, and institutional capacity. By analysing ICT development and e-government expansion as interconnected dimensions of governance transformation, the paper contributes to a deeper understanding of digital governance and provides an analytical foundation for future comparative research on digital governance and AI-enabled public administration, particularly across different institutional contexts such as South Korea and the United Kingdom.
Authors - Gulshat Amirkhanova, Alikhan Amirkhanov, Gulnur Tyulepberdinova, Bauyrzhan Amirkhanov, Yenlik Faruzkyzy Abstract - Digital twins for manufacturing usually rely on a Manufacturing Execution System (MES) that records what each operation does and when. Many small and medium-sized enterprises (SMEs) lack such systems yet increasingly meter the electricity of individual machines. This paper asks whether a plant's process can be reconstructed, simulated and optimised from electrical metering alone, and formalises the answer as Energy-Driven Process Reconstruction (EDPR). The subject is a commercial bakery in Kazakhstan whose fifteen machines are metered and integrated through OpenEgiz, a digital-twin platform built on the open-source OpenTwins framework; about 65 million readings span 198 days. EDPR detects machine states from active power, abstracts them into events, and links events into per-batch chains by a batch-anchored lead-lag operator with one-to-one assignment in O(N log N) time. With no MES ground truth available, EDPR is scored on a labelled synthetic benchmark, where it reaches an eventdetection F1 of 0.97 and, against three baselines, is the only method that combines competitive chain accuracy with a valid one-to-one batch partition. On the real plant, conformance checking raises model precision from 0.22 for a naive day-case model to 1.00 for the EDPR reconstruction, and a discrete-event twin parameterised only by the energy-derived lead times reproduces the observed throughput of about 21 batches per day. Using the twin, three demand-side measures are estimated to give a potential electricity-cost reduction of 20.5 per cent at constant output. The pipeline uses only open-source software and permachine power.
Authors - Ekene Michael Ogbejesi, Ming Jiang Abstract - The rapid growth of e-commerce and smart city infrastructures has intensified the operational challenges associated with last-mile delivery systems. Last-mile logistics remains the most expensive and operationally complex component of modern supply chains due to traffic congestion, inefficient routing, fluctuating environmental conditions, and increasing customer delivery expectations. This study presents a scalable data fusion and machine learning platform designed to optimize smart city last-mile delivery operations through predictive analytics and route optimization techniques. The proposed framework integrates heterogeneous datasets including geospatial coordinates, delivery agent profiles, weather conditions, traffic density, and vehicle characteristics to improve delivery efficiency and operational decision-making. A comprehensive preprocessing and feature engineering pipeline was developed, incorporating Haversine distance computation and categorical feature transformation. Multiple machine learning models including XGBoost, LightGBM, Random Forest Regression, and Neural Networks were evaluated for delivery time prediction using historical logistics data. The dataset comprised 1,000 delivery records with features including distance, traffic conditions, vehicle type, and time-based variables. Data preprocessing involved handling missing values, outlier treatment, and feature scaling. Experimental evaluation demonstrated that the Random Forest model achieved the strongest predictive performance with a Root Mean Square Error (RMSE) of 25.15 minutes, Mean Absolute Error (MAE) of 19.28 minutes, and an R² score of 0.76. The predictive model was integrated with a Traveling Salesman Problem (TSP)-based optimization module using a nearest-neighbor heuristic algorithm to generate efficient multi-stop delivery routes. The framework was deployed as an interactive Streamlit application supported by Folium geospatial visualization. Results demonstrate that the proposed framework improves delivery prediction accuracy, route planning efficiency, and operational scalability within intelligent transportation systems.
