Generative artificial intelligence (GenAI) is reshaping higher education, yet its educational value, governance, inclusion and responsible implementation remain contested. This structured integrative literature review used a PRISMA-informed search and selection process and thematic synthesis to examine pedagogical opportunities and risks, student–staff partnership, assessment and academic integrity, institutional governance, privacy, accessibility and digital equity, with particular attention to UK higher education. The detailed synthesis drew on a 27-report focused corpus formed through an investigator-dependent purposive selection process from 1011 broadly eligible reports; it is therefore neither exhaustive nor statistically or descriptively representative of the broad eligible register. The evidence suggests that GenAI may support explanation, writing, planning, language assistance and flexible academic engagement, but these benefits remain conditional on disciplinary context, user capability, verification and continued human support. The evidence base is stronger for reported use, perceptions and implementation concerns than for causal improvements in learning outcomes. Evidence of student–staff partnership was clearest where participation influenced defining curriculum, assessment or policy decisions; consultation alone should not be presented as co-creation. Responsible adoption also requires programme-level assessment review, clear permitted-use guidance, accessible alternatives, equitable access, data protection, meaningful human oversight and institutional accountability. Drawing these findings together, the review develops an evidence-informed TPACK–Responsible AI–Sociotechnical framework based on pedagogical and capability alignment, Responsible AI safeguards and sociotechnical readiness. The framework supports transparent decisions to approve, approve with controls, pilot with monitoring, redesign or reject proposed uses. Responsible GenAI integration should therefore be use-specific, participatory, proportionate and subject to continued review.
Retrieval-augmented generation (RAG) systems for scientific literature require evidence-based choices of document segmentation, representation, retrieval, and generation components, particularly when the source collection varies in topical specificity and document structure. This study addresses the lack of an end-to-end, component-level comparison of these choices for soil science question answering. A three-stage evaluation was conducted across general, domain-specific, and geospatial soil science corpora. The corpus combines foundational soil science books, peer-reviewed research articles, European soil monitoring material, and geospatial mapping publications, thereby covering both broad disciplinary concepts and specialized scientific evidence. The study compares four chunking strategies, three embedding models, five retrieval methods, and five large language models. In Experiment 1, semantic chunking with text-embedding-3-large achieved the highest aggregate retrieval scores (recall@1 = 0.824; MRR = 0.819), whereas text-embedding-3-small delivered practically comparable performance at lower cost. In Experiment 2, hybrid reciprocal rank fusion achieved recall@5 values of 0.957, 0.960, and 0.647 for the general, domain-specific, and geospatial corpora, respectively; the cross-encoder reranker showed weaker rank quality on scientific content. In Experiment 3, model responses attained BERTScore values of 0.909–0.927 and faithfulness of at least 0.993; these automated measures indicate low contradiction with retrieved context but do not establish answer completeness or human-perceived correctness. The study provides a reproducible component-level evaluation design, characterizes the effect of corpus specificity on RAG retrieval, and identifies a practical configuration for soil science literature retrieval. Among the models retained for direct aggregate comparison, Llama 3.1 8B offered the most favorable observed balance of answer quality, latency, cost, and model openness.
Libraries are increasingly recognized as strategic institutions that support innovation, creativity, and knowledge-based economic development. This study presents a narrative and conceptual review of scholarship on the role of libraries in supporting innovation and creative industries. Using a purposive rather than exhaustive search, the review examines a corpus of 26 peer-reviewed studies published between 2010 and 2025. Across the reviewed studies, libraries are represented as increasingly active knowledge institutions supporting innovation through technological infrastructure, digital services, makerspaces, collaborative environments, entrepreneurship support, and community engagement. Within this purposively selected corpus, public-library contexts and qualitative or case-based approaches are substantially represented, while fewer included studies address academic-library contributions to entrepreneurship, research commercialization, and knowledge transfer. These patterns describe the reviewed set and are not intended as prevalence estimates for the wider literature. The principal contribution is a structured synthesis and organization of existing knowledge: the review identifies recurring themes, clarifies relationships among previously separated strands of scholarship, highlights unresolved questions, and develops a future research agenda. An integrative library-specific framework is offered as a conceptual organizing device that brings these strands into a common input–process–outcome structure; it is not presented as a novel or empirically validated theory.
