		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
March 2026 | Vol. 104 No.6

Title:	THE ROLE OF DIGITAL LITERACY IN STRENGTHENING CIVIL SERVANTS’ COMPETENCIES AND PUBLIC GOVERNANCE OUTCOMES
Author:	IHOR LUKIANENKO, TETIANA NYCH, YEVHEN HREBONOZHKO, MAKSYM SIKALO, LIUDMYLA TREBYK
Abstract:	Digital transformation calls for the enhancement and cultivation of public servants’ competencies, thereby enabling them to more fully realize their intellectual potential and contributing to the overall improvement of the public administration’s efficacy. The aim of this study was to assess the influence of the E-Government Development Index and the E-Participation Index on public government’s effectiveness at a global scale. The research applied the methods of regression and correlation analyses, as well as ANOVA testing. Utilizing panel data from 190 countries during 2014-2024, the study examined the E-Government Development Index, E-Participation Index, and Government Effectiveness. The findings suggest that in nations exhibiting a high degree of e-government advancement and active e-participation, the digital literacy among public servants and citizens is correspondingly elevated. This observation substantiates the rationale for employing the E-Government Development and E-Participation indices as indicators that indirectly reflect the level of digital literacy within both civil servants and the general public. The study revealed a close positive correlation among all indicators of e-governance, e-participation, and governance efficacy. The strongest correlations were observed between the E-Participation Index and the Online Service Index (correlation coefficient 0.95), followed by the E-Government Development Index and the E-Participation Index (0.84), and the E-Government Development Index and the Government Effectiveness Index (0.81). A regression model elucidating the influence of e-governance sub-indices, with an explanatory power of 66.04%, demonstrated that all three independent variables exert a direct and statistically significant impact on governance efficacy. An analysis concerning the influence of e-participation indicated that the E-Participation Index also possesses a statistically significant and direct effect on governance effectiveness, albeit its impact is partially overlapped by other components of e-governance. The insights derived from this study may support the formulation of strategies aimed at the countries’ digital advancement, integrating initiatives to reinforce the digital literacy of both the population and civil servants.
Keywords:	E-governance, E-participation, Online Services, Human Capital, Telecommunication Infrastructure, Government Effectiveness
DOI:	https://doi.org/10.5281/zenodo.19365432
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	MITIGATING MODE COLLAPSE TO IMPROVE DIVERSITY IN TEXT-TO-IMAGE GAN OUTPUTS: STRATEGIES IN ARCHITECTURAL DESIGN, TRAINING METHODOLOGIES, AND EVALUATION TECHNIQUES
Author:	SUBUHI KASHIF ANSARI, MANAL AL KHAMMASH, ANJALI APPUKUTTAN, ANNE ANOOP, SANDEEP KUMAR MATHARIYA, SHEELA D V, MOHAMMED SALEH AL ANSARI
Abstract:	Text-to-image generation using Generative Adversarial Networks (GANs) has advanced significantly in recent years. This enables image synthesis from textual descriptions. However, mode collapse remains a critical challenge that limits output diversity. This systematic review analyzes strategies to mitigate mode collapse in text-to-image GANs. It examins architectural designs, training methodologies, latent-space techniques, and evaluation metrics. The review covers 45 studies published between 2015 and 2025, categorized into: architectural innovations (18 papers), training-based strategies (12 papers), latent-space and loss function methods (10 papers), and evaluation-centric approaches (5 papers). Findings show that attention-based models, multi-scale architectures, and semantic-spatial models enhance semantic alignment and diversity, with specific limitations. Training-based approaches, including curriculum learning, adaptive training, gradient penalties, and progressive growing of GANs, help stabilize training and mitigate collapse. Latent-space techniques, such as mode-seeking losses, contrastive losses, and noise manipulation, promote output diversity. However, evaluation metrics like Fréchet Inception Distance (FID), Inception Score (IS), Learned Perceptual Image Patch Similarity (LPIPS), and Multi-Scale Structural Similarity Index (MS-SSIM) show limitations in capturing semantic diversity. Progress in mitigating mode collapse depends on combined architectural design, training stability, and loss-function engineering. Future priorities include developing unified benchmarks for evaluating semantic diversity, exploring hybrid architectures, and designing adaptive training protocols to enable more robust text-to-image models generating diverse, semantically coherent outputs.