Authors - Tlangelani Promise Mlambo, Tranos Zuva, Andrew Brown, Ramadile Moletsane Abstract - South African organizations are experiencing rapid digital transformation as digital technologies increasingly shape service delivery, data management, and interactions between the state and citizens. In this time where machines such as Generative Artificial Intelligence (GenAI) can generate answers from the user prompt, there’s a need to control their use in order not to harm others. This study examines digital governance in the generative AI era, focusing on how privacy and trust are managed within organizations in South Africa. The study adapts a qualitative research approach utilizing the traditional literature review and secondary data analysis. There’s a disconnect between the initiated policies and the implementation or regulation of these policies, therefore, to develop a comprehensive digital governance framework for South African organizations to help minimize any harm that can be caused by use of Generative Artificial Intelligence. The research shows that in the generative AI era, is necessary for digital governance to move from only technology-centric strategies to a comprehensive framework that will combine institutional, ethical and operational aspects. The study contributes theoretically and practically to a deeper understanding of governance challenges and opportunities in South Africa. The study is limited to South African organizations and on secondary data from existing literature. Future work can extend the scope of the study beyond South Africa and conduct empirical and longitudinal studies to gain more insights on digital governance in the generative AI era.
Authors - Oussama H Hamid, Ayman Ahmed, Arif Al-Nahdi Abstract - The increasing reliance on web-based services has intensified the need for secure and reliable authentication mechanisms. Facial recognition has emerged as a widely adopted biometric modality owing to its convenience and contactless operation. However, static facial recognition systems remain vulnerable to presentation attacks such as printed-photo spoofing, screen replay, and mask-based impersonation. Existing face anti-spoofing techniques typically implement liveness detection as a preliminary filtering stage rather than as an integral component of the authentication decision process, creating a structural security gap that this paper addresses. A threat-model-driven architectural framework is proposed that integrates motion-based behavioural verification as a second authentication factor within facial recognition systems deployed in web-based environments. The framework introduces a system-generated challenge–response mechanism in which users perform randomised facial actions; facial landmark tracking and temporal motion analysis verify challenge execution in real time. A formal threat model distinguishes remote attackers, limited physical attackers, and generative adversarial attackers, and maps each adversary class to specific architectural countermeasures. A structured security analysis evaluates the framework against six presentation attack scenarios, including deepfake-based adaptive attacks. The proposed design operates on commodity hardware without specialised sensors, and the paper discusses biometric template protection, client–server deployment models, and privacy compliance in detail. This work contributes a principled architectural foundation for multi-factor biometric authentication in web environments and identifies concrete directions for future empirical validation
Authors - Andrea Piroddi, Maurizio Torregiani Abstract - This paper introduces an information-theoretic framework for bounding the mutual information (MI) between the two-dimensional physical position of a user and the observed multiple-input multiple-output (MIMO) channel state information (CSI). Unlike classical Cram´er–Rao bounds or I-MMSE relations, the proposed approach explicitly incorporates the spatial variability of the channel via its Jacobian with respect to position. Two complementary results are derived: a local Jacobian-based characterization, valid under small prior spatial uncertainty with explicit validity conditions, and a global nonlinear upper bound via path-integral formulations that extends the analysis to arbitrary spatial distributions. Both are derived from first principles under clearly stated assumptions, and their relationship to classical Cram´er–Rao bounds, Fisher information, and Ziv–Zakai bounds is discussed. A quantitative tightness analysis across array sizes M ∈ {8, 16, 32, 64} shows that the local bound is nearly exact for small arrays and low SNR, while becoming progressively more conservative for larger configurations. A sensitivity analysis of the global bound to the reference point choice is provided, together with a principled method to reduce this sensitivity via the channel manifold centroid. Validation using Sionna RT ray tracing on the Munich urban scene reveals that realistic channels exhibit lower spatial variability than synthetic line-of-sight models, with direct implications for CSI-based localization design.
