Artificial Intelligence is reshaping global socio-economic structures, yet existing studies often analyze its economic, social, and market impacts in isolation, lacking a unified framework that handles multi-dimensional interactions. This paper constructs an end-to-end machine learning pipeline - from raw heterogeneous data input, through three cascaded modules (micro: Random Forest + SHAP; meso: XGBoost; macro: K-means++ clustering), to actionable policy outputs - to quantify the complex socio-economic effects of AI adoption across countries and industries. Using a comprehensive global dataset (2020-2025), this research finds that: (1) Human-AI Collaboration Rate (HAICR) is the dominant mitigator of AIdriven job loss, outweighing AI Adoption Rate (AIAD), while high AIAD significantly exacerbates job displacement in automated sectors like manufacturing and retail; (2) Strict regulation is the strongest driver of consumer trust, but its efficacy is moderated by cultural contexts and tool transparency; (3) China and the U.S. lead AI-driven markets, whereas India and South Korea require infrastructure and policy upgrades. By coupling machine learning methods in a tripartite analytical pipeline, this study moves beyond isolated econometric models and provides quantifiable, actionable insights for balancing AI innovation with social welfare.
Visual Question Answering (VQA) aims to achieve cross-modal semantic understanding through joint modeling of visual content and natural language. Although existing attention-based approaches effectively align features, they struggle to simultaneously accommodate global semantic modeling and local fine-grained perception within single-path architectures, leading to semantic drift and underutilization of information when addressing multi-perspective reasoning tasks. To mitigate this limitation, we propose a Decoupled Global-Local Collaborative Network (DGL-Net) that structurally partitions the cross-modal reasoning process into two complementary branches: global semantic understanding and local fine-grained modeling. Specifically, the global branch employs hierarchical attention fusion to model long-range dependencies and holistic scene semantics, providing contextual constraints for reasoning. The local branch introduces semantic relation enhancement and relevant region screening to explicitly model intra-modal structural relationships, strengthening fine-grained region-word alignment by injecting structural information. During training and inference, a branch-decoupling and collaborative fusion strategy enables independent yet complementary feature learning. Extensive experimental results demonstrate accuracies of 72.15% and 57.87% on VQA-v2 and GQA benchmarks, respectively. Notably, the proposed method exhibits enhanced robustness and generalization in complex semantic and compositional reasoning tasks, validating the efficacy of the decoupled collaborative strategy.
Data validation in ETL workflows is most often performed after transformation, even though many data quality issues originate earlier in the pipeline. In practice, this means that errors are detected late, when tracing their source and applying corrections already requires additional effort. In this work, we explore whether virtual reality (VR) could be used as an alternative interface for addressing this problem earlier, at the stage of raw data inspection. We present VR-ETL, a proof-of-concept system that allows users to inspect and correct tabular data before transformation by interacting with it inside an immersive 3D environment. The system imports CSV datasets and applies a set of rule-based validation checks targeting common data quality problems, such as missing values, format inconsistencies, and basic referential integrity violations. Detected issues are visualized directly within the 3D representation of the data, using color-coded cells and simple spatial summaries that reflect error distribution across columns and records. A functional prototype was implemented using Unity and OpenXR and tested on structured datasets containing up to 10,000 records. During controlled technical testing, the system supported automated validation, real-time visual feedback, and immediate re-validation after user corrections, while maintaining performance suitable for immersive use on consumer-grade hardware. All intentionally seeded rule-based errors were detected. Although usability and workflow efficiency were not evaluated in this study, the prototype demonstrates that pre-transformation data validation in VR is technically feasible and provides a concrete starting point for future empirical investigations.
