		FeedBack \| Contact Us \| Links \| Site Map

Welcome to Journal of Theoretical and Applied Information Technology

Submit Paper / Call for Papers

Journal receives papers in continuous flow and we will consider articles from a wide range of Information Technology disciplines encompassing the most basic research to the most innovative technologies. Please submit your papers electronically to our submission system at http://jatit.org/submit_paper.php in an MSWord, Pdf or compatible format so that they may be evaluated for publication in the upcoming issue. This journal uses a blinded review process; please remember to include all your personal identifiable information in the manuscript before submitting it for review, we will edit the necessary information at our side. Submissions to JATIT should be full research / review papers (properly indicated below main title).

Journal of Theoretical and Applied Information Technology
April 2026 | Vol. 104 No.7

Title:	THE ROLE OF DIGITAL TRANSFORMATION AUDIT IN MANAGING DIGITAL PAYMENT VALIDATION COMPLEXITY AND ITS IMPLICATIONS FOR AUDIT QUALITY
Author:	AUFA SYAHDURANI, AUDREY JADE SUKANDI, BAMBANG LEO HANDOKO
Abstract:	This study analyzes the role of digital transformation audit in managing the complexity of digital payment validation and its implications for external audit quality. With the growing use of digital transactions such as QRIS in the Society 5.0 era, auditors face challenges validating real-time, large-scale data. A digital transformation audit leveraging AI, Big Data Analytics, Machine Learning, and other technologies is proposed to enhance efficiency and accuracy. Using an explanatory quantitative approach based on the DeLone and McLean IS Success Model, data were collected from external auditors in Indonesia and analyzed with PLS-SEM. The findings revealed that System Quality, Information Quality, and the Complexity of Digital Payment Validation exerted significant direct effects on Audit Quality. In contrast, the mediation role of complexity was not supported. These results indicated that technological quality contributes to audit quality primarily through direct pathways, underscoring the importance of improvements in system performance and information accuracy for strengthening external audit outcomes in digital transaction environments.
Keywords:	Digital Transformation Audit, External Audit, Digital Payment, Complexity Validation, Audit Quality
DOI:	https://doi.org/10.5281/zenodo.19593224
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	DOMAIN-INVARIANT REPRESENTATION LEARNING FOR GENERALIZABLE TEXT MINING ACROSS MULTIPLE DOMAINS
Author:	SRINADH SWAMY MAJETI, BOMMA RAMAKRISHNA, PEDDADA NAGAMANI, MAREPALLI RADHA, CH. V. RAVI SANKAR, S. JAYANTH, LAKSHMI PRASANNA BYRAPUNENI, GUNTI SURENDRA, MEDIKONDA ASHA KIRAN, MANYAM THAILE, RAMESH BABU PITTALA
Abstract:	Text mining models often struggle to generalize across domains due to domain-specific linguistic and contextual biases. This limitation significantly affects real-world deployment where data distributions continuously shift. In this work, we hypothesize that explicitly learning domain-invariant representations can substantially improve cross-domain robustness without requiring target-domain fine-tuning. To address this, we propose a Domain-Invariant Representation Learning (DIRL) framework that integrates shared–private representation decomposition, adversarial domain confusion, and contrastive semantic alignment within a transformer-based architecture. The proposed method is evaluated on Amazon Reviews, MDSD, and 20 Newsgroups under cross-domain and unseen-domain generalization settings. Experimental results demonstrate consistent improvements of 4–9% in Macro-F1 over strong baselines, along with a significant reduction in domain generalization gap. These findings confirm that enforcing domain invariance at the representation level enhances scalability, robustness, and real-world applicability of text mining systems.
Keywords:	Domain-Invariant Representation Learning; Text Mining; Cross-Domain Generalization; Multi-Domain Learning; Transfer Learning; Transformer-Based Models; Adversarial Learning; Contrastive Representation Learning; Document Classification; Sentiment Analysis
DOI:	https://doi.org/10.5281/zenodo.19593257
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	TEMPORAL-AWARE DATA AUGMENTATION VIA VARIATIONAL MODE DECOMPOSITION AND CONDITIONAL GANS FOR EEG-BASED SEIZURE DETECTION
Author:	MR. KONDANNA KANAMANENI, DR. VENKATA RAJU K
Abstract:	Automatic detection of epileptic seizures based on electroencephalogram (EEG) signal has been an essential challenge issue in automated clinical diagnostic systems. The imbalance issue in the first place and the lack of labeled seizure data seriously hinder the progress of effective deep learning models. The proposed paper introduces a temporal-aware data augmentation system that fuses Variational Mode Decomposition (VMD) and the Conditional Generative Adversarial Networks (cGANs) in order to produce realistic synthetic EEG seizure data and maintain important temporal and spectral features. The VMD-cGAN model breaks down EEG signals into intrinsic mode functions (IMFs) that represent various frequency components, and the generator learns and reproduces the physiologically realistic seizure patterns. We assess our method using CHB-MIT scalp EEG database, proving that augmented training data significantly enhances seizure detection rates. Our experimental findings indicate 97.84% accuracy, 96.73% sensitivity, and 98.52% specificity, which is 3-5% higher than state-of-the-art techniques. The framework manages the lack of data while preserving temporal relationships and morphological properties required to properly identify seizures.
Keywords:	Electroencephalogram, Seizure Detection, Data Augmentation, Variational Mode Decomposition, Conditional Generative Adversarial Networks
DOI:	https://doi.org/10.5281/zenodo.19593288
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	SCALABLE PARKINSONS DISEASE PREDICTION USING MATHEMATICAL MODELING, HILBERT TRANSFORMS, AND TRANSFORMER-BASED DEEP LEARNING
Author:	THALAPATHIRAJ.S, VIVEKANANDAM B, RAJENDRA KUMAR TRIPATHI
Abstract:	Parkinsons disease (PD) is a progressive neurological disorder necessitating early and precise diagnosis for effective therapy. However, traditional machine learning and convolutional neural network (CNN) methods generally have problems when used on heterogeneous biomedical data since they don't scale well, don't generalize well, and are harder to understand. To tackle these issues, this research presents a scalable hybrid system that combines mathematical modeling, Hilbert transform-based spatial embedding, and transformer-based deep learning architectures for predicting Parkinson's disease. The suggested Hilbert-based embedding adds biologically inspired spatial correlations that keep structural information and make features more stable. To efficiently capture both local and global dependencies, these improved features are subsequently processed using advanced transformer architectures including Swin Transformer and Vision Transformer (ViT). Testing the proposed framework on multimodal datasets that include spiral drawings, wave patterns, and functional MRI (fMRI) pictures shows that it is more accurate, precise, and recall than traditional CNN and machine learning models. The Swin Transformer with Hilbert embedding had the best performance, with an accuracy of 97.96%. This shows that it is more general and more robust. The findings demonstrate that the suggested mathematically based framework offers a scalable, interpretable, and clinically significant approach for the early prediction of Parkinson’s disease.
Keywords:	Mathematical Modeling, Machine Learning, Mortality Diseases, Parkinson’s Disease, Hilbert Matrix.
DOI:	https://doi.org/10.5281/zenodo.19593314
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	STRUCTURAL DEFENSE-ORIENTED GCR-SCAPS METHOD WITH SPARSE CAPSULE ENCODING AND CONSISTENCY-BASED LEARNING
Author:	HEMASHREE P , N VALLIAMMAL
Abstract:	Adversarial perturbations compromise the structural reliability of remote sensing image classification, challenging the stability of decision-making systems in security-critical contexts. The proposed GCR-SCaps model introduces a Geometric Consistency Reinforced Sparse Capsule Network tailored to counter structural attacks through object-aware representations. Initial convolutional layers extract shallow spatial cues, which are then transformed into capsule vectors with enforced sparsity to promote discriminative feature isolation. Dynamic routing captures inter-capsule agreement, while final class capsules preserve semantic integrity. A parallel stream generates geometrically altered inputs to compute a consistency loss, aligning original and transformed predictions. This dual-path strategy ensures robustness against geometric distortions while maintaining classification fidelity. The joint optimization of structural margin and geometric consistency losses strengthens spatial reasoning, minimizing confusion caused by adversarial interference. GCR-SCaps establishes a structurally grounded defense architecture that enhances resilience, preserves spatial coherence, and ensures dependable classification performance across diverse aerial scenarios prone to manipulation or ambiguity.