The notion of “AI natives” is increasingly used to describe young learners’ presumed ease with generative artificial intelligence (GenAI), but the label often turns age and exposure into proxies for competence. This study re-examines that assumption through an exploratory secondary analysis of two open student datasets and proposes a practice-based relational framework for AI nativeness. In this framework, AI nativeness is not a generational identity, but a testable configuration of competence foundations, GenAI use practices, access conditions, and institutional mediation. Because the two datasets do not measure all four dimensions within the same learners, the framework is motivated rather than fully tested here. The results show that basic operational skills do not, on their own, explain AI readiness; critical information literacy is the most consistent positive predictor in the regression models. Student GenAI use and access conditions are also heterogeneous, forming four interpretable GenAI use/access profiles: low-adoption learners, high-intensity multi-taskers, mobile-dependent moderate users, and balanced cognitive adopters. These findings do not establish who is or is not an AI native. They show why the age-based label is analytically insufficient and why future research should examine AI nativeness through competence, practice, access, and institutional mediation together.
Digital technologies, particularly artificial intelligence (AI) and big data analytics (BDA), are increasingly recognized as drivers of competitive advantage, yet the mechanisms through which they enhance organizational performance remain insufficiently understood. Grounded in Teece’s dynamic capabilities framework, this scoping review examines how digital technologies enable sensing, seizing, and reconfiguring capabilities and their impact on organizational performance. Following PRISMA-ScR guidelines, 1349 records from Scopus and Web of Science were screened, resulting in 21 eligible studies (2017–2025). A thematic synthesis and a complementary random-effects meta-analysis of 15 quantitative studies (N = 6924) were conducted, representing an integrated scoping review with embedded quantitative synthesis. The findings show that AI and BDA strengthen sensing through advanced data analytics, AI and digital platforms enhance seizing via real-time decision support, and IoT and blockchain facilitate reconfiguring by improving process optimization and organizational flexibility. Dynamic capabilities consistently mediate the relationship between digital technologies and performance. The meta-analysis confirms a significant positive overall effect (β = 0.356, 95% CI: 0.258–0.454), remaining robust after publication bias correction (β = 0.338). This review further identifies three generative tensions, breadth–depth, speed–deliberation, and flexibility–rigidity, that underpin capability development. An integrative framework and future research agenda are proposed, extending dynamic capability theory by conceptualizing digital technologies as active enablers of organizational adaptation and sustained performance.
Agile project monitoring commonly relies on numerical metrics and visual artifacts, such as Burndown charts, to assess Sprint progress and identify potential deviations. However, most automated approaches focus on structured data, while the visual patterns contained in agile charts remain underexplored. This paper presents an exploratory study on visual-based Sprint evaluation using convolutional neural networks and agile project metrics. The proposed approach uses Burndown and Completed vs. Uncompleted Work chart (TTvsNT) images to classify Sprint performance into four categories: Poor, Regular, Good, and Excellent. A transfer-learning strategy based on MobileNetV2 was applied, including image preprocessing, Sprint-level data partitioning, two-phase training, and multiclass evaluation. The model achieved an overall accuracy of 70.33% on the evaluation set. Class-level results showed better performance for the Poor and Excellent categories, while the intermediate classes presented greater ambiguity. The main contribution of this study lies in evaluating Sprint monitoring charts as a complementary visual representation to traditional metric-based models. The findings provide preliminary evidence that these images contain useful performance-related patterns; however, the limited dataset size and current accuracy do not support production-level deployment. Further research with larger datasets, additional architectures, and multimodal approaches is required.