Cloud computing has become a dominant paradigm for delivering scalable and flexible on-demand resources; however, efficiently executing high performance computing (HPC) workloads remains challenging, particularly in heterogeneous environments. Conventional static scheduling methods often lead to poor resource utilization and increased makespan, while dynamic approaches improve load distribution but introduce significant overhead due to continuous monitoring and real-time decision-making. To address these challenges, this paper proposes an SLA-aware Dynamic Enhanced Resource-Aware Load Balancing Algorithm (SLADE- RALBA). The algorithm minimizes load imbalance by considering the computational capacities of virtual machines and ensures Service Level Agreement (SLA) compliance through a three-tier priority-based workflow. The proposed approach is evaluated using CloudSim Plus on two benchmark datasets: Heterogeneous Computing Scheduling Problem (HCSP) instances and the Google Cloud Jobs dataset. Results demonstrate that SLA-DE-RALBA consistently outperforms baseline algorithms, including RALBA, DRALBA, DE-RALBA, SLA-RALBA, Dynamic Max- Min, PSSLB, and PSSELB, across key metrics such as makespan, resource utilization, job rejection, throughput, execution time, and cost. Notably, it achieves zero job rejection, reduces energy consumption by up to 85%, improves resource utilization by 11.9%, lowers makespan by 41-45%, and decreases execution time by up to 57%, making it a robust and efficient solution for HPC workload scheduling in cloud environments.
Energy-aware path planning for industrial autonomous vehicles requires more than a geometric shortest path: terrain, payload, drivetrain efficiency, turning, and a hard task-level energy budget jointly affect feasibility. We present Neuro-PathNet-EC, a physics-aware constrained planner with learned multimodal guidance. The planner operates on heading-aware grid states and evaluates directed edge energy from friction and elevation. Exact Pareto labels enforce the hard energy budget, while a goal-conditioned ResNet-18 network fuses four raster channels with vehicle semantics and predicts an eight-channel dense cost-to-go map. The learned heuristic is used only for guidance; it is not assumed admissible or consistent. Exact Energy-A* obtains its guarantee from an analytic admissible anchor, and a FOCAL extension provides the bound C ≤ wC*. On a layout-disjoint synthetic industrial dataset, five training seeds achieve a test relative heuristic MAE of 0.0289 ± 0.0014 and Spearman correlation of 0.9963 ± 0.0004. Across 512 paired binding-budget queries per seed (2,560 seed-query runs), full multimodal neural-constrained search returns the exact constrained objective in every run while reducing expanded labels by 10.73% relative to Energy-A*. The semantic branch also reduces labels relative to an otherwise matched image-only model. FOCAL satisfies its declared bound in all evaluated cases, although its current correctness-first OPEN management does not improve runtime. The evaluation is based on synthetic layouts and does not constitute public-benchmark or real-vehicle validation.
Mental health disorders affect over one billion people worldwide, with treatment gaps particularly severe in resource-constrained regions such as Pakistan, where only 0.19 psychiatrists are available per 100,000 population. This study presents PocketPsyc, a mobile-based system designed to deliver scalable, evidence based mental health support through Cognitive Behavioral Therapy (CBT). The system integrates three specialized AI models: a fine-tuned BART model for therapeutic response generation (88.2% BLEU score), a RoBERTa-base classifier for real-time emotion recognition (F1-score of 0.89 across seven categories), and a TinyLlama-1.1B model for personalized mindfulness guidance (rated 4.3/5 in human evaluations). To ensure user privacy, the platform employs client-side AES-256 encryption and row-level security mechanisms. Additionally, a hybrid crisis detection module combines clinical threshold monitoring with linguistic cue analysis to identify high-risk scenarios. The system achieves an average end-to-end response latency of 3.2 seconds. These findings demonstrate that PocketPsyc provides a technically feasible, privacy-preserving, and scalable solution for delivering AI-assisted mental health support in underserved populations.
We prove the existence of a maximal solution to any lattice-valued fuzzy relational equation, inequation, or system, in which infimum, supremum, and the standard composition are used. This result holds under the assumption that, in the codomain lattice, the infimum operation commutes with suprema of chains. To establish this result, we first recall and extend several lattice-theoretic properties under the assumption of meet-continuity. In particular, we recall that meet-continuity of the codomain lattice ensures composition-continuity in the lattice of L-valued relations. Building on these properties, we introduce algorithms for constructing the least solution to a broad class of equations and inequations, and prove that these algorithms terminate after at most countably many steps.