Keywords:	Remote Sensing Classification, Capsule Networks, Geometric Consistency, Adversarial Defense, Sparse Routing, Structural Robustness
DOI:	https://doi.org/10.5281/zenodo.19593327
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	AN INTEGRATED ANN FRAMEWORK USING XGBOOST AND LSTM FOR CONDITION-BASED MAINTENANCE OF ROTATING EQUIPMENT
Author:	BONTHA SUSMITHA, PUTTI SRINIVASARAO
Abstract:	Preserving reliability and efficiency in industrial applications including rotating machinery depends on a fundamental method called Condition-Based Maintenance (CBM). Sometimes, two of the most often used traditional techniques of maintenance -reactive and preventive maintenance can cause needless downtime or failures. Existing research on CBM explored the Machine learning and deep learning models for Diadnosis of faults in machine for their Remaining Useful life (RUL), even though the models are existing from past years many of these not encounter challenges with high dimensional sensor data and often fail to adequately model temporal dependencies. This gap limits predictive accuracy and interpretability in real-world industrial applications. To address this, this paper suggests an ANN framework for enhanced fault diagnosis and RUL prediction in rotating machinery combining eXtreme Gradient Boosting (XGBoost) and Long Short-Term Memory (LSTM) networks. This will help one to handle these challenges. To maximize the input parameters of the LSTM model, the suggested system uses XGBoost for strong feature selection and importance ranking. XGBoost high-dimensional data handling and overfitting control features provide the LSTM network only the most relevant features. The LSTM model is appropriate for precise fault classification and RUL prediction since it naturally detects temporal dependencies in sensor data. When combined, deep sequential modeling and ensemble learning improve predictive performance as they draw on one another's strengths. Experimental vibration data collected from vibration sensors demonstrates that the proposed XGBoost-LSTM model consistently outperforms conventional ML and DL models with performance measures including recall, accuracy, accuracy, F1-score, and mean absolute error (MAE). The experimental results show that, contributes new knowledge by showing that coupling feature selection with temporal modeling yields a scalable, interpretable, and high-performing CBM solution s. By providing a fast, scalable AI-driven framework that reduces maintenance costs, improves operational safety, and minimizes downtime, this work advances predictive maintenance.
Keywords:	Condition-Based Maintenance, Rotating Equipment, XGBoost, LSTM, Predictive Maintenance, Fault Diagnosis, Remaining Useful Life.
DOI:	https://doi.org/10.5281/zenodo.19593352
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	PRIVACY PRESERVING HIGH-DIMENSIONAL DISTRIBUTED LEARNING VIA CLIENT-SIDE DEEP DENOISING SPARSE AUTOENCODERS AND SPLIT LEARNING
Author:	MR. LINGAM SUMAN , DR. S. VENKATA LAKSHMI
Abstract:	Privacy preservation (PP) in deep learning (DL) is becoming increasingly important, particularly when working with high-dimensional datasets that pose considerable computing hurdles. Traditional approaches frequently experience the "curse of dimensionality," resulting in inefficiencies and increased privacy threats. To address these issues, this study offers an innovative process that combines Split Learning (SL) with Client-side Deep Denoising Sparse Autoencoders (CDDSAEs), resulting in a more efficient and privacy-conscious solution. The technique begins with attribute filtering, which uses a threshold-based mechanism to keep just the most important data properties, hence lowering dimensionality from the start. The data is then vertically shard to distribute computational burdens and let each client to focus on a single data shard. In the SL architecture, DDSAEs are used on the client side to compress and anonymized data before transmission, guaranteeing that only low-dimensional, PP representations are exchanged with the server. Performance investigation shows that the suggested model reduces latency by 17.7%, runs faster by 40.3%, and has a minimum information loss of 0.22, demonstrating its superiority over existing techniques.
Keywords:	Curse of Dimensionality, Split Learning, Attribute Filtering, Vertical Sharding, Denoise.
DOI:	https://doi.org/10.5281/zenodo.19593374
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	APPLICATION OF DIGITAL FORENSIC TECHNOLOGIES TO DETECT AND PRESERVE DIGITAL TRACES IN PRE-COURT INVESTIGATIONS
Author:	VIACHESLAV KULIUSH, OLEKSANDR SHEVCHENKO, VLADAS TUMALAVIČIUS, VALERII NIKITCHENKO, OLEKSANDR HRUZD
Abstract:	The article presents the results of a comprehensive study of the application of digital forensic technologies for the detection and preservation of digital traces in pre-trial investigation. The cases of Ukraine, Lithuania, and Latvia were used in the study. The relevance of the study is determined by the growing need for the implementation of modern digital forensic technologies capable of ensuring reliable detection, fixation, and preservation of digital traces in pre-trial investigation. The aim of the study is to assess the effectiveness of forensic platforms (Autopsy, FTK, X-Ways, Magnet AXIOM), cryptographic algorithms SHA-256 and SHA-3, as well as blockchain solutions to maintain the continuity of the “chain of evidence” in three countries with different levels of digital maturity. The object of the study was forensic platforms (Autopsy, FTK, X-Ways, Magnet AXIOM), cryptographic algorithms SHA-256 and SHA-3, as well as blockchain solutions to maintain the continuity of the “chain of evidence”. The research employed the following methods: experimental testing of forensic tools on simulated datasets, performance assessment of hash algorithms and blockchain logs, and comparative analysis of practices in Ukraine, Lithuania, and Latvia. The results of the study revealed significant differences between countries. In Ukraine, certified forensic tools are used in only 35% of units, and the reliability of the “chain of evidence” does not exceed 0.72, which reduces the level of trust in digital evidence in more than 40% of criminal cases. In Lithuania, the level of certification reaches 85%, compliance with EU directives and GDPR ensures the stability of procedures at the level of 0.91, and the speed of digital data processing is 28% higher than in Ukraine. Latvia occupies an intermediate position: the level of certification is 65%, independent verification of digital evidence exceeds 75%, and the transparency of blockchain logs is estimated at 0.87 with an average transaction delay of 250 ms. The academic novelty of the study is the comprehensive technical and legal analysis of digital forensic technologies in three countries with different levels of maturity of digital infrastructure and the development of a model of optimal workflow for pre-trial investigation.
Keywords:	Digital Forensics, Electronic Evidence, Blockchain, SHA-256, SHA-3, Forensic Platforms, Criminal Justice, Rule Of Law, Ukraine, Lithuania, Latvia
DOI:	https://doi.org/10.5281/zenodo.19593411
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	TRANSFORMATION OF THE DEVELOPMENT STRATEGY OF HIGH-TECHNOLOGICAL PRODUCTION SYSTEMS
Author:	SVITLANA FILYPPOVA, VOLODYMYR FILIPPOV, SVITLANA YERMAK, LIUBOV NIEKRASOVA, OLEKSANDR BALAN, OLEKSIY SOLOHUB
Abstract:	In the article, the features of the formation and development of high-tech production systems in countries with different levels of economic development are examined. The purpose of the study is a comprehensive analysis of the prerequisites, structure and results of the functioning of high-tech production systems in countries with different levels of economic development based on a system of indicators of innovative activity, human capital, high-tech production and international trade in high-tech products, as well as substantiation of the directions of transformation of strategies for the development of high-tech production systems in modern conditions of global economic competition. The methodological basis of the study is the methods of theoretical generalization, statistical and comparative analysis, correlation analysis and conceptual modeling. The empirical basis of the study was formed based on the analysis of indicators of research and development costs, employment in knowledge-intensive types of economic activity, the share of high-tech production, as well as indicators of import, export and trade balance of high-tech products in Sweden, Romania, Ukraine and the Republic of Moldova for 2020–2025. It was found in the study that countries with a high level of investment in scientific research and development and a significant share of employment in knowledge-intensive sectors demonstrate a more developed structure of high-tech production and higher competitiveness in international markets for high-tech products. It was found that Sweden is characterized by a balanced structure of high-tech trade and a high level of innovative activity, while Romania is characterized by positive dynamics of development of high-tech production, which gradually strengthens its export potential. In contrast, Ukraine and the Republic of Moldova demonstrate significant dependence on imports of high-tech products, which indicates limited domestic production and innovation potential in the relevant segment. The correlation analysis confirmed the presence of a close relationship between investments in scientific research, development of human capital and the results of functioning of high-tech production. Based on the integrated analysis, a conceptual model of the transformation of the development strategy of high-tech production systems was formed, which reflects the logical transition from the formation of innovative resources to ensuring production results and strengthening the country's competitiveness in international trade in high-tech products. The practical effect of the study is the possibility of using the obtained results to form an effective state policy to support innovative development, stimulate investment in scientific research and the development of high-tech production, as well as develop strategies to increase the technological competitiveness of countries with transformational economies.