Speaker recognition models are typically enrolled using neutral speech, yet real users rarely speak under emotionally neutral conditions. Emotion alters prosody, spectral structure, articulation, and speaking rate, shifting utterances away from the acoustic distribution observed during enrolment. Conventional augmentation with noise or speed variation introduces generic acoustic variability but does not explicitly model these coordinated, emotion-specific changes. This study therefore investigates whether synthetic emotional speech representations can improve speaker recognition when authentic emotional enrolment recordings are unavailable. Three conditional generative adversarial networks translate neutral mel spectrograms into angry, happy, and sad variants. The translation models are trained using lexically parallel neutral and emotional recordings from ESD and SAVEE. The generated spectrograms are subsequently added to neutral training data, while evaluation is conducted exclusively on genuine recordings from the independent RAVDESS and CREMA-D corpora. This protocol assesses whether learned emotional transformations transfer across speakers, lexical content, and recording conditions. Combined emotional augmentation improves speaker identification and verification on both corpora, with a larger effect on RAVDESS. Identification accuracy increases from 65.00% to 71.83% on RAVDESS and, by a smaller margin, from 56.04% to 60.20% on CREMA-D. The experiments further indicate that augmentation effectiveness depends on the target emotion and corpus, and that single-emotion augmentation can degrade recognition on some corpora.
AI-driven smart healthcare platforms increasingly combine multimodal data acquisition, remote monitoring, predictive analytics, longitudinal dashboards, and role-specific information services. While these capabilities enable more continuous and personalized health monitoring, they also introduce accessibility challenges that extend beyond conventional user-interface design, particularly in how monitoring results and AI-generated outputs are structured, contextualized, and communicated. This paper addresses the gap in operationalizing accessibility at the system level by proposing a standards-mapped Accessibility-by-Design framework for AI-enabled smart healthcare platforms. Requirements derived from international accessibility, human-centred design, and software quality standards are translated into traceable architectural and component-level constraints and integrated into the development workflow. The methodology combines standards-to-requirement-to-component traceability mapping, architectural documentation analysis, component inspection, and controlled workflow walkthroughs. The framework is applied to the NeuroPredict platform, an AI-based environment for longitudinal and multimodal monitoring of neurodegenerative disorders. Accessibility is operationalized through cross-layer architectural mechanisms, including semantic interface components, predictable interaction workflows, accessible presentation mechanisms, and structured representation of monitoring and AI outputs. AI-generated results are presented in a structured form together with contextual, temporal, source-related, and explanatory information to support role adaptation; uncertainty-related fields are reserved for future validated user-facing integration. Implementation evidence from selected platform workflows demonstrates that the proposed mechanisms can be incorporated into the current platform architecture and provides preliminary support for their technical feasibility. The evidence remains limited to implemented and inspected mechanisms, without user-based accessibility validation, formal conformance auditing, clinical deployment, or regulatory certification. These results illustrate that, while maintaining the distinction between accessibility mechanisms and the underlying analytical computation, accessibility-by-design can offer an engineering foundation for integrating accessibility requirements into the architecture and development lifecycle of AI-driven smart healthcare platforms.
Business Process Analytics (BPA) has become increasingly important for improving organizational processes and supporting data-driven decision-making. However, existing Business Process Management and Data Analytics methodologies provide limited support for stakeholder participation, User-Centered Design (UCD), and collaborative implementation, creating barriers for organizations with limited analytical maturity. This study presents MIDA5, an initial methodological framework that integrates BPA, Data Analytics, Business Process Management, UCD, and gamification into a participative implementation methodology. Developed following a Design Science Research approach, MIDA5 comprises five phases, 14 stages, 33 activities, and 61 methodological artifacts. The framework was evaluated through a 3-month-and-12-day organizational case study conducted at a public university involving four core organizational participants, seven organizational stakeholders, and complementary organizational applications. The implementation formalized an undocumented process, integrated heterogeneous data sources, developed an analytical database and ETL workflow, and produced three interactive Power BI dashboards. The analytical solution achieved a System Usability Scale score of 84.2 and a SERVQUAL score of 4.35/5, providing initial evidence of usability, stakeholder acceptance, and organizational applicability. MIDA5 contributes an initial, collaborative, user-centered methodological framework that operationalizes BPA through structured stakeholder participation, standardized artifacts, and iterative organizational validation, while reducing methodological and technological barriers.