Authors - Gulshat Amirkhanova, Alikhan Amirkhanov Abstract - Wearable and Internet-of-Medical-Things (IoMT) devices now stream cardiac signals continuously, yet turning that flow into an early warning still needs a model that says when an adverse event is likely, not just whether one will occur. We treat that question as a time-toevent problem and benchmark five survival models on the UCI Heart Failure Clinical Records cohort (299 patients, 96 deaths, follow-up 4– 285 days): Cox proportional hazards, penalised Cox, random survival forest (RSF), gradient-boosted survival analysis, and the deep survival network DeepSurv. Models are compared under 5×5 repeated stratified cross-validation with Harrell’s and IPCW concordance, the integrated Brier score, and time-dependent AUC. The two tree-ensemble survival models lead: gradient boosting reaches a C-index of 0.731 and RSF 0.725, both ahead of Cox (0.709) and well ahead of DeepSurv (0.645), which also shows the worst calibration. On a cohort of this size the deep model does not pay its way. Permutation and SHAP analysis of the RSF point to serum creatinine, ejection fraction and age as the dominant risk drivers, which agrees with established cardiology. We frame the survival learner as the analytics stage of an IoMT cardiovascular-monitoring pipeline and release code, data split and seeds for full reproducibility.
Authors - Aidana Karibayeva, Oleg Myssov Abstract - This paper presents a systematic experimental evaluation of the SeamlessM4T speech-to-text translation system on four Turkic languages — Kazakh, Tatar, Turkish, and Uzbek — across nine language pairs: kaz-tur, kaz-uzb, tat-kaz, tat-tur, tat-uzb, tur-kaz, tur-uzb, uzb-kaz, and uzb-tur. The study compared the base model (v1) and its fine-tuned version (v2). A closed parallel corpus of 1000 sentences per language pair was used. For evaluation, the BLEU, chrF, and WER metrics were used. The results showed a moderate but uneven improvement in scores after fine-tuning. The best results were observed in pairs where Uzbek was used as the source language (uzb–kaz: +0.78 BLEU; uzb–tur: +0.95 BLEU). However, in some Turkic language pairs, the scores decreased, which may be related to the model forgetting its prior knowledge during multi-task learning. When Tatar was the source language, both models performed poorly. This is explained by the scarcity of data and the phonological differences between languages. Furthermore, it was found that the model's architecture is limited when Tatar is used as the target language. The chrF metric was found to be more effective than BLEU for morphologically complex agglutinative languages. Therefore, it is recommended for use as the primary metric in the S2TT evaluation for Turkic languages.
Authors - Zubaria Inayat, Hanin Esawi, Maya Daneva, Marten van Sinderen, Giancarlo Guizzardi, Luiz Bonino da Silva Santos Abstract - Explainable artificial intelligence (XAI) is becoming an important component of healthcare systems, supporting transparent and trustworthy AI-assisted decision-making. However, existing explainable AI approaches are mainly designed around healthcare professionals, while patient needs and expectations regarding AI-generated explanations remain insufficiently explored. This study investigates how patients perceive the quality of explanations provided by XAI healthcare systems and identifies the challenges that influence their understanding, trust, and engagement. A mixed-method research approach was adopted, combining (i) a rapid literature review, (ii) an online survey with 33 participants, and (iii) expert consultation for validation. The literature review identified nine quality dimensions for patient-oriented XAI explanations. The survey findings revealed eight challenges experienced by patients when interacting with explainable AI systems. Based on the synthesis of these findings, a patient-centred evaluation matrix was developed, linking explanation quality dimensions with patient-related challenges. The proposed matrix was validated through expert feedback. The results highlight the importance of moving beyond developer and clinician-centric XAI design towards patient-centred explainable healthcare systems. This work contributes to trustworthy artificial intelligence in healthcare by providing guidance for evaluating explanation quality, improving usability, supporting patient trust, and enabling informed shared decision-making.