Keywords:	High-Tech Production, Innovative Development, Scientific Research And Development, Knowledge-Intensive Industry, International Trade In High-Tech Products, Economic Competitiveness.
DOI:	https://doi.org/10.5281/zenodo.19593436
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	EFFECT OF EMBEDDING MODELS IN RAG SYSTEM FOR PEDAGOGICAL QUESTION GENERATION
Author:	JAYASHREE GANESHKUMAR, M. KRISHNAVENI
Abstract:	The need for cognitively aligned assessments has highlighted the disadvantages of traditional question generation systems, which often fail in pedagogical structures such as Bloom’s taxonomy and suffer from factual hallucinations. This study aims to develop an automated question generation system that integrates a pedagogical framework using a Retrieval-Augmented Generation (RAG) approach. This work uses three open-source embedding models – MiniLM-L6, MiniLM-L12 and msmarco-distilbert to encode educational content and retrieve relevant information using FAISS indices. Retrieved content is then used by T5-based generator, guided by Bloom’s taxonomy-specific prompts, to produce pedagogically aligned questions which are then evaluated using standard NLP metrics and validated against manually created ground-truth dataset. This study provides a comparison of embedding models for pedagogically aligned question generation within RAG framework. Findings show that MiniLM-L12 outperforms other embedding models across all the levels of Bloom’s taxonomy. The results suggest that educators can use the RAG-based question generation to reduce assessment design workload while ensuring generation of cognitively appropriate questions. Future research will explore larger datasets, fine-tuning strategies, multi-domain scalability to further advance pedagogically aligned automated assessment tools. This work contributes to enhanced learning outcomes and broader educational impact.
Keywords:	Retrieval-Augmented Generation, Bloom’s Taxonomy, Educational Technology, Automated Assessment, Question Generation.
DOI:	https://doi.org/10.5281/zenodo.19593463
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	A SIMULATION-BASED OPTIMIZATION FRAMEWORK FOR BALANCING EFFICIENCY AND INTRA-GROUP EQUITY IN HEALTHCARE RESOURCE ALLOCATION
Author:	AHLAM BAHARI, LINA ABOUELJINANE, MARIA LEBBAR
Abstract:	Purpose: Efficient resource allocation is fundamental to mitigating Emergency Department (ED) congestion. However, global optimization often masks significant service disparities across patient groups. This study investigates the complex "Efficiency-Equity-Cost" trilemma by developing a dual-phase optimization framework designed to minimize boarding times and budgetary expenditures while explicitly enforcing intra-group equity across different patient severity levels. Methodology: A Discrete Event Simulation (DES) model was developed to capture the stochastic dynamics of patient flow across an ED and three specialized Inpatient Units (IUs). The study employs a Simulation-Based Optimization (SBO) approach using the ε-constraint method to identify Pareto-optimal trade-offs between budget and efficiency. In the second phase, an intra-group equity constraint was integrated to stabilize service performance within each Emergency Severity Index (ESI) category. Findings: Results indicate that systemic boarding reduction is primarily achieved by expanding capacity in downstream IUs rather than the ED itself, identifying these units as the primary structural bottlenecks. While efficiency-only optimization yielded significant performance gains, the introduction of equity constraints revealed a critical feasibility ceiling. However, a "sweet spot" was identified at a budget of 150 units, where substantial improvements in equity are achieved with negligible impacts on global system efficiency. Originality/Value: This research contributes a robust decision-support tool that quantifies the "Price of Fairness" in healthcare. By identifying the specific budget thresholds where equity and efficiency converge, this work provides hospital administrators with a theoretically grounded and practically viable strategy for ethically aligned resource allocation in high-pressure stochastic environments.
Keywords:	Emergency Department, Inpatient Unit, Simulation-Based Optimization, Multi-Objective Approach, Boarding Time, Equity.
DOI:	https://doi.org/10.5281/zenodo.19593719
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	RSNT-SSAE: A HYBRID DEEP LEARNING AND SPARSE FEATURE LEARNING FRAMEWORK FOR ROBUST ORANGE FRUIT DISEASE DETECTION AND CLASSIFICATION
Author:	C. MAHALAKSHMI, Dr. S. SARAVANAN
Abstract:	Accurate and timely detection of fruit disease is essential for maintaining crop quality and minimizing economic losses in citrus production. Orange fruits are susceptible to various diseases that directly affect their surface appearance, leading to reduced market value and increased post-harvest losses. Manual inspection methods are often time–consuming, subjective, and impractical for large–scale applications. To address these challenges, this paper presents an automated framework for detecting and classifying orange fruit diseases based on deep learning and machine learning techniques. The proposed model integrates a ResNet50-based deep feature extraction approach with a Stacked Sparse Auto Encoder (SSAE) classifier, referred to as the RSNT -SSAE framework. Before classification, orange fruit images undergo Wiener filter – based noise removal to suppress image noise, followed by Contrast Limited Adaptive Histogram Equalization (CLAHE) to enhance visual contrast. Color – based segmentation using RGB thresholding is employed to isolate disease – affected regions from the fruit surface. Deep features extracted using the ResNet50 architecture are then classified using the Stacked Sparse Auto-Encoder (SSAE) model to distinguish between healthy and diseased fruit categories. Experimental evaluation conducted on a publicly available orange fruit dataset demonstrates the effectiveness of the proposed approach, achieving a classification accuracy of 98.10%. The results indicate that the RSNT – SSAE framework provides robust and reliable performance, making it a promising solution for automated orange fruit disease diagnosis and precision agriculture applications.
Keywords:	Image processing, orange fruits, Deep learning, ResNet50, CLAHE, Machine Learning.
DOI:	https://doi.org/10.5281/zenodo.19593742
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	AN ADAPTIVE HYBRID META-LEARNING FRAMEWORK FOR CONTEXT-AWARE INVENTORY FORECASTING AND OPTIMIZATION
Author:	MOHAMED ALI, THANAA MOHAMED, SAHAR EL TAYEB3, MOHAMED ROUSHDY
Abstract:	Accurate inventory forecasting remains challenging in dynamic market conditions, where demand patterns are shaped by external factors such as competitor pricing, promotional events, and supply volatility. Traditional forecasting models often treat algorithms as static, ignoring contextual adaptation and the integration of forecasting with operational decision-making. This paper presents an adaptive hybrid intelligent framework (AHIF) that dynamically integrates XGBoost, LightGBM, and CatBoost within a learning-driven selection architecture, enabling context-aware demand forecasting across different product categories, time frames, and market conditions. Unlike existing approaches, AHIF unifies predictive analytics with an adaptive economic order quantity (EOQ) model that adjusts purchasing strategies. Empirical evaluation on a multi-sector granular dataset shows strong predictive performance, achieving R² values between 0.985 and 0.987 with reduced prediction errors across multiple metrics, including MAE, RMSE, and RMSLE. This framework effectively reduces stockouts and overstocking, highlighting its practical impact on inventory management. The theoretical contribution lies in extending hybrid gradient boosting strategies to a generalizable and context-adaptive decision support framework that bridges the gap between prediction accuracy and inventory optimization. The practical contribution lies in providing a deployable system that enables companies to make adaptive, data-driven purchasing decisions under uncertainty.
Keywords:	Inventory Management, Demand Forecasting, Machine Learning, XGBoost, LightGBM, CatBoost.