Recent industry reporting indicates that meantime to exploit has become negative in several observed datasets, implying that exploitation may occur before patch availability for some classes of vulnerabilities. Adversarial use of artificial intelligence (AI) is a documented accelerant of this trend. This paper addresses the operational problem that follows in healthcare cybersecurity: the volume and velocity of vulnerability disclosure exceed human analytic capacity, which leads practitioners to under-prioritize, or defer entirely, individual Common Vulnerabilities and Exposures (CVEs) at precisely the moment their risk is rising. Adopting the design science research paradigm of Hevner et al. [8], the study develops and evaluates a purposeful information technology artifact intended to resolve this problem within a mid-sized United States healthcare system. The artifact is a three-application automated CVE intelligence, prioritization, and remediation-tracking pipeline implemented in Microsoft Azure Logic Apps, integrating the National Vulnerability Database (NVD), the CISA Known Exploited Vulnerabilities (KEV) catalog, the Microsoft Security Response Center (MSRC) CVRF API, Microsoft Defender, Claroty xDome, Microsoft Security Copilot, and ServiceNow, and operationalizing the four risk factors codified in CISA Binding Operational Directive (BOD) 26-04. In naturalistic operation across three CISA Weekly Vulnerability Summary bulletins, the artifact processed 5,216 unique CVE references and reduced them to 534 environment-relevant findings, an 89.8 percent exposure-first reduction, before expensive per-CVE enrichment and ticketing. Findings indicate that governed automation demonstrably increases CVE coverage, reduces low-value enrichment volume, and creates an auditable prioritization record. Because no controlled before-and-after time-and-motion study was conducted and no independent ground-truth severity labels were collected, remediation-speed and analyst-productivity outcomes remain future validation targets rather than demonstrated results.
Machine-learning intrusion detection is challenged by attacks resembling legitimate traffic and by class imbalance. This study evaluates Random Forest detection on the UNSW-NB15 dataset for five binary attack-versus-normal tasks: DoS, Exploit, Backdoor, Analysis, and Reconnaissance. Categorical attributes were label-encoded, features selected using mutual information, hyperparameters optimized by randomized search, decision thresholds tuned on validation data, and SMOTE applied to training data at a fixed target of 56,000 minority samples. SVM and KNN results are descriptive because the shared label-encoded representation precludes model-independent comparison. Without SMOTE, Random Forest achieved F1-scores of 0.9468 for DoS, 0.9504 for Exploit, 0.9839 for Backdoor, 0.9317 for Analysis, and 0.9720 for Reconnaissance. Fixed-SMOTE pipelines achieved 0.9539, 0.9516, 0.9668, 0.8370, and 0.9651, respectively. Thus, SMOTE improved DoS, marginally improved Exploit, and reduced performance for Backdoor, Analysis, and Reconnaissance. Confusion matrices showed fewer false positives and false negatives for DoS, while Analysis gained recall but produced substantially more false positives. These results compare separately optimized complete pipelines and do not isolate SMOTE’s effect. They characterize filtered target-attack-versus-normal streams, not operational mixed-attack traffic. Under this protocol, performance and precision–recall trade-offs were attack-dependent under the evaluated fixed-SMOTE experimental configuration, indicating that oversampling should be evaluated separately for each attack category.
General-text translation is converging toward near-full automation by large language models (LLMs), yet the translation of OTT multimedia content still demands a deliberately designed division of labor between humans and AI; this asymmetry is the starting point of this review. Korea occupies a distinctive position in this transformation: the global success of Korean content has created an asymmetric, outbound-heavy translation market in which linguistically complex Korean-source material must be localized into dozens of languages at speed, while the domestic language-service industry undergoes a structural shift from human translation to machine-translation post-editing. This paper reviews three research streams that have developed largely in isolation, namely LLM translation quality and automatic evaluation, human–AI collaboration and knowledge-worker productivity, and audiovisual translation, and integrates them through an operations-management input–process–output–outcome framework tailored to the Korean OTT context. The review finds that LLM-based post-editing improves draft quality and productivity, but that honorific register, culture-bound expressions, and speaker-relational meaning, all pervasive in Korean dialogue, lie largely outside current LLM competence. Three testable propositions are derived on how Korean-specific linguistic density conditions the productivity–quality trade-off, and a research agenda is proposed with implications for job redesign and capability development in language service providers, OTT platforms, and language-focused higher education. Designing this division of labor is a management problem before it is a technological one, and its answers lie where language expertise and operations management meet.