Authors - Harsh Sudhir Shah, Ozlem Mis, Bosen Zhou, Christian Zimmer, Frank Feinbube Abstract - Immersive digital twins use virtual reality, augmented reality, and Internet of Things technologies to create interactive virtual replicas of physical assets and environments that users can enter and explore. Across Europe, experts report only limited progress toward immersive digital twins that are scalable, sovereign, and AI-augmented. The literature does not yet offer a stakeholder-grounded account of the barriers. This paper reports a reflexive thematic analysis of 11 semistructured interviews with experts. The sample spans industrial manufacturing, automotive, telecommunications, public infrastructure, 2 platform or studio vendors, a research institute, a technology consultancy, 2 universities, and an industry association. The transcripts were coded collaboratively by 3 researchers who reached consensus through structured discussion. The qualitative analysis is corroborated by a Likert layer reported alongside the themes. The analysis identifies 8 master barrier themes. Interoperability across data formats and standards is the central constraint. Content creation is the dominant scaling bottleneck. The remaining themes cover hardware and platform volatility, the absence of a shared immersive user experience grammar, the role of artificial intelligence inside the runtime, trust and explainability limits, open source as a sovereignty question, and the demand for ready-to-run workflow scaffolding. A single-source observation concerns the erosion of opensource community review quality by AI-generated commits. The themes are translated into 7 candidate open-source building blocks that would relieve the common barriers. The contribution is empirical and forward looking. It is intended to guide open-source contribution priorities in the European immersive digital twin community.
Authors - Ahmed Abdelmoula, Balint Molnar Abstract - Honey fraud, including adulteration, origin mislabeling, and unsuitable thermal handling, remains difficult to control because analytical verification and digital traceability are rarely integrated. Laboratory methods can assess honey quality and authenticity, while blockchain systems can protect traceability records; however, ledger immutability does not ensure that recorded evidence is scientifically valid. This paper introduces a preliminary framework for converting honey-quality metrics into states suitable for permissioned blockchain environments. The architecture combines IoT sensing, edge-level validation, metric computation, offchain storage, cryptographic referencing, and rule-based smart-contract logic. The main contribution is a metric-to-state mapping mechanism that encodes environmental, biochemical, and process-related evidence as blockchain events, hashes, quality indicators, and certification states. A simulation-based prototype evaluates a 24-hour hive-monitoring scenario using synthetic temperature and humidity readings collected every 15 minutes. Of 96 generated readings, 93 passed validation and were aggregated into 24 hourly quality-state transactions, reducing potential ledger entries by 75% while preserving SHA–256 hash-chain integrity. The results provide initial feasibility evidence for the proposed sensor-to-ledger pipeline, while remaining limited to a proof-of-concept setting rather than a full industrial deployment or permissioned blockchain benchmark. Future work will incorporate real hive data, laboratory measurements, and complete permissioned blockchain implementation.
Authors - Malgorzata Pankowska, Michal Dimmich, Radosław Pacud, Paweł Lula Abstract - In this study, the authors provide the text anonymization model and present a statistical analysis of the anonymization results. The purpose of this study is to: ensure compliance of data processing with General Data Protection Regulation (GDPR) in the activities of a law firm - removing direct and indirect personal data from documents used for further work of a lawyer as templates and know-how; ensure the protection of professional secrets in the activities of a law firm - removing all identifying data and elements covered by professional secrecy within documents used for further work of a lawyer as templates and know-how, i.e., leaving the names and surnames of the authors of scientific publications that confirm a given view; ensure the protection of personal and identification data when sending documents. The authors concluded that the Long Short Term Memory (LSTM) model application provided high performance results for legal document anonymization
Authors - More Ickson Manda Abstract - The rapid advancement of digital technologies and artificial intelligence has created significant opportunities and challenges for developing countries such as South Africa. To harness these opportunities and mitigate associated risks, South Africa must adopt strategic policy and investment interventions that promote innovation, strengthen digital capabilities, and support inclusive and sustainable socio-economic growth. Technological developments such as Artificial Intelligence, robotics, internet of things and cloud-based technologies have disrupted every sector. Concerns around ethics, sovereignty, security, privacy, skills, affordability, governance and infrastructure have hindered developing countries from fully leveraging the benefits of AI. The purpose of this paper is therefore to identify key pillars for building an inclusive digital society in South Africa in the age of Artificial Intelligence where citizens can fully access and benefit from using digital services. The study found that developing an inclusive digital society, digital leadership, digital infrastructure development, embracing new and emerging technologies, digital policy and governance and digital resilience are key pillars of an inclusive digital society.