DOI:	https://doi.org/10.5281/zenodo.19593756
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	A TOPIC–SENTIMENT–INTEGRATED NLP FRAMEWORK FOR CTQ-BASED EARLY DETECTION IN THE GLOBAL SHIPPING MARKET
Author:	KIM, DONGWON , KIM, YEONJOO
Abstract:	This study develops an integrated natural language processing (NLP) framework for the early detection of Critical-to-Quality (CTQ) risks in the global shipping market. The shipping industry has recently experienced heightened uncertainty driven by geopolitical disruptions, climate-related constraints, and persistent supply–demand imbalances. Under such conditions, conventional quantitative indicators often fail to capture emerging quality risks that are embedded in market narratives and stakeholder perceptions. This study addresses this gap by transforming unstructured shipping-related text into actionable, forward-looking quality risk indicators. The empirical analysis is based on 155 weekly maritime market reports published by the Korea Maritime Institute (KMI) from January 2022 to June 2025. The proposed framework integrates sentiment analysis, topic modeling, and quantitative forecasting. A domain-tuned sentiment lexicon is constructed by extending the Loughran–McDonald financial dictionary with shipping-specific terminology, enabling more accurate interpretation of industry-contextual sentiment. Using ProdLDA-based topic modeling and expert validation, six core CTQ dimensions are identified: freight rate stability, schedule reliability, lead time, vessel utilization, equipment availability, and eco-efficiency. Topic–sentiment relationships are quantified using pointwise mutual information (PMI) and network centrality measures, and composite CTQScores are derived by estimating topic-to-CTQ impact weights through ElasticNet regression combined with Bayesian linear modeling. The predictive performance of CTQScores is evaluated using ElasticNet and vector autoregression (VAR) models to examine co-movement and lead–lag relationships with key performance indicators, particularly the Shanghai Containerized Freight Index (SCFI). The results demonstrate that CTQScores provide statistically significant early signals of market instability. The integrated model achieves strong explanatory power (R² ≈ 0.68) and high directional accuracy (hit ratio ≈ 0.74) for one-week-ahead SCFI movements, outperforming traditional time-series benchmarks. These findings highlight the value of domain-tuned NLP approaches in enhancing early-warning systems and supporting proactive quality risk management in the global shipping industry.
Keywords:	Critical-To-Quality (CTQ), Natural Language Processing (NLP), Sentiment Analysis, Shipping Market Early Warning System, Topic Modeling.
DOI:	https://doi.org/10.5281/zenodo.19593790
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	FEDERATED INTELLIGENCE IN OPHTHALMOLOGY: PRIVACY-PRESERVING COLLABORATION FOR MULTICENTER AMBLYOPIA MODEL DEVELOPMENT
Author:	JAYA LAKSHMI C , DR. N. SATHEESH , DR. M. KUMARASAN
Abstract:	The paper under consideration introduces a federated system of deep-learning to support privacy-safe and multicenter cooperation in the automated process of diagnosis and prognosis of amblyopia. Its key objective will be to create a standard, multimodal ophthalmic corpus (fusing fundus-photography, optical coherence tomography (OCT), eye-tracking signals, visual induced potentials (VEP), and demographic data). Through generic adversarial networks (GANs) augmentation and harmonization, such a layer of data acts out to reduce class imbalance and boost the generalizability of downstream models. The second goal is focused on the advanced feature representation based on the work of a hybrid Convolutional Neural Network (CNN)-Transformer structure. This type of architecture is able to encode both spatial and temporal relations both in imaging and behavioral modalities simultaneously, and therefore forms more comprehensive multimodal embeddings that enhance the discrimination of disease. The explainable artificial intelligence (XAI) methods such as Grad-CAM and SHAP are included in the tertiary objective to enhance the interpretability and clinical transparency of diagnostic predictions. Federated learning (FedAvg, FedProx) allows learning model optimization among participating institutions without the need to share their data, hence, protecting patient privacy and regulating compliance with regulations. Experimental confirmation using synthetic and real-world pediatric datasets have shown a higher diagnostic accuracy (AUC > 0.95), high across-center generalizability and readable visual reasoning maps to aid the decision-making of clinicians. The given work, thus, predetermines the course of safe, transparent, and cooperative, AI-driven ophthalmic diagnostics, accelerating the process of clinical implementation and universal transfer of amblyopia screening.
Keywords:	Federated Learning, Amblyopia Diagnosis, Multimodal Deep Learning, Explainable AI (XAI), Privacy-Preserving Collaboration, Ophthalmic Intelligence.
DOI:	https://doi.org/10.5281/zenodo.19593814
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	A SECURE AND IMPERCEPTIBLE IMAGE STEGANOGRAPHY SCHEME BASED ON LSB-NEW
Author:	R.GANESH , DR.S.THABASUKANNAN
Abstract:	Image steganography plays a vital role in secure multimedia communication by concealing sensitive information within digital images. Conventional least significant bit (LSB) techniques provide high embedding capacity but often suffer from visual distortion and weak resistance to analysis under increased payload conditions. This paper presents an enhanced adaptive LSB-NEW technique for secure image-in-image data hiding, integrating secret image encryption with pixel-aware embedding [1]. The proposed method adaptively selects embedding locations based on local intensity characteristics, thereby improving imperceptibility while maintaining computational efficiency. The significance of this work lies in its ability to achieve a practical balance between security, embedding capacity, and visual quality without introducing high computational complexity, making it suitable for real-world applications such as secure image transmission and digital content protection. Experimental evaluation using standard benchmark images demonstrates that the proposed approach consistently achieves higher PSNR and SSIM values with lower MSE compared to the conventional LSB method. Visual analysis further confirms minimal perceptual distortion. These results indicate that the proposed LSB-NEW technique offers an efficient and scalable solution for secure image-in-image data hiding in modern communication systems.
Keywords:	Image Steganography, LSB-NEW, Image-In-Image Hiding, PSNR, MSE, SSIM, Encryption
DOI:	https://doi.org/10.5281/zenodo.19593832
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	IDENTIFYING A TYPOLOGY OF MOROCCAN HEALTH PROVINCES IN URBAN AREAS BASED ON MATERNAL AND CHILD HEALTH INDICATORS USING A CLUSTERING APPROACH
Author:	MERYEM CHAKKOUCH, HICHAM SADIKI, MEROUANE ERTEL, AZIZ MENGAD, SAID AMALI, ABDELMALEK AZZOUZI
Abstract:	Maternal and child health remains a public health priority, particularly given significant spatial disparities in healthcare accessibility and outcomes. It is essential to clearly identify these differences in order to refine intervention strategies and optimize resource allocation. In this article, we propose a clustering approach to group and identify typologies of Moroccan health provinces based on MCH indicators. We conducted a comparative analysis to evaluate five dimensionality reduction (DR) techniques, including PCA, ICA, and nonlinear methods (KPCA, LE, Isomap), combined with three clustering methods (K-Means, agglomeration clustering, GMM) on Moroccan urban areas data characterized by a high dimensionality, low sample size (HDLSS) structure on MCH indicators. The model selection was guided by internal validation indicators (Silhouette, Calinski-Harabasz, Davies-Bouldin). Laplacian Eigenmaps followed by K-means achieves superior performance measures reflecting the most robust and consistent clustering structure. Three clusters of provinces were identified in this comparative study: moderate-performing provinces characterized by the presence of both public and private urban healthcare systems; high-performing provinces with robust urban public services; and lower-performing provinces due to rural dependence on urban facilities and resource constraints. The main contribution of this study is to identify disparities in maternal and child health across Moroccan provinces/prefectures using urban-area indicators and a clustering-based approach. It thus makes it possible to propose a territorial typology that can support public health interventions and improve the strategic allocation of resources.
Keywords:	Clustering; PCA, ICA, KPCA, LE, Isomap, K-Means, Agglomerative Clustering, GMM, MCH Indicators
DOI:	https://doi.org/10.5281/zenodo.19593857
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	TEXT-TO-SPEECH (TTS) ANALYSIS SYSTEM WITH CLIENT-SIDE PROCESSING
Author:	TASKEOW SRISOD, PRACHYANUN NILSOOK, SASITORN ISSARO, ORAPHAN AMNUAYSIN3 THANANAN AREEPONG, ORAWAN SAEUNG, THANI JINTASUTTISAK, THAMASAN SUWANROJ
Abstract:	In todays digital age, text-to-speech (TTS) technology has become a crucial tool for increasing accessibility and enhancing user experiences across platforms, from devices for the visually impaired to smart assistants. Developing this technology efficiently and quickly is essential. This article focuses on developing a text-to-speech analysis and synthesis system using client-side processing technology, an approach that enables TTS conversion to occur directly on a user's web browser, thereby reducing server load and increasing response speed. The work covers everything from the process of TTS, user interface (UI) development, to Web Speech API implementation. Furthermore, to ensure the quality of the synthesized voices, a systematic evaluation was conducted using the internationally-standardized Mean Opinion Score (MOS) for Thai voices from Microsoft client-side TTS voices, namely Pattara and Kanya, to measure clarity, naturalness, and fluidity. The results of this project not only serve as a prototype for an effective TTS system, but also provide valuable insights for the future development of synthetic voices that are more natural and closely approximate human speech.