Music-to-dance generation requires precisely aligning movement dynamics with musical rhythm, yet existing methods rely on shallow conditioning or auxiliary beat-alignment objectives that fail to establish stable beat–motion correspondences. We present ChoreDiffusion, a diffusion-based framework that integrates explicit beat guidance directly into the denoising process. Central to our approach is a beat-enhanced cross-modal attention mechanism that injects beat-salience cues at every refinement step, promoting fine-grained synchronization beyond the reach of conventional conditioning pipelines. To support multiple dance styles within a unified model, we incorporate lightweight low-rank adaptation (LoRA) modules that encode style-specific motion signatures with only a small set of additional parameters per style, and a three-stage progressive curriculum stabilizes the joint learning of rhythmic alignment and stylistic expressivity. Experiments on two public multi-style dance benchmarks (AIST++ and FineDance) show that ChoreDiffusion achieves the lowest FID values among the compared generation methods on both benchmarks, while maintaining competitive rhythm alignment and multi-style controllability. These results indicate that embedding beat-aware guidance during generation, rather than applying it afterwards, is an effective route toward human-like musicality in music-driven choreography.
Educational assessment systems have primarily relied on episodic forms of assessment, including examinations, assignments, grades, and credentials. These approaches provide efficient and scalable summaries of achievement and yet capture only part of the developmental process through which learners build competence. Moreover, learning increasingly unfolds across digital platforms, workplaces, collaborative networks, and AI-mediated environments, generating rich evidence of learner development that remains fragmented across systems and contexts. Advances in artificial intelligence, learning analytics, multimodal analytics, learner modeling, and semantic interoperability make it increasingly feasible to connect, integrate, and interpret this evidence across contexts and over time. This paper introduces the AI-Mediated Continuous Assessment Infrastructure (AIM-CAI), a sociotechnical framework supporting longitudinal, probabilistic interpretation of distributed evidence of learning. Within AIM-CAI, continuous assessment refers to the ongoing accumulation and dynamic interpretation of evidence generated through learning activities. The framework integrates distributed evidence systems, evidence serialization mechanisms, AI-mediated semantic translation, probabilistic learner models, dynamic competency profiles, and federated governance architectures to support context-sensitive interpretations of learner development while maintaining human judgment, privacy, accountability, and learner agency. The authors examine implications for assessment, credentialing, lifelong learning, institutional roles, interoperability, and governance and outline a research agenda addressing key psychometric, ethical, and governance challenges, including validity, fairness, surveillance, semantic instability, and ownership of learning evidence.
For every prime-power dimension, a complete set of d+1 mutually unbiased bases (MUBs) is known, but in non-prime-power composite dimensions, the maximum number N(d) remains open. The tensor-product construction supplies mini(piai)+1 MUBs for d=∏ipiai, and this is the best lower bound currently known for every non-prime-power composite d≤100. We do not attempt to determine N(d); we ask whether these specific tensor-product sets admit one additional basis. We build and verify the sets for all 64 such dimensions, with pairwise overlap deviations below 10−15; run a construction-free joint search for d≤7; compare twelve optimizers; and test construction, convergence, and success thresholds. For a specified protocol A—algorithm, initialization distribution, success criterion, and stopping rule—let qA(d) denote the probability that one descent recovers a provably existing extension. Among the successful local-descent protocols tested, the inferred probabilities have comparable order, whereas the number of descents per computational budget differs much more strongly. With maxfun unbounded, exact gradients and optimization on U(d) recover guaranteed extensions at d=10,12, and 16. Raising the Riemannian CG iteration limit from 6000 to 150,000 changes none of the success counts or extension-search summaries. In d=6, the numerical failure to extend the tensor-product-basis triple reproduces a known analytic unextendibility theorem; it does not resolve the general four-MUB problem. Moreover, the actual extension target is maximally entangled. Haar-random initialization has zero probability of lying exactly on that structured submanifold, and the unrestricted search has not been validated for convergence to it, so recovery probabilities from the tensor-basis validation task cannot be transferred directly to the extension problem. The constructed bases, numerical summaries, per-restart arrays for the optimizer and threshold studies, complete per-descent arrays, and all code are publicly archived.