Authors - Andrei-Horia Ignat, Dan-Matei Popovici Abstract - Large language models solve tasks across many domains, but their size and reliance on hosted providers place them out of reach of organisations that must own their deployment. We present a multi-model, multi-adapter framework for building and serving a conversational multiexpert search platform entirely on commodity hardware. Each capability is realised as a small, task-specialised expert: a parameter-efficient (QLoRA) adapter over one of a few shared base models, trained on data synthesised offline by a large model rather than collected from users. A lightweight router directs each query to the relevant expert, which answers by emitting an executable tool call against the service’s data. Because every expert is a few-megabyte adapter over a shared base, the dominant memory cost is the base models, not the number of experts: a single workstation GPU holds an estimated several hundred to several thousand experts, and adding a service requires only a short, repeatable recipe rather than an advanced MLOps stack. We organise the design around eight requirements spanning commodity-hardware operation, extensibility, accessibility, multilingual access, data-grounded answers, transparency, self-hosting, and graceful integration depth, and we describe serving optimisations (multi-adapter co-residency, prefix caching, function-call routing) that keep it efficient. We instantiate the framework on the FUTURAL Metasearch Platform, the smart-solution registry of an EU agricultural project; a representative wildlife expert reaches about 95.1% token-level accuracy after roughly twenty minutes of training on one GPU.
Authors - AA Adebomehin, FJ Ibrahim, AS Dahiru, TK Akinduyite, TO Obinna-Esiowu, I Ofodile, OA Odeyemi, T Salman Abstract - This paper presents novel AI-enabled Kalman filter techniques that enhance positioning accuracy in inertial navigation systems (INS). Essentially, accurate positioning remains a core challenge in multi-sensor INS; since performance of traditional Kalman filtering degrades in nonlinear environments. Our method significantly improves the precision of INS by integrating adaptive AI-based machine-learning Kalman filters with classical estimation theory. The approach focuses on multi-sensor fusion, adaptive noise modeling, and robust improvements for INS applications. Simulation results demonstrate the effectiveness of the proposed techniques, especially airborne INS. This is significant in that achievement of precision without sacrificing overall accuracy is essential to sensor data in view of factors like sensor thermal noise, lower inference quantization, and tolerance which could affect real–world performances. Additionally, in a multi–sensor fusion setting, effectiveness of INS hinges on crucial and dependable filter systems for real data integration, noise reduction, reliable predictive analysis, and real-time processing. Consequently, this research developed AI-based models and utilized them for all filter GPS positioning data to update the INS states covering INS position output, INS internal states (position, velocity and heading), and finally for estimation of bias on the accelerometer sensor. It is believed that the approach has possible applications in diverse fields like defense & security with favorable implications for autonomous systems; as well as surveys & disaster management. Key highlights of the improved multi-sensor INS algorithm models are presented in this paper.