Keywords:	Text to Speech (TTS), Client-Side Processing, Natural Language Processing (NLP)
DOI:	https://doi.org/10.5281/zenodo.19593882
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	AN INTEGRATED REDUNDANT RELIABILITY OPTIMIZATION FRAMEWORK FOR COHERENT HYBRID SYSTEMS USING LAGRANGE MULTIPLIER, HEURISTIC, AND INTEGER PROGRAMMING APPROACHES
Author:	SAI UMA SANKAR MANDAVILLI, SIDHAR AKIRI, SRINIVASA RAO VELAMPUDI, ARUN KUMAR SARIPALLI, BHAVANI KAPU, RAMADEVI SURAPATI
Abstract:	The overarching purpose of reliability engineering is to guarantee, within a certain time frame and set of circumstances, the proper operation of systems and its components. Integrating redundancy strategically while taking restrictions like mass, cost, and space needs into coherent system topologies improves system robustness, according to reliability theory. The impact of these limitations volume, weight, dimensions, and space on the improvement of system reliability is the primary emphasis of this study. Parts used in a typical representative coherent hybrid system are the focus of the inquiry. These parts are really taken from coherent systems, where dimensions, mass, and cost are critical for maintaining effective operation. The motor control unit, internal combustion engine, and electric generator must all go through thorough reliability testing. To solve this problem, we need to research and construct an integrated redundant reliability model in a coherent systems structure utilizing the Lagrange multiplier approach. This approach generates continuous-valued results for significant variables, including component count, individual and stage reliabilities, and overall system dependability. The work takes it a step further by using heuristic methods and integer programming to come up with integer-based solutions that are both realistic and useful. These strategies make the dependability research much more accurate and valuable.
Keywords:	IRR Model, Coherent Systems, System Reliability, LAM Approach, HAM Approach, IP Approach
DOI:	https://doi.org/10.5281/zenodo.19593900
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	SPEECH EMOTION ANALYSIS ON MULTIPLE DATASETS USING OPTIMIZED DEEP LEARNING MODEL
Author:	P.SUBHADASREE, D.USHARANI
Abstract:	Emotions are fundamental to human communication and play a critical role in guiding both rational behavior and social interaction. Speech Emotion Recognition (SER) aims to automatically identify human emotions from spoken audio, which is a challenging task due to the variability in speech patterns and the subtle acoustic differences between emotional states, SER are widely used in many real world applications including health care, stress treatment, ICU monitoring and other surveillance. This paper presents a Long Short-Term Memory (LSTM) based approach, enhanced through hyperparameter optimization, for accurate classification of speech emotions. The model is trained and evaluated on multiple benchmark datasets, including RAVDESS, KBES, nEmo, VESUS, and BANSpEmo, each offering diverse linguistic and emotional content. Key acoustic features such as pitch, energy, and Mel-frequency cepstral coefficients (MFCCs) are extracted and used to train the model. Unlike traditional Convolutional Neural Network (CNN) methods that focus on spatial features, the proposed LSTM architecture effectively captures temporal dependencies within the audio signals. Emotions such as happiness, surprise, anger, sadness, and neutrality are classified with high accuracy. Experimental results indicate that the proposed LSTM-based model, optimized via hyperparameter tuning, outperforms baseline methods and demonstrates improved generalization across datasets, making it a promising solution for real-world emotion recognition applications.
Keywords:	Artificial Intelligence, Machine Learning, Deep Learning, Grid Search Cross Validation, Random Forest, Hyper Parameter Tuning, Best Fit Parameters.
DOI:	https://doi.org/10.5281/zenodo.19593942
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	MODIFIED SHUFFLENET WITH TRANSFER LEARNING MODEL FOR BLOCKCHAIN ASSISTED SECURE LUNG CANCER CLASSIFICATION IN CLOUD-IOT ENVIRONMENT
Author:	AFFROSE, ASHWANI KUMAR YADAV, AND ARCHEK PRAVEEN KUMAR, AND CHERUKU SANDESH KUMAR
Abstract:	Globally, lung cancer (LC) is the much prevalent source of cancer-based deceases. In recent years, classifying lung cancer using CT has become a new area of study in the area of medical image systems. A key factor in the LC diagnosis is the capacity to precisely determine the location and size of the disease. Rapid diagnosis, detection, classification, and evaluation of CT images are therefore necessary. The developed study is made to classify the LC using CT imageries. For classifying LC, the data (CT images) will be acquired from Cloud-IoT. The acquired data is stored in Blockchain to ensure its security. After acquiring the data, the LCC process begins by pre-processing the CT images using Processed Pixel-based Threshold-based Median Filtering (PPT-MF) technique. Subsequently, modified Res U-Net is used for segmenting the images. Then, features namely, Modified MRELBP, Local Monotonic Pattern (LMP), and PHOG are derived. After that, data augmentation is carried out using random sampling technique. Finally, classification is done using Transfer learning-based modified Shufflenet (TL-MSNet) model. From the analysis, a higher specificity of 98% is gained with TL-MSNet, whereas, existing methods score lower specificity values.
Keywords:	Lung cancer; PPT-MF Technique; Modified Res U-Net; Modified MRELBP; Modified Shufflenet.
DOI:	https://doi.org/10.5281/zenodo.19593959
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	ITERATIVE LLM-GUIDED RESIDUAL REASONING FOR SPEECH ENHANCEMENT IN REAL-WORLD NOISY ENVIRONMENTS
Author:	AWS I. ABUEID
Abstract:	Despite significant advances in speech enhancement, most existing approaches rely on single-pass processing and closed black-box models, limiting progressive refinement, transparency, and reproducibility. This paper presents a fully local and reproducible Iterative LLM-Guided Speech Enhancement Framework that reformulates denoising as a multi-step residual reasoning process operating on log-Mel spectral representations. Unlike traditional single-pass digital signal processing methods, the proposed approach employs a 14B-parameter open-weight transformer (Qwen2.5) to perform structured residual spectral refinement across multiple iterations, enabling progressive noise suppression while preserving harmonic continuity and formant structure under real-world acoustic conditions. The framework was evaluated on 45 paired recordings collected across three environments: clean indoor speech, outdoor street noise, and dynamic walk-and-talk conditions. Quantitative evaluation over 30 noisy samples demonstrates that moderate refinement depth (K = 2) achieves optimal performance, improving SNR from 3.1 dB to 10.4 dB, PESQ from 1.42 to 2.21, and STOI from 0.61 to 0.79. Semantic assessment using a locally executed Whisper ASR model shows a reduction in mean Word Error Rate (WER) from 1.534 (noisy) to 1.507 at K = 2, confirming improved linguistic recoverability. Deeper refinement (K = 4) yields diminishing returns, indicating stable convergence of the residual update mechanism. All experiments were conducted entirely offline on a consumer-grade RTX 3060 GPU (12GB VRAM), demonstrating controlled iterative convergence under constrained hardware resources. These findings establish open-weight transformer-based residual reasoning as a scalable and transparent paradigm for intelligibility-centred speech enhancement in realistic environments.
Keywords:	Speech Enhancement, Large Language Models, Iterative Refinement, Semantic Intelligibility, Residual Learning.