Keywords:	Text-To-Image GANS, Mode Collapse, Output Diversity, Architectural Design, Training Methodologies, Evaluation Techniques, Attention Mechanisms, Latent-Space Techniques
DOI:	https://doi.org/10.5281/zenodo.19365469
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	ADAPTIVE FEATURE EXTRACTION FRAMEWORK FOR ROBUST FACE ANTI-SPOOFING WITH CROSS-DOMAIN GENERALIZATION
Author:	KARTHIKA S, G PADMAVATHI
Abstract:	Face spoofing detection is crucial for the security of facial recognition systems; however, many existing methods struggle to generalize across varying acquisition conditions such as changes in lighting, camera angles, and multiple types of spoofing attempts. To address this challenge, this work proposes a novel feature extraction framework that combines three advanced components: Adaptive Kernel Generator (AKG), Discrete Style Assembly (DSA), and Adaptive Style Transfer (AST). AKG dynamically adjusts feature extraction based on instance-specific characteristics, enhancing sensitivity to subtle variations in spoofing attacks. DSA categorizes input samples into distinct style categories, enabling synthesis of style-specific features that are resilient to different presentation attack instruments (PAIs) and environmental conditions. AST further refines feature representations by adaptively transferring stylistic information from reference images, ensuring consistency and accuracy across diverse scenarios. The framework is built upon a modified ResNet-18 backbone optimized for single-channel face inputs, serving as the initial feature extractor before enhancement by the AKG, DSA, and AST modules. The model is evaluated on Replay-Attack, SiW-Mv2, and OULU-NPU datasets, each offering unique variations in spoofing scenarios. Experimental results demonstrate that the proposed model outperforms a baseline ResNet-18 model, achieving up to 94% accuracy in cross-dataset testing, highlighting its effectiveness and improved generalization performance in real-world face spoofing detection. Unlike conventional face anti-spoofing methods that rely on fixed feature extractors or classifier-level adaptation, this work introduces a style-aware, feature-level adaptation strategy that improves cross-domain generalization without requiring target-domain data.
Keywords:	Face Spoofing, Domain Generalization, Adaptive Kernel Generator, Discrete Style Assembly, Adaptive Style Transfer
DOI:	https://doi.org/10.5281/zenodo.19365489
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	TOWARDS ROBUST AND INTRINSICALLY INTERPRETABLE BRAIN TUMOR MRI CLASSIFICATION VIA ADAPTIVE ATTENTION-GUIDED FUSION
Author:	MORSA SRUTHI, VEERRAJU GAMPALA
Abstract:	The correct and sensible categorization of brain tumors through magnetic resonance imaging (MRI) is still a major problem with a heterogeneous tumor morphology, overlapping intensity patterns and the shortcoming of fixed feature fusion techniques in deep learning models. Current convolutional and transformer-based architectures tend to focus on raw accuracy but do not have adaptive fusion and inherent interpretability, which restricts their clinical reliability. In this research work, whether hierarchical attention integration and spatially adaptive dual-backbone fusion with Particle Swarm Optimization (PSO)-based scaling of features are examined that can enhance robustness and explainability in the process of brain tumor MRI classification. In a bid to resolve this issue, the Enhanced Spatial Attention Enhanced (SAE) Hybrid architecture is developed that includes the ConvNeXt and EfficientNetB0 backbones with multi-stage Convolutional Block Attention Modules (CBAM), spatial gating fusion, and a learnable PSO Weighted Scaling layer. The model was tested on a stringently partitioned Kaggle brain tumor MRI dataset of 4,117 images of three tumor subtypes, where data were strictly partitioned as training and validation and hold-out independent test (906 images). It has been experimentally shown that the proposed framework attains 95.03% accuracy, 95.03% F1-score, and 0.9949 ROC-AUC, which is statistically significantly higher than the baseline concatenation (p = 0.012). The fusion of spatially adaptive methods was better than scalar weighting methods, thereby establishing that features balancing regionally is more effective than global mixing in the classification of MRI. Moreover, the presence of embedded attention mechanisms offered inherent interpretability of dynamically focusing on regions of the tumor of interest without the use of post-hoc explanation.
Keywords:	Attention-Based Feature Fusion, Convolutional Block Attention Module, Deep Neural Network Architecture, Particle Swarm Optimization, Spatial Gating Mechanism, Transfer Learning.