Authors - Mahnoor Raghib, Muhammad Usman Noor, Muhammad Usman, Ahmedullah Mohammed, Bushra Naeem, Shovro Baidya, Farae Naeem Abstract - The fashion industry and fashion education sector are undergoing signif- icant transformation through the adoption of Artificial Intelligence (AI) and advanced cloud computing technologies. While cloud computing provides scalable infrastructure, real-time collaboration, and data processing capabilities, AI introduces intelligent au- tomation, predictive analytics, generative design, and personalized decision-making. In fashion education, AI-powered tools enhance creativity, virtual prototyping, digital de- sign capabilities, and industry readiness by enabling students to experiment with inno- vative concepts in simulated environments. Simultaneously, cloud-based platforms fa- cilitate collaborative learning, scalable storage, and access to professional design and simulation software. Within the fashion industry, cloud computing supports digital transformation across supply chain management, manufacturing operations, customer experience, and sustainability initiatives. AI-driven analytics further enable demand forecasting, inventory optimization, waste reduction, and customer personalization. However, challenges relating to cybersecurity, privacy protection, algorithmic bias, transparency, organizational resistance, and legacy system integration continue to hin- der implementation. This paper synthesizes current literature and proposes an AI– Cloud Fashion Transformation Framework that conceptualizes cloud computing as the infrastructure layer and AI as the intelligence layer supporting transformation across educational and industrial domains. The findings suggest that successful digital trans- formation requires not only technological adoption but also effective governance, workforce development, ethical implementation, and organizational commitment. The proposed framework contributes to understanding how AI and cloud computing collec- tively drive innovation, sustainability, and operational excellence within fashion eco- systems.
Authors - Caitlin Loh, Susmitha Vekkot, Pancham Shukla Abstract - Deepfakes generated using modern machine learning techniques pose growing risks to digital trust by enabling realistic manipulation of both audio and video content. Many existing detection approaches rely on a single modality, limiting robustness when confronted with increasingly sophisticated forgeries. This paper evaluates a multimodal deepfake detection framework that integrates audio and visual information using deep learning. The proposed system combines stateof-the-art audio and visual encoders within a modular architecture and conducts a systematic comparison of three fusion strategies: early fusion, late fusion, and cross-attention. Experiments are conducted on the PolyGlotFake dataset, a multilingual benchmark containing synthetic and authentic audio–visual media. Results show that multimodal approaches substantially outperform unimodal baselines, with late fusion achieving an AUROC of 0.955 and cross-attention models reaching accuracies of up to 0.996. These findings provide a controlled comparison of fusion strategies and demonstrate that multimodal fusion significantly improves detection performance and highlights its potential for building more robust deepfake detection systems.
Authors - Bayangali Abdygalym, Madina Sambetbayeva, Aigerim Yerimbetova, Elmira Daiyrbayeva, Duman Telman Abstract - Dense retrievers achieve strong results on lexical synonymy but degrade on hierarchical (BT/NT) and associative (RT) relations in specialized multilingual thesauri. This paper addresses typed thesaurus relation extraction by introducing a sparse–neural pipeline that combines a domain-tokenized BM25 index with morphological prefix tokens for agglutinative Kazakh and inflectional Russian, one-hop graph expansion over partially observed thesaurus structure, a KazBERT relation classifier over six labels (BT, NT, RT, SYN, LE, NONE) trained with hard-negative mining over taxonomic siblings, and an LE-mediated synonym closure that propagates synonymy through translation edges. The pipeline is evaluated on a curated trilingual military thesaurus of 5,262 canonical entries and 64,374 manually validated relation edges, using a 15% term-level holdout (711 query terms). The proposed pipeline reaches BT Hit@5 = 0.796, NT Hit@5 = 0.602, RT Hit@5 = 0.726, and SYN Hit@5 = 1.000. The results indicate that sparse lexical retrieval remains highly competitive for hierarchical thesaurus relation discovery in low-resource multilingual settings.
Authors - Muhammad Sohaib Ayub, Sikandar Bakhat, Mian Muhammad Awais, Umair ul Hassan Abstract - Smallholder farmers in Pakistan make irrigation, fertiliser, and crop-selection decisions largely on inherited practice rather than field-specific evidence, even though the satellite data needed to inform those decisions is now free and globally available. This paper presents the Agriculture Portal, a role-based web platform that turns raw Sentinel2 imagery into decision-ready guidance for farmers and the extension officers who support them. The system computes NDVI, EVI, NDMI, NDWI, and a radar vegetation proxy from Sentinel-2 surface reflectance through Google Earth Engine, complements them with ground-deployed IoT sensors and an AI land classifier, then fuses the signal through an ensemble of the Analytic Hierarchy Process, fuzzy TOPSIS, and VIKOR into two scores: a Crop Health Index describing vegetation vigour and a Soil Suitability Index describing cultivation potential. A fixed agronomic normalisation scheme prevents the degenerate behaviour that field-relative normalisation produces over spatially uniform plots. Beyond the satellite layer, extension officers can author standardised crop advisories and farmers can log field activities against them, with the portal automatically generating a compliance report that surfaces the gap between guidance and practice. Field trials across cultivated and forested sites in Pakistan confirm that the dual-score design correctly separates vegetation vigour from cultivation suitability and that the compliance tracker correctly reconciles recommended and logged activity.