The effects of transcutaneous electroacupuncture stimulation (TEAS) on large-scale brain function remain insufficiently characterized. This study employed a graph-theoretical approach to analyze electroencephalogram (EEG) data from 48 healthy participants in the Pilot-6 TEAS study. Participants received sham (0 pps), 2.5 pps, 10 pps, and 80 pps stimulation during baseline, stimulation, and recovery phases. Functional connectivity was assessed using coherence and the weighted phase-lag index, followed by calculation of global and nodal graph measures from thresholded weighted undirected sensor-level networks. Descriptive analysis indicated potential frequency-related differences in EEG network organization. The 2.5 pps condition exhibited the highest average degree, whereas the 80 pps condition demonstrated the highest average clustering coefficient. At 10 pps, sensor-level maps revealed a distinct frontal–central betweenness-centrality pattern. Although 48 participants provided usable EEG data for descriptive analysis, only 3 participants had complete matched graph-metric data for all four stimulation conditions, limiting repeated-measures statistical validation. After correction for multiple comparisons, no statistically significant frequency-related effects were observed, and nodal hub differences were not independently confirmed. Consequently, these patterns should be interpreted as descriptive and exploratory rather than established group-level effects. These findings indicate that graph-theoretical EEG analysis may facilitate the identification of candidate network features for future investigations of TEAS-related brain network organization.
Security teams decide which vulnerabilities to patch first, which alerts to trust, and whether social media warns of new threats earlier than the official feeds. We answer these questions by directly measuring EdgeGuard, a deployed cyber-threat knowledge graph that merges eleven public threat feeds into one Neo4j database via MISP (an open threat-sharing platform) and the STIX 2.1 exchange format, recording for every entry which feed reported it and when. These records let the graph be read as a time series. Read this way, it shows that half of the vulnerabilities known to have been exploited were listed as exploited within five days of their publication (352 cases), and that a large ingestion spike in early 2026 came from a single feed rather than a real attack wave. Benchmarked against 10,000 threat-related social-media posts, the graph already held 96% of the actionable vulnerabilities the posts discussed and reported them at least as quickly, while most posts carried no actionable signal and social media led only in early warning of active exploitation. A crowd-sourced community layer additionally supplies the only intelligence tagged by industry sector. The deployed graph is thus a clean, timely, and comprehensive base, and live social ingestion a small, targeted enhancement.
Dense small traffic object detection is essential for intelligent transportation systems but remains challenging because distant targets contain limited visual details, densely distributed objects frequently overlap, and complex road backgrounds introduce substantial interference. To address these limitations, this study proposes GAD-YOLO, a multi-level feature enhancement network based on YOLOv9. Ghost-MSConv performs lightweight multi-receptive-field feature extraction in the backbone, Mixed Local Channel Attention combines local spatial relationships with global channel dependencies during feature refinement, and DySample performs content-adaptive point sampling during feature upsampling. In the primary experiments on a six-class traffic object dataset derived from UA-DETRAC, GAD-YOLO achieved a precision of 78.9%, a recall of 76.4%, an mAP50 of 82.8%, and an mAP50:95 of 65.5%. Compared with YOLOv9c, precision, recall, mAP50, and mAP50:95 increased by 5.4, 0.5, 3.1, and 4.8 percentage points, respectively. Under the complexity statistics used in the primary experiments, GAD-YOLO contains 25.455 M parameters and requires 102.4 GFLOPs, compared with 25.442 M parameters and 103.2 GFLOPs for YOLOv9c. Additional experiments on the public VisDrone2019-DET benchmark were conducted to evaluate cross-dataset applicability, small-object performance, scene-density sensitivity, and standardized inference efficiency. On the VisDrone2019-DET test-dev set, GAD-YOLO improved mAP50 and mAP50:95 from 26.5% and 15.7% to 27.1% and 16.3%, respectively. A COCO-style analysis further showed that APS increased from 6.72% to 7.31%, while the dense-subset mAP50:95 increased from 13.95% to 14.44%. Under an RTX 4090, batch-size-one, 640×640, FP32 inference protocol, GAD-YOLO achieved a mean latency of 9.98 ms and a throughput of 100.20 FPS. These results show that GAD-YOLO improves the primary traffic object detection task and yields modest positive performance differences on an independent public benchmark under the fixed experimental setting, while maintaining real-time inference capability.