DOI:	https://doi.org/10.5281/zenodo.19365523
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	SCIBERT- DRIVEN GNN MODELS FOR DETECTING COMMUNITIES FROM SCOPUS KEYWORD CO-OCURRENCE NETWORK
Author:	KIRUTHIKA. R., KRISHNAVENI SAKKARAPANI
Abstract:	Identifying the research communities from the large-scale bibliometric data is essential for understanding and interpreting the trends among recent research topics. This article presents a novel framework for identifying communities from the Scopus bibliometric data by constructing a keyword co-occurrence network. The deep learning based scientific articles extracted from the Scopus bibliographic database were selected for this work. These collected data are segmented into five different time frames named as Scopus Bibliographic Dataset (SBD), SBD_1 (2006-2013), SBD_2 (2014-2016), SBD_3 (2017), SBD_4 (2018) and SBD_5 (2019). This work is developed as a hybrid framework by integrating the traditional Louvain algorithm with various Graph Neural Network (GNN) models, which helps to improve the performance through the extraction of the best textual information features from the keywords by applying the SciBERT model as node features. GNN techniques like GCN, GraphSAGE, FeaStConv, APPNP, WLConvContinuous and AGNN were implemented to compare among the models to find the best model for developing this novel framework. These integrated frameworks are named as SciBLoGCN, SciBLoGS, SciBLoFSC, SciBLoWLC, SciBLoAPPNP, and SciBLoAGNN. The experimental findings determine the presence of meaningful and effective research communities based on the structure and interconnection between the nodes within the network. This work supports scholars to understand the interconnection among their domains based on recent research topics, which enables strategic prioritization of their research areas.
Keywords:	Community Detection, Scopus Bibliometric Data, Keyword co-occurrence Networks, SciBERT, Complex Network.
DOI:	https://doi.org/10.5281/zenodo.19365549
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	TRANSPARENT WILDFIRE DETECTION SYSTEMS: THE ROLE OF FIREDETXPLAINER IN EXPLAINABLE AI MODELS
Author:	CHITTAKULA ROHINI, LAXMAIAH KOCHARLA, A. GEETHA, KONETI VARALAKSHMI, JAGADEESAN S, PRAVEENA MANDAPATI, B. SRINIVASA RAO, VENKATA SRINIVASU VEESAM
Abstract:	Advanced detection systems are required to improve the efficacy of response efforts in the face of the growing threat posed by wildfires to ecosystems and communities. Wildfires' dynamic nature frequently renders conventional detection methods inadequate. This paper introduces FireDetXplainer, a novel framework that is intended to enhance wildfire detection by incorporating transparent and explainable AI techniques. FireDetXplainer guarantees interpretability and clarity in decision-making processes by employing state-of-the-art machine learning models. Our strategy is designed to improve the accuracy of detection and foster stakeholder trust by offering actionable insights into AI predictions. The model's overall accuracy is significantly enhanced by the integration of convolutional blocks and advanced image pre-processing techniques. FireDetXplainer implements Explainable AI (XAI) tools to guarantee comprehensive result interpretation by utilizing a variety of datasets from Kaggle and Mendeley. The FireDetXplainer outperforms current top models and achieves remarkable accuracy, as evidenced by the extensive experimental results. This renders it a highly efficient method for image classification in wildfire management.
Keywords:	Wildfire Detection, Explainable AI, LIME, Meteorological Data, Deep Learning
DOI:	https://doi.org/10.5281/zenodo.19365585
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	DEVELOPMENT OF AN EDUCATIONAL MOBILE APPLICATION WITH INTELLIGENT AUTOMATION FOR MATHEMATICS TEACHING: INTEGRATION OF FLUTTER, N8N AND OPEN SOURCE AI
Author:	CARLOS J. LAVADO AYALA, LUIS E. TORIBIO PALACIOS, FERNANDO W. BALBIN URCUHUARANGA, ALAN M. INFANTE VIDALON
Abstract:	Design, develop and evaluate an educational mobile application that integrates Flutter, n8n and open source AI models (Router, Llama 2, Mistral) for automated mathematical assistance with teacher monitoring. Methods: A dual application was developed combining guided assistance and automatic resolution via AI, with automated notifications and reporting via n8n. Quasi-experimental evaluation with 180 students and 12 teachers over 12 weeks. Results: 87% adoption rate, average 23% improvement in mathematics grades, significant increase in self-efficacy (d=1.18), and 40% reduction in teacher time for diagnostic identification. AI models successfully processed 94% of mathematical queries (n=1,247 total queries). Conclusions: Intelligent automation with n8n and open source AI represents an effective innovation for personalizing mathematical learning and optimizing pedagogical monitoring.