DOI:	https://doi.org/10.5281/zenodo.19593981
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	STTN-CP: A SPATIAL-TEMPORAL TRANSFORMER WITH CONTRASTIVE PRETRAINING MODEL FOR CREDIT CARD FRAUD DETECTION
Author:	KATHIRESAN JAYABALAN , SETHURAMAN RADHAKRISHNAN
Abstract:	Credit card fraud detection remains a critical challenge due to highly imbalanced datasets, evolving fraud patterns, and complex transaction behaviors, leading to significant financial losses and reduced user trust. A novel Spatial-Temporal Transformer Network with Contrastive Pretraining (STTN-CP) is proposed to effectively detect fraudulent transactions. The STTN-CP model is designed to capture temporal dependencies throughout the entire transaction sequence and to identify spatial correlations among the transaction features. This study introduces a unified framework that integrates spatial–temporal transformer modeling with contrastive representation learning to enhance feature discrimination and detection accuracy. A two-step methodology is employed to tackle the class imbalance issue commonly found in credit card datasets. Initially, Min-Max normalization is conducted, followed by the application of the Synthetic Minority Oversampling Technique (SMOTE) in the data preprocessing phase. Subsequently, the spatial-temporal transformer blocks collaboratively generate hierarchical embeddings, while the contrastive pretraining module enhances feature discrimination by clustering analogous transactions and separating dissimilar ones inside the feature space. The proposed framework is evaluated using the Credit Card Fraud Detection dataset that is publicly available from Kaggle, and achieves an accuracy of 99.12%, precision 99.00%, recall 98.86%, F1-score 98.92%, and specificity 97.96%. The ablation studies show the importance of spatio-temporal modeling, contrastive pretraining, and data preprocessing with respect to high detection performance. The proposed model provides a scalable and robust solution for real-time fraud detection, improving financial security and decision-making reliability. Future work will focus on multi-class fraud detection, lightweight architectures for real-time deployment, and explainable AI integration.
Keywords:	Credit Card Fraud Detection; Deep Learning; Spatial–Temporal Transformer; Contrastive Pretraining; SMOTE
DOI:	https://doi.org/10.5281/zenodo.19593993
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	DIGITAL ECOSYSTEM OF THE STATE AS A DRIVER OF PROFESSIONALIZATION OF PUBLIC SERVICE: THE EXPERIENCE OF UKRAINE AND EU COUNTRIES
Author:	OKSANA EVSYUKOVA, NATALIIA OBUSHNA, TETIANA VASYLEVSKA, IRYNA DYNNYK, IRYNA LAZEBNA, VIKTORIIA PETRYNSKA
Abstract:	In the context of the public administration transformation in Ukraine, digital public services are becoming a determining factor in increasing the transparency, efficiency and professionalism of the public service, providing new formats of interaction between the state and citizens, optimizing management processes, contributing to the formation of modern competencies of civil servants and creating conditions for the transition to a service- oriented model of the state. The need for research is due to the fact that existing scientific approaches consider digitalization mainly as a tool for increasing management efficiency, not sufficiently taking into account its systemic impact on the professionalization of the public service. In the context of active digital transformation in Ukraine, a comprehensive analysis of the relationship between the development of digital services, the level of digital competencies and the institutional capacity of the public service is relevant. The purpose of the article is to explore the potential and effectiveness of digital public services as a tool for increasing the professionalism of the public service, to determine their impact on the competence, effectiveness and openness of the public service, and to outline promising ways of their improvement. The study uses systemic, institutional and structural-functional approaches, as well as analysis, synthesis, comparison, content analysis of official digital platforms, analysis of statistical data and monitoring of digital governance practices. As a result of the study, it was established that digital public services are a system-forming factor in the professionalization of the public service of Ukraine. It is proven that the introduction of digital tools, digital ethics and digital communications forms new competencies of civil servants, expands the range of managerial roles and increases the need for digital literacy, data management, cybersecurity and electronic systems. A comparison of the level of digital skills of citizens and e - government development indicators revealed interdependence: a higher level of competence of the population ensures a faster implementation of digital services, and electronic services themselves stimulate further growth of digital literacy of society and public service employees. The analysis of public investment in digitalization has shown that the effectiveness of the digital services development depends not only on the amount of funding, but also on the ability of the state to integrate digital solutions into personnel policy, educational programs and the management system. The article demonstrates that digital public services are not just automation tools, but a fundamental mechanism for the professionalization of the public service. Their implementation contributes to the renewal of the competencies of officials, improvement of management, improvement of the quality of public services and the formation of a service- oriented model of the state. This ensures sustainable progress in reforming Ukraine's public service and its compliance with European standards.
Keywords:	Digital Services; Digital Ecosystem Of The State; Digital Ethics; Digital Communications; Public Service; Administrative Services; Professionalization; Digital Competencies; E-Government; ICT Sector; Digitalization; Innovation; Digital Transformation; Service- Oriented State.
DOI:	https://doi.org/10.5281/zenodo.19594018
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	A NOVEL META-REINFORCEMENT FAST LEARNING NETWORK FOR AN ENERGY EFFICIENT FOG ENABLED IOT ENVIRONMENT
Author:	G. PRABHAKAR, MEERAVALI SHAIK
Abstract:	Internet Enabled Devices has opened its wide dimension of applications in the multiple areas such as healthcare, transportation and automation systems. As the computing devices increases, Internet connected devices generates the more number of data which leads to the scarcity of energy and computation. Fog computing has emerged as an innovative paradigm to enhance these devices in terms of energy and time by bringing the computation closer to the data sources. However, a dynamic characteristics of the networks and heterogeneous data in the Internet enabled network poses the serious challenge that affects the network performance to achieve the low latency and energy efficiency simultaneously. To overcome this limitations, this research introduces a novel meta-reinforcement fast learning algorithm for energy-efficient routing and scheduling in the Fog enabled IoT environment. The suggested framework integrates the principle of reinforcement learning and fast neural networks to enable rapid adaptation of routing policies and optimizes the networks by minimizing the transmission latency and energy consumption in the Fog gateways. The framework was deployed in the flexible python based environment (FogBus2 and SimPy) and demonstrated its strong performance by evaluating the quality of services(QoS) metrics such as average latency, energy and resource utilization process. To prove the efficacy of the suggested reinforced technique, its performance was compared with the other techniques. Evaluation results demonstrates that the suggested model balances network characteristics and its performance by achieving the lower transmission latency(54% lesser than traditional methods),consuming only 40% of energy than the conventional learning techniques. The suggested research provides the better insights for scalable and intelligent solutions for an energy efficient Fog computing systems.
Keywords:	Fog Computing, Meta-Reinforcement, Fast Learning Networks,Internet Enabled Devices, Quality Of Services.
DOI:	https://doi.org/10.5281/zenodo.19594039
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	A UNIFIED BENCHMARKING FRAMEWORK FOR CLASSICAL AND QAOA-BASED MULTI-OBJECTIVE ROUTING IN COMMUNICATION NETWORKS
Author:	KALAVATHI S , BALAMURUGAN N M
Abstract:	Multi-objective routing in communication networks requires simultaneous optimization of delay, congestion, and reliability—posing challenges for classical single-metric routing protocols. This paper presents a rigorous quantitative benchmarking framework comparing classical deterministic routing paradigms, namely Routing Information Protocol and Open Shortest Path First, against a quantum-assisted routing formulation based on the Quantum Approximate Optimization Algorithm (QAOA). The routing problem is encoded as a constrained Quadratic Unconstrained Binary Optimization (QUBO) model incorporating flow-conservation and path-continuity constraints. Random connected weighted graphs (10–20 nodes) were evaluated across 50 independent runs per topology size under a unified scalarized cost function (0.5D + 0.3C + 0.2L). Results demonstrate that deterministic link-state routing achieves optimal path selection with minimal computational complexity O(m log n), while QAOA attains near-optimal solutions (3–8% deviation) at significantly higher computational overhead due to qubit scaling and variational optimization loops. Scalability analysis reveals quadratic qubit growth with network size, limiting NISQ-era applicability. The study establishes a structured comparative benchmark and positions quantum routing as a complementary exploratory mechanism rather than a replacement for classical routing.
Keywords:	Multi-objective routing, QAOA, QUBO, RIP, OSPF, quantum networking, combinatorial optimization, SDN.