This study investigates students’ acceptance of AI-assisted translation tools by proposing an extended Technology Acceptance Model (TAM-AI) for AIassisted translation learning contexts. Unlike prior additive approaches, the proposed model explains how translation-specific perceptions influence behavioral intention (BI) through underlying cognitive mechanisms. A survey was conducted among undergraduate translation students and analyzed using Structural Equation Modeling (SEM). The results indicate that perceived usefulness (PU) and perceived ease of use (PEOU) significantly predict BI. In addition, translation-specific factors influence technology acceptance indirectly: perceived translation quality operates through trust, feedback clarity through cognitive understanding, and cognitive load reduction through effort reduction. The TAM-AI model demonstrates greater explanatory power than the baseline TAM. These findings provide a deeper understanding of technology acceptance in AI-assisted learning environments and offer practical implications for the design of AI-assisted translation tools. Future studies with larger and more diverse samples are encouraged to further validate the proposed model.
We consider the problem of optimizing a complex function f(・) given a sample consisting of M input-output pairs (argument, function value), S = {(xn,yn)}_N=1^M , where x ∈ \mathbb{R}^{d} and y ∈ \mathbb{R} . We assume that the sample is available only in a distributed manner at K entities, where each entity controls a sample part Sk and S = ∪_k=1^K Sk. The problem is addressed employing a machine learning approach. Our goal is to generate a surrogate of the function based on the given sample, and to perform the optimization on the surrogate function. Due to the distributed nature of the sample we employ Federated Machine Learning (FML) to generate the surrogate function by joint efforts of all K entities that possess the sample parts. We call the entities involved in FML clients, and we assume that certain clients participate in FML in a malicious manner by poisoning training data or the model parameters they generate locally. Systematic experiments have been performed to study the characteristics of the surrogate function and its capabilities to provide reliable analysis of the original function depending on the sample available for training, and impacts of the training data poisoning, as well as locally generated model parameters.
Analyzing open-ended student feedback in course evaluations is a laborintensive task due to the unstructured and complex nature of natural language. While Large Language Models (LLMs) offer significant potential for automation, a welldefined methodology for their application in analyzing student feedback remains underdeveloped. This paper addresses this gap by proposing an LLM-based feedback analytics pipeline designed to transform students’ open-ended feedback into structured, actionable insights. The pipeline consists of three sequential stages: (i) segmenting student feedback into semantic units and assigning polarity (sentiment) to those units; (ii) topical classification of semantic units, and (iii) summarization of units within each topical category. By systematizing these processes, the proposed method enables educators and course managers to efficiently derive meaningful patterns from vast datasets of student opinions. We evaluated the proposed method using a comprehensive dataset from several editions of a U.S. university course, yielding encouraging results of the method’s effectiveness. This research provides a scalable, generic methodology for (semi-)automated feedback analysis, ultimately supporting data-informed improvements in teaching and course management.
The rapid development of artificial intelligence (AI) and large language models (LLMs) is reshaping assessment practices in higher education, particularly in grading open-ended questions. Traditional grading is time-consuming, prone to inconsistency, and difficult to scale for large classes. LLMs offer an opportunity to improve efficiency, consistency, and feedback frequency by leveraging their ability to understand and reason about natural language. This paper introduces EduGrader, an open-source, multi-model grading platform integrating several LLMs, including GPT-4o, GPT-4o-mini, DeepSeek-Chat, GPT-5.1, and DeepSeek-Reasoner. EduGrader supports three grading strictness levels (lenient, neutral, strict) and generates numerical scores with concise explanations highlighting correct reasoning, missing concepts, and errors. The system operates in two modes: a provided reference-answer grading mode guided by instructor solutions and a generated reference-answer grading mode that automatically creates reference answers from course materials. All experiments are conducted in Serbian, a comparatively low resource language underrepresented in mainstream AI training data, extending AI assisted grading research beyond the dominant English-language context. EduGrader is evaluated on 686 real student responses from six university courses across two institutions, including free-text explanations and code-based answers. In its best performing configuration, the system achieves a Pearson correlation coefficient of 0.90 with human graders while reducing grading workload and improving reliability, though it is intended to assist rather than replace instructors.