Keywords:	Flutter, N8n, Educational Artificial Intelligence, Llama 2, Mistral, Workflow Automation, Mathematics Learning, Educational Technology.
DOI:	https://doi.org/10.5281/zenodo.19365615
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	AN INTELLIGENT NHPP-BASED SOFTWARE RELIABILITY GROWTH MODEL ENHANCED WITH DEEP LEARNING FOR CROSS-PLATFORM FAILURE PREDICTION
Author:	DR B. SUVARNA MUKHI, DR.P. NAMRATHA, DR SUKANYA K, DR. E. HARIPRASAD, A. HEMANTHA KUMAR, DR. NICHENAMETLA RAJESH, DR PRAVEEN KULKARNI, DR.P. NARESH
Abstract:	Modern software ecosystems especially those driven by open-source collaboration and continuous integration/continuous deployment (CI/CD) practices exhibit rapidly evolving and non-stationary behavior. Such dynamics pose significant challenges to conventional Software Reliability Growth Models (SRGMs). Traditional models based on the Non-Homogeneous Poisson Process (NHPP) typically assume fixed or smoothly varying failure detection rates, limiting their effectiveness in capturing real-time failure trends within highly dynamic and frequently updated code repositories. To overcome these limitations, this study proposes a hybrid predictive framework that integrates NHPP-based reliability modeling with Long Short-Term Memory (LSTM) networks for adaptive and real-time failure forecasting. The proposed Deep Learning Augmented NHPP (DL-NHPP) model enhances the classical failure intensity function by incorporating temporal patterns learned from repository-level signals, including commit frequency, issue reports, and developer activity metrics. Through this integration, the model dynamically adjusts failure rate estimations in response to evolving development behaviors. The framework is evaluated on five large-scale open-source repositories, demonstrating substantial predictive improvements. Experimental results show a reduction of more than 30% in Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) compared to conventional NHPP-based SRGMs. These findings indicate that the DL-NHPP approach provides a scalable, interpretable, and robust solution for real-time software reliability prediction in modern, continuously evolving development environments.
Keywords:	Software Reliability Growth Model (SRGM), Non-Homogeneous Poisson Process (NHPP), Deep Learning, LSTM, Failure Prediction; Temporal Modeling, Open-Source Repositories
DOI:	https://doi.org/10.5281/zenodo.19365744
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	AN INTERPRETABLE TRANSFORMER–GRAPH FUSION FRAMEWORK FOR MULTI-MODAL CARDIOVASCULAR DISEASE PREDICTION
Author:	CHANDRASEKHARA REDDY T , Dr M.PURUSHOTHAM
Abstract:	The rapid rise of urbanization and the connected mobility demonstrates the serious limitations facing traditional traffic management systems. Solutions previously focused on vehicle flow. Others relied on manual inspection for road maintenance. Both these were not comprehensive. They also caused delays in answering urban transport issues. In this paper, we will introduce DU-Net (Dual-Stream Urban Network), an intelligent deep-learning framework for real-time detection of road anomalies and vehicle analytics in a smart city. The design utilizes high-definition cathode ray tube roadside camera visual sensing and IOT sensor data from embedded infrastructure for multi-task learning to detect potholes and classify vehicles and dynamically estimate traffic density. A convolutional pipeline with two streams can capture spatial dependencies and subsequent temporal dependencies. A probabilistic fusion model views the hypotheses for sensor signals and video signals as being generated from a probabilistic mixture model. It can align the video-based hypotheses with the sensor-based hypotheses to achieve consistent decisions. The system uses a smart technique that gives an appropriate weight to each task in order to enhance robustness over weather and lighting and traffic conditions. Experimental analysis using urban road datasets shows DU-Net to outperform state-of-the-art detectors like YOLOv8 and Faster R-CNN in terms of evaluation metrics, inference speed, and computational cost. Moreover, its prediction analytics component facilitates scheduling of maintenance and traffic congestion prediction. The suggested DU-Net framework provides an extensible foundation for smart transportation systems in the future settings through data-driven, adaptive and climate-resilient urban traffic governance.
Keywords:	Deep Learning, Dual-Stream Architecture, Intelligent Transportation Systems, IOT Data Fusion, Multi-Task Learning, Pothole Detection, Smart Cities, Traffic Analytics, Urban Governance, Vehicle Classification.