DOI:	https://doi.org/10.5281/zenodo.19594057
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	UNDERSTANDING DIGITAL ADOPTION AMONG INDONESIAN MSMES: THE ROLE OF BEHAVIOURAL BIASES, FINANCIAL LITERACY, AND HEDONIC MOTIVATION
Author:	KIMBERLY IRENE MANGITUNG, CAREN IMANUEL, FRANSISCA HANITA RUSGOWANTO
Abstract:	This study is quantitative research that examines the role of financial literacy and behavioural biases among MSME actors in Indonesia in shaping decisions related to digital adoption. This study uses data from 180 MSME owners and managers operating in Indonesia and applies Partial Least Squares Structural Equation Modelling (PLS-SEM) for data-analysis. The findings show that financial literacy has a positive influence on digital adoption, while heuristics, loss aversion, and overconfidence have negative effects on digital adoption. Hedonic motivation, as a moderating variable, strengthens the relationship between financial literacy and digital adoption and reduces the negative effects of heuristics and overconfidence. However, hedonic motivation does not reduce the influence of loss aversion on digital adoption. This study contributes to the literature by developing an integrated framework that combines financial literacy, behavioural biases, and hedonic motivation in explaining digital adoption decisions among MSMEs, while extending behavioural research through the identification of the moderating role of hedonic motivation in shaping bias-driven decision-making, and providing empirical evidence from the Indonesian MSME context, where such integrated analyses remain limited. Digitalisation should be accompanied by improvements in financial literacy and the development of user-friendly and enjoyable digital systems.
Keywords:	Digital adoption, Behavioural Biases, Financial Literacy, MSME, Hedonic Motivation
DOI:	https://doi.org/10.5281/zenodo.19594079
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	ML-ER-OPE: JOINT PREDICTION OF SOFTWARE DEVELOPMENT EFFORT AND RELIABILITY VIA OPERATIONAL PROFILE–DRIVEN MULTI-TASK LEARNING
Author:	SUDHAKAR KAMBHAMPATI, Dr. G VAMSI KRISHNA
Abstract:	Precise estimation of software development effort and software reliability remains a significant challenge in today’s non-stationary workloads and heterogeneous patterns of use. Currently, these tasks are generally modelled independently and based on static or past failure data, such that these methods do not provide optimal performance when the system is operating normally. In this work, we introduce an operational-profile-driven multi-task learning model, ML-ER-OPE, for the simultaneous prediction of development effort and software reliability. The approach builds dynamic operational profiles from execution logs to derive scenario-level execution probabilities that behaviorally scale structural metrics and reliability signals. The features are then pooled and concatenated, and the multi-modal representation is fed into a shared deep encoder with task-specific regression heads (under a composite objective), which together make feature-level consolidation and associated model optimization. A flexible optimization engine also adjusts the level of testing and resource distributions based on estimated risk. The framework is tested on four NASA benchmark datasets (JM1, KC1, KC2, PC1) using repeated stratified 10-fold cross-validation. ML-ER-OPE significantly outperforms state-of-the-art baselines in terms of effort and reliability, where the null hypothesis is rejected using paired t-tests and Wilcoxon tests at a 95% confidence level, as well as medium-to-large effect sizes
Keywords:	Software Effort Estimation, Software Reliability Prediction, Multi-Task Learning, Operational Profiles, Behavior-Aware Modeling, Joint Effort–Reliability Modeling, Dynamic Resource Optimization, Empirical Software Analytics.
DOI:	https://doi.org/10.5281/zenodo.19594101
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	ENHANCING DEPENDENCY PARSING FOR TELUGU-ENGLISH CODE-MIXED TEXT: TREEBANK CREATION, PARSER ADAPTATIONS AND POS TAGGING INTEGRATION
Author:	SANDEEP MADDU , VIZIANANDA ROW SANAPALA
Abstract:	Code-mixed text from social media poses significant challenges for syntactic analysis due to irregular grammar, non-standard usage, and frequent language switching. For Telugu-English code-mixed text, the absence of large-scale syntactic resources and specialized parsing models limits progress in downstream multilingual NLP applications. In this work, we address this gap by introducing the first substantial manually annotated Telugu-English code-mixed dependency treebank of 4,152 sentences, developed using Universal Dependencies (UD) 2.0 guidelines. We further propose enhancements to a biaffine dependency parser by incorporating a language-aware head-dependent bias and relation-specific structural weights to better capture cross-lingual syntactic patterns. Our approach improves parsing performance, achieving 75.53% UAS and 61.86% LAS, with consistent gains over a strong baseline. In addition, we demonstrate that integrating dependency-derived syntactic features into a BiLSTM-CRF model improves part-of-speech tagging, achieving a macro-F1 score of 83.73%, with statistically validated gains. We also re-annotate an existing Telugu-English dataset using UD 2.0 to ensure compatibility with modern syntactic frameworks. Overall, this work provides new annotated resources and modeling strategies that advance syntactic processing for Telugu-English code-mixed text, with broader implications for developing robust NLP systems in low-resource and multilingual settings.
Keywords:	Dependency Parsing, Code-Mixing, Telugu, Part-Of-Speech Tagging, Language Aware Head Bias, Dependency Aware Weights
DOI:	https://doi.org/10.5281/zenodo.19594110
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	DUAL-THRESHOLD FEATURE SELECTION AND LIGHTWEIGHT ADAPTIVE SOFT-VOTING ENSEMBLE FOR NETWORK INTRUSION DETECTION IN IoT ENVIRONMENT
Author:	VEENA S BADIGER , GOPAL K SHYAM
Abstract:	Network intrusion detection systems (NIDS) is an important tool for protecting networking devices. In an resource constrained network environment such as IoT safeguarding these devices is essential. The use of lightweight NIDS is necessary to find intrusion in resource-constrained IoT environments. Achieving high detection accuracy with real-time efficiency is challenging therefore, a novel two-stage lightweight NIDS is developed to improve intrusion detection in IoT networks. The study presents three important novel components. First Dual-Threshold Random Forest Feature Selection (DT-RFFS) methodology decreases the feature space by 68-71 percent and increases the rate of inference by 12-15. Confidence weighted adaptive soft voting (CW-ASV) strategy adapts the weight of the classifiers based on the performance and improves the minority attack detector by 2.3%. Third Confidence Based Early Exit (CBEE) mechanism leaves earlier high confidence benign traffic with a significant reduction in the computational overhead and a high detection accuracy. Proposed architecture performs binary classification in first stage to distinguish benign traffic from malicious network activity. Subsequently, the second stage identifies specific intrusion categories using Soft Voting ensemble learner consisting Decision Tree, Random Forest, LightGBM, XGBoost, and AdaBoost as Base learners. The dataset is balanced using SMOTE. The experiment was conducted on CIC-ToN-IoT and RT-IoT2022 datasets. CIC-ToN-IoT and RT-IoT2022 experiment results achieved the accuracy of 99.8 and 98.6-99.7 in stage 1 and stage 2 respectively, and a precision, recall and F1-score of 99 percent. The log-loss values are between 0.01 and 0.10, and AUC-ROC values are between 0.9 and 1.0 indicating high reliability in probability prediction and overall class differentiation. Cross dataset testing has a generalization accuracy of 90-91%. The experiment demonstrated inferences at low computation price. Model achieved a latency of less than 20 ms, and a throughput of 690 to 720 packets/second, showing its real-time performance and deployment in resource-constrained IoT network conditions.
Keywords:	Cyber-attacks, Ensemble Learning, Intrusion Detection, IoT Security, Lightweight Classifier, Soft Voting.