In the context of today's big data and numerical intelligence era, this study explores an incremental network information extraction technology for Web topic detection characterized by the semi-structured or unstructured big data as important research object to promote network information detection application. This study takes Web big data as the main research object and proposes an incremental network information extraction idea for Web topic detection. In this idea, the designed algorithm of theme similarity measurement for incremental network information extraction can extract Web instances related to theme, and calculate importance of Web instances related to theme, furthermore, the designed algorithm of incremental instance extraction for Web topic detection can analyze Pattern and BasePattern according to extracted Web instance URL, and conduct segmentation for Web instance title and text content, extract keywords, which are capable of describing Web topic. Experimental results demonstrate that the framework, method, and algorithm proposed in this paper significantly outperform traditional methods in network information extraction. Particularly, the accuracy rate of extracted Web instances that are similar to the theme can reach 0.833, the F-Measure value of extracted Web instances that are similar to the theme under different threshold adjustment is close to 0.83, the accuracy rate of topic detection under the condition of determining the number of Web news instances extracted, the threshold and the parameter value is close to 0.82. The study concludes that the incremental network information extraction idea proposed in this paper is feasible, verifiable, and superior, and can play an important role in reconfiguring numerical intelligence warehouses for detecting Web topic, inferring the Web hierarchical big data propagation path.
Remote sensing image classification is a fundamental task for Earth observation and environmental monitoring. However, conventional convolutional neural networks (CNNs) are limited by computational capacity and struggle to efficiently process the rapidly growing volume of remote sensing data. To address this limitation, we propose HAQCCN (Hybrid Asymmetric Quantum-Classical Convolutional Network), a novel hybrid architecture that integrates quantum computation into the classical convolutional framework through asymmetric quantum convolutional circuits. In HAQCCN, the asymmetric quantum circuits enable a limited number of qubits to process more classical data while maintaining excellent feature extraction capability. Experiments conducted on the IBM Qiskit platform using the Overhead-MNIST, PatternNet, and RSI-CB256 datasets demonstrate that HAQCCN outperforms conventional CNNs and existing quantum models. Furthermore, we systematically investigate the effects of encoding schemes, the number of quantum convolutional kernels, and the number of qubits on model performance, confirming the effectiveness and scalability of the proposed method for remote sensing image classification.
Governmental policies for transparency and reuse of public sector information have encouraged the launch of open government data portals around the world. Many of these portals are based on pyramidal structures: national open data portals are aggregators of the contents harvested from open data portals maintained by governments in charge of administrative areas with a narrower scope. Taking into account this hierarchical organization, these open data portals lack consistent and scalable mechanisms for thematic annotation, limiting dataset discoverability. This work proposes a framework for the automated thematic classification of open government data. The framework integrates (i) thematic annotation quality assessment, (ii) supervised machine learning models trained on annotated metadata corpora, and (iii) embedding-based semantic similarity methods for theme assignment in the absence of reliable annotations. The framework is evaluated using 29,793 datasets from data.europa.eu, the European open data portal. Experimental results show that supervised models achieve high classification performance, with Support Vector Machines reaching an accuracy of 93.65%, while unsupervised embedding-based approaches achieve substantial semantic agreement with portal-assigned themes (74.56%) using transformer-based representations. These results demonstrate that the proposed framework enables scalable, consistent, and interoperable thematic annotation, offering both theoretical contributions to automated metadata enrichment and practical value for integration into large-scale open data portal infrastructures.