Authors - Mussa Turdalyuly, Aigerim Yerimbetova, Bakzhan Sakenov, Ulmeken Berzhanova, Orken Mamyrbaev, Duman Telman Abstract - In this paper, we examine the effectiveness of word- and phonemelevel automatic speech recognition (ASR) models for translating speech into sign language in the presence of limited language resources. A speech-to-sign system requires not only accurate transcription but also reliable preservation of semantic information for the correct selection of gestures. We focus on the Kazakh language and compare two ASR approaches based on the XLS-R architecture, namely, a word-level model and a phoneme-level model. Both models were studied and evaluated using the same Kazakh voice dataset and experimental settings. Its performance was assessed using traditional recognition metrics such as word error rate (WER) and phoneme error rate (PER), as well as gesture token accuracy (GTA), a follow-up evaluation index that measures the accuracy of gesture selection. The word-level ASR model achieved the best recognition performance with a WER of 0.360, while the phoneme-level model achieved a PER of 0.573. The phoneme-level model demonstrated excellent robustness in follow-up tasks despite low transcription accuracy and achieved high gesture identification accuracy of 0.673 versus 0.686 for the word-level model. These results suggest that phoneme-based expressions are suitable for supporting communication systems where preserving meaning is more important than accurate transcription. This study highlights the importance of evaluating speech recognition systems using the following target metrics in addition to traditional recognition metrics.
Authors: Johnson Masinde, Mugambi Frankline, Daniel Wambiri Abstract: Over time, the field of librarianship has had to adapt due to factors such as technological advancements, which have significantly impacted the library landscape. This evolution has prompted librarians to reconsider their roles and responsibilities, as well as the future and long-term sustainability of conventional library practices and services. The recent past has seen a significant intensification of the debate surrounding the role and future of academic libraries, driven by the rapid advancements in technology. Traditionally, academic libraries have served as core pillars of university education, promoting learning, teaching, and research. However, institutions of higher learning are transitioning from traditional in-person teaching methods to online methods of instruction. Furthermore, an increasing number of students are opting for online modes of instruction instead of the traditional in-person classroom. This study was motivated by the need to explore the changing dynamics of academic libraries, influenced by the opportunities and challenges present in contemporary society. The research study conducted a foresight analysis of the potential characteristics at the convergence of academic libraries and the fourth industrial revolution (4IR) to uncover the significant trends and uncertainties that might not be immediately visible. Furthermore, it utilized both qualitative and quantitative research methods, providing a thorough analysis of the potential future of the academic library landscape. The Shawaz scenario planning process was utilized to assess the key drivers of change in academic libraries over time enabling the study to paint a picture of the plausible features with the study structured in three distinct phases: (i) the literature review to uncover the mega trends and uncertainties, (ii) the delphi survey involving 33 participants (iii) and the formulation of the scenario. Study findings from the first phase show a complex network of relationships among the major trends and uncertainties influencing the academic library landscape in both the short term and long-term future. In addition, the findings indicate that economic, technological, and political factors significantly influence the academic library landscape. Academic libraries are evolving into a more dynamic landscape adopting new roles and responsibilities such as digital archiving, data analysis support, and research data management services. The findings show that while earlier studies suggested that technological factors shape the academic library landscape, it is evident that economic and political factors also play significant roles in influencing this environment.