DOI:	https://doi.org/10.5281/zenodo.19594128
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	YOUTH SOCIAL MEDIA BEHAVIOR AND ETHICAL INFORMATION VERIFICATION: AN EMPIRICAL STUDY OF THE DIGITAL TABAYYUN MODEL
Author:	ABDUL RAUF RIDZUAN, KHAIRUL AZHAR MEERANGANI, MOHAMMAD FAHMI ABDUL HAMID, MUHAMMAD HILMI MAT JOHAR, ABDUL SATTAR
Abstract:	The rapid growth of social media has intensified concerns about misinformation and the ethical responsibilities of users, particularly among youth who constitute one of the most active online populations. While technological solutions such as automated fact-checking have gained prominence, less attention has been given to the behavioral and ethical dimensions of information verification. This study examines the influence of youth social media behavior on ethical information verification practices and proposes a structured value-based framework to address this gap. Using a quantitative cross-sectional survey design, data were collected from 215 young social media users. Behavioral constructs grounded in the Theory of Planned Behavior; attitude, subjective norm, and perceived behavioral control were employed to examine their relationship with ethical verification practices, which were operationalized through self-control, open-mindedness, critical thinking, and information-seeking behaviors. Descriptive, correlation, and regression analyses were conducted using SPSS. The findings reveal a strong and significant positive relationship between social media behavior and ethical information verification (r = 0.767, p < .01). Regression analysis further indicates that social media behavior is a substantial predictor of verification practices, explaining 59% of the variance (R² = .59). These results demonstrate that ethical verification is not merely a technical or cognitive skill, but a behavioral practice shaped by attitudes, social influence, and perceived capacity for action. Building on these empirical insights, the study introduces the Digital Tabayyun Model as a value-based ethical verification framework that integrates source evaluation, reflective verification behavior, and impact awareness. The model complements existing technological and behavioral approaches by emphasizing the human-centered dimensions of ethical decision-making in digital environments. This study contributes to research in information and education technology by providing empirical evidence on behavioral determinants of ethical verification and offering a culturally responsive, value-based model to support digital literacy and responsible social media engagement among youth.
Keywords:	Ethical Information Verification; Youth Social Media Behavior; Digital Literacy; Theory of Planned Behavior; Digital Tabayyun
DOI:	https://doi.org/10.5281/zenodo.19594147
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	HARNESSING UNCERTAINTY-AWARE AND SCALABLE GMM-RF FOR CROSS-DOMAIN SENTIMENT INTELLIGENCE IN HETEROGENEOUS DIGITAL PLATFORMS
Author:	RAMAJAYAM G , Dr. VIDYABANU R
Abstract:	Accurate sentiment classification across diverse online shopping domains remains a complex challenge due to shifting vocabulary, variable sentiment intensity, and non-linear emotional expression. This paper presents a domain-adaptive framework Gaussian Mixture Model-enhanced Random Forest (GMM-RF)—designed to overcome ambiguity and variability in customer reviews by integrating probabilistic clustering with ensemble learning. The GMM component models review as soft sentiment distributions, capturing nuanced emotional overlaps that are often misrepresented by rigid classifiers. These probabilistic encodings guide the Random Forest classifier, which applies cluster-aware bootstrapping, entropy-filtered sampling, and GMM-driven thresholding to form precise and interpretable decision boundaries. Evaluation on Amazon product reviews—Books, DVDs, Electronics, and Kitchen Appliances—demonstrates an average classification accuracy of 95.705%, marking a 7.28% improvement over MMASA. The model also surpasses EBC with a 9.28% gain, confirming its resilience in sentiment-rich, high-entropy environments. GMM-RF requires no domain-specific retraining and is fully scalable, making it ideal for deployment in real-time e-commerce systems. Its design supports adaptive, interpretable sentiment analytics, enabling more reliable customer insight extraction from dynamic, cross-category review streams.
Keywords:	Sentiment Classification, Gaussian Mixture Model, Random Forest, Cross-Domain Analysis, Online Reviews)
DOI:	https://doi.org/10.5281/zenodo.19594160
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	A CNN-DRIVEN FRAMEWORK FOR ACCURATE DIABETIC FOOT ULCER DETECTION IN CLINICAL IMAGING
Author:	TALASILA VEENA , BABU REDDY MUKKAMALLA
Abstract:	One of the most serious side effects of diabetes mellitus is diabetic foot ulcers (DFUs), which, if left untreated, can result in infection, delayed wound healing, and lower limb amputations. Because manual DFU assessment is subjective, time-consuming, and heavily reliant on clinical knowledge, it is not as scalable in actual healthcare settings. This study suggests a deep learning-based intelligent and automated DFU diagnosis system that improves clinical reliability and detection accuracy in order to address these issues. The suggested approach successfully extracts discriminative features from DFU photos by combining Convolutional Neural Networks (CNNs) with transfer learning architectures. To adjust to DFU-specific visual patterns, pre-trained models like ResNet and EfficientNet are refined. A Generative Adversarial Network (GAN)-based data augmentation technique is used to address the problem of small and unbalanced medical datasets, allowing the creation of realistic DFU images and enhancing model generalization. Benchmark DFU datasets under standardized preparation are used to train and assess the system and protocols for augmentation. According to experimental results, the suggested method outperforms traditional CNN-based and transfer learning baselines with a classification accuracy of 97.8%, sensitivity of 96.9%, specificity of 98.4%, and an AUC score of 0.99. Particularly in minority ulcer classes, the addition of GAN-generated data greatly increases resilience and decreases overfitting. This work's main contribution is the efficient combination of adversarial learning and transfer learning for DFU diagnosis, which provides a scalable, precise, and clinically interpretable approach. By enabling early intervention, lowering clinician workload, and enhancing patient outcomes in diabetic care, this approach has great potential for implementation in computer-aided diagnostic systems.
Keywords:	CNN, DFU, Multitask Deep Learning, Transfer Learning
DOI:	https://doi.org/10.5281/zenodo.19594171
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	AN INTERPRETABLE DEEP LEARNING FRAMEWORK FOR GASTRIC POLYP DETECTION AND SEGMENTATION USING GRAD-CAM AND LIME
Author:	JYOTI GARG , MOHIT CHHABRA
Abstract:	Correct identification and segmentation of precancerous polyps in endoscopic images takes the first steps toward the early detection of stomach cancer. While Deep Learning has greatly improved in the field of medical image processing, clinical trust in such models is limited. This paper develops an explainable deep learning model using a publicly available Kvasir-SEG dataset for the segmentation of stomach cancer polyps. In this methodology, it encompasses preprocessing of images, residual U-Net segmentation, and evaluation of the model using IoU, Dice Score, accuracy, and loss metrics. For interpretability, LIME provides local explanations at the pixel level for individual predictions, whereas Grad-CAM has been used to highlight class-discriminative regions that influence segmentation decisions. The integrated XAI framework in the study offers global interpretability and instance-specific interpretability; therefore, clinicians are able to visually confirm why the model designated a certain area as suspected of carrying cancer. The obtained experimental consequences showed impressive configuration of model-highlighted regions and actual polyp structures, with great segmentation capability. The proposed approach provides relevancy and transparency. The proposed model has been optimized by using multiple optimization models such as Attention UNet, EfficientNet, ResUNet, DenseNet, and Transformer. This integrated ensemble model incorporated with LIME and Grad-CAM produces a deep learning based sudsier that could easily be used for stomach cancer diagnosed through endoscopic analysis.
Keywords:	LIME, Grad-CAM, Gastric Cancer, Polyp Segmentation, Explainable AI(XAI)
DOI:	https://doi.org/10.5281/zenodo.19594184
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text

Title:	TRANSFORMER-BASED CYBERBULLYING DETECTION IN MOROCCAN DARIJA: A WILLARD TAXONOMY APPROACH
Author:	ABDELKARIM TAHIRI , BOUBKER SBIHI, RACHID GOUARTI , FATIMA MOURCHID , ALI EL KSIMI
Abstract:	Cyberbullying remains a serious threat, yet detection tools for Arabic dialects are scarce. Moroccan Darija poses particular challenges for automated systems. This paper evaluates three transformer models on the Offensive Moroccan Comments Dataset. Results show MARBERT achieves 85.07% F1-score, outperforming AraBERT (83.84%) and the multilingual baseline (80.10%), confirming that dialectal pre-training matters for low-resource varieties. We also propose a severity framework based on Willard's (2007) cyberbullying taxonomy, which distinguishes eight behavioral types. Our system maps these into three levels: CRITICAL (cyberstalking, harassment, outing) for immediate action, MODERATE (flaming, denigration, exclusion, trickery, impersonation) for standard review, and NONE for safe content. Dataset analysis shows 9.6% of comments fall into the CRITICAL category. On this high-risk class, MARBERT reaches 86.05% F1-score with 84.80% recall, ensuring most dangerous content gets flagged. These results offer practical guidance for deploying content moderation systems for Arabic communities.
Keywords:	Cyberbullying Detection, Arabic NLP, Willard Taxonomy, Transformer Models, Severity Classification
DOI:	https://doi.org/10.5281/zenodo.19594204
Source:	Journal of Theoretical and Applied Information Technology 15^th April 2026 -- Vol. 104. No. 7-- 2026
Full Text