The secure data sharing across data domains in cloud-edge collaborative environment faces challenges of low data access control efficiency and a lack of dynamism. In this work, we proposed FPCA, a fully constant-length and policy updating cross data domain access control for cloud-edge collaborative environment. FPCA address the above challenges by proposing three algorithms: the Multi Data Domains Key Generation (MDKG) algorithm maintains constant secret key length and enables cross data domain access without attribute conversion, the Constant-Length Ciphertext Encryption (CLCE) algorithm maintains constant ciphertext length, reducing decryption overhead, and the Access Policy Update (APU) algorithm updates the access policy with constant-length update message and low computational complexity. Security analysis demonstrates FPCA resist key forgery, collusion, and chosen-key attacks while ensuring backward security. Simulation experiment shows that FPCA outperforms existing schemes. For FPCA, the storage costs are only 50% to 80% of other constant-length ABE schemes, the communication costs remains at 0.38-1.27 KB between entities, the time costs of access policy update and cross data domain access remain around 8.4 ms and 30.4 ms, respectively, and are lower than other similar schemes. These results confirm the efficiency of FPCA in achieving secure and dynamic cross data domain access in cloud-edge collaborative environment.
The detection of epileptic tonic-clonic seizures during everyday life based on accelerometric data from wearable devices would enhance the diagnostic and the follow-up of the epileptic patients. We develop an algorithm which may contribute to recognition of tonic-clonic epileptic seizure based on accelerometer data that can be collected from mobile and wearable devices. We consider this task to be a multivariate time-series classification problem. State-of-the art solutions to this problem are based on machine learning techniques, such as Random Convolutional Kernel Transform (ROCKET). We enhance ROCKET by replacing standard convolution with dynamic convolution. Dynamic convolution was originally defined for univariate time series, therefore, we extend it to multivariate time series. We perform experiments on two publicly available real-world datasets related to tonicclonic seizures. The experimental results show that the proposed enhancements of the ROCKET algorithm significantly reduce the average classification error. Moreof deep neural networks that are commonly used in the domain of time-series classification. An enhanced version of the ROCKET algorithm is proposed for the automated detection of epileptic tonic-clonic seizures using accelerometer data. To assist reproducibility and follow-up works, we made our implementation publicly available at https://github.com/kr7/seizure .
This pilot study offers a linguistic evaluation of six machine translation systems: GPT-4o, GPT-5, Gemini 2.5 Flash, Google Translate, Microsoft Translator, and NLLB-600M applied to the translation of a short excerpt of Orwell's "1984" into Macedonian. The analysis consisted of three interconnected experiments: manual annotation of translation errors and comparison with human output, evaluation using eight popular MT metrics, and sentence-level similarity analysis via cosine similarity, Jaccard similarity, and Levenshtein distance. Manual annotation revealed that stylistic errors (48.47%) and linguistic errors (34.54%) were the most common. The LLMs outperformed other systems, particularly GPT-5, while NLLB-600M performed poorly, often introducing incomprehensible sentences or non-existent words. Metrics-based evaluation showed that lexical metrics sometimes penalized fluent and accurate translations that deviated from the reference. Sentence similarity analysis confirmed that accurate translations were more consistent, while wrong-wrong sentence pairs were more divergent, especially in Levenshtein scores. The findings underscore the importance of combining manual and metric-based evaluation to fully understand MT quality, particularly in low-resource settings.
Existing deep learning models in digital pathology typically require extensive labeled data and show limited generalization across organs. In contrast, large vision models exhibit effective feature extraction capabilities, enabling pathological image analysis for gastrointestinal cancer with relatively small sample sizes. In this study, we developed a screening framework leveraging a large vision model for coarse-grained classification of gastric and colorectal tissues. The model was evaluated on multicenter cohorts and under limited-data conditions. Using labeled tiles from only 76 whole-slide images, the model achieved class-averaged sensitivity and precision of 0.9816 and 0.9808 on the internal test set, and 0.9161 and 0.9179 on the external test set. When trained with only 200 tiles per class from 20 whole-slide images, the model maintained comparable performance, achieving sensitivity and precision of 0.9548 and 0.9518. These findings suggest that the model has reliable performance across multicenter cohorts and potential applicability in clinical pathology workflows.