Context: The Agile methodology has been prevalent in the software industry for more than two decades, marking a shift from plan-driven to market-driven approaches and introducing various challenges. While the literature identifies numerous challenges in Agile development, little attention has been given to their ranking and prioritization, which are critical for effective project management and decision making. This study fills this gap by combining empirical evidence from the literature and practitioners. Objectives: This study aims to identify and hierarchically prioritize the most recent challenges faced by Agile practitioners during product development. To achieve this, a Systematic Literature Review (SLR) was conducted using 115 published studies between 2010 and 2025 followed by empirical data collection from 30 Agile experts through semi-structured interviews conducted with practitioners from Agile companies and an online survey. This study applies Cumulative Voting (100-Dollar Test) and Multi-Criteria Decision Making (MCDM) techniques to rank and prioritize these challenges. Results: The SLR identifies several recurring Agile challenges; however, limited research has focused on their ranking and prioritization. The present study reveals new challenges, such as user interface complexities, lack of pre-development and pre-operational cost information, and lack of cost scalability at the module and feature levels. The current study identifies Inadequate Architecture (22%), Lack of Standardized Framework (18%), Communication and Coordination (16%), Poor Requirement Verification (13%), and Minimum Documentation (8%) as the most significant challenges. Conclusions: This study provides valuable insight for Agile practitioners and organizations, enabling more informed project planning, resource allocation, and strategic decision making. By focusing on the most critical challenges, teams can enhance software quality, streamline processes, and improve overall productivity in Agile environments.
Customer segmentation and purchase behavior prediction are fundamental tasks in intelligent e-commerce systems, enabling personalized marketing strategies and data-driven customer relationship management. However, conventional machine learning and graph neural network approaches primarily model pairwise interactions and often fail to capture higher-order relationships among customers, products, brands, and purchase contexts, limiting predictive performance and model interpretability. To address these challenges, this study proposes an Explainable Hypergraph Neural Network (EHGNN) framework that integrates higher-order hypergraph representation learning with post hoc explainability using SHAP. The proposed framework constructs a heterogeneous hypergraph from customer transaction data, learns informative customer embeddings via hypergraph convolution, segments customers via clustering, and predicts purchase behavior using an embedding fusion network. Comprehensive experiments were conducted to compare the proposed framework with conventional clustering algorithms, deep clustering methods, graph neural networks, and hypergraph neural networks. Experimental results demonstrate that the proposed EHGNN consistently achieved superior performance, obtaining a Silhouette Coefficient of 0.824, Davies–Bouldin Index of 0.336, and Calinski–Harabasz Index of 2815 for customer segmentation. For purchase behavior prediction, the proposed framework achieved an Accuracy of 97.30%, Precision of 97.00%, Recall of 96.80%, F1-score of 96.90%, Area Under the Receiver Operating Characteristic Curve (AUC) of 99.20%, and a Matthews Correlation Coefficient (MCC) of 0.942, outperforming all benchmark methods. These findings demonstrate that modeling higher-order customer relationships using hypergraph learning substantially improves both customer segmentation quality and purchase behavior prediction, while maintaining model transparency via explainable artificial intelligence. The proposed EHGNN framework provides an effective, robust, and interpretable solution for intelligent customer analytics and personalized decision support in modern e-commerce environments.