DOI:	https://doi.org/10.5281/zenodo.19366388
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	KIDNEY STONE DETECTION FROM DICOM IMAGES USING IMAGE PROCESSING TECHNIQUES
Author:	YUTI DEWITA ARIMBI, SARIFUDDIN MADENDA, LUSSIANA ETP
Abstract:	Kidney stone detection from medical images is a critical step in early diagnosis and treatment planning. In this study, an image processing approach is proposed to identify kidney stones from DICOM (Digital Imaging and Communications in Medicine) files, focusing on both right and left kidney regions. The methodology includes grayscale conversion, elliptical area cropping, thresholding, and pixel counting to calculate kidney stone size. The results indicate significant differences in processing time and threshold values depending on the image characteristics. Based on the analysis, the shortest processing time was 1.30 seconds, while the longest reached 1.46 seconds, with an average of 1.32 seconds. It was also observed that images without kidney stones required approximately 1 second of processing time, while those containing kidney stones took between 20 to 40 seconds, depending on the size and number of stones. In one case (Sample 14), the processing time reached 2.21 seconds, reflecting the high pixel count (542 pixels) and larger stone area (322.59 mm²). The correlation analysis showed that the right kidney had a moderate positive relationship (R = 0.57, R² = 0.33), while the left kidney had a stronger correlation (R = 0.94, R² = 0.88), indicating more consistent pixel-to-area relationships. These findings demonstrate the effectiveness of the proposed image processing approach in identifying kidney stones with high accuracy, providing valuable insights for clinical decision-making and future improvements in automated kidney stone detection systems.
Keywords:	DICOM, Kidney Stone Detection, Image Processing, Pixel Analysis, Medical Imaging
DOI:	https://doi.org/10.5281/zenodo.19366416
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	HYBRID GLOBAL–LOCAL FEDERATED LEARNING FOR PRIVACY-PRESERVING TRAFFIC PREDICTION UNDER NON-IID URBAN DATA
Author:	CHOPPARAPU GOWTHAMI, S KAVITHA
Abstract:	Rapid urbanization and increasing vehicular density require scalable and privacy-preserving traffic forecasting systems capable of operating across heterogeneous cities. Conventional centralized deep learning approaches demand raw data aggregation, creating significant privacy, governance, and scalability concerns, particularly in multi-city collaborations. This study proposes a Privacy-Preserving Federated Learning (PP-FL) framework for collaborative urban traffic flow prediction and disaster-aware mobility modeling without sharing sensitive local datasets. The framework introduces a hybrid global–local model architecture in which shared layers capture universal spatio-temporal traffic dynamics while city-specific layers preserve localized mobility characteristics, effectively mitigating non-identically distributed (non-IID) data heterogeneity. To ensure strong confidentiality guarantees, differential privacy with calibrated Gaussian noise and secure aggregation are integrated into the federated optimization process, protecting against gradient inversion and membership inference attacks.Extensive experiments conducted on heterogeneous multi-city traffic datasets and 40,000 surveillance images demonstrate that the proposed model achieves an RMSE of 12.94, MAE of 6.01, and R² of 0.91, outperforming conventional local and centralized baselines. The framework reduces privacy leakage risk from 91% to 5% while lowering communication overhead by 34% through compressed update transmission. Additionally, the model enhances anomaly detection performance (F1-score = 0.94) and improves disaster-response sensitivity across flood and evacuation scenarios.
Keywords:	Federated Learning, Traffic Flow Prediction, Privacy Preservation, Intelligent Transportation Systems, Deep Learning.
DOI:	https://doi.org/10.5281/zenodo.19366441
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	A DYNAMIC PRICING FRAMEWORK FOR TRAIN TICKETS USING MACHINE LEARNING PREDICTION AND RULE-BASED ADAPTATION
Author:	KIKI WIJAYA, YULYANI ARIFIN
Abstract:	The railway transportation system in Indonesia has experienced a significant increase in demand, especially during holiday seasons, indicating the need for ticket price optimization to maximize revenue and balance passenger distribution. This study aims to develop a simple and efficient dynamic pricing model for train tickets, addressing the issues of static pricing and the confusing complexity of ticket subclasses for passengers and management. The methods employed include identifying key factors influencing ticket prices (booking time, route, service type, demand) and building a robust price prediction model using the XGBoost algorithm. Train ticket purchase transaction data from 2020 to 2025, including details like purchase time, route, ticket class, and schedule popularity, were utilized to generate accurate base prices. These base prices are then adjusted in real-time considering current demand and seat availability. Dynamic pricing simulations will evaluate price increases based on demand percentage and train occupancy rates. Model evaluation will use R-squared (R²), Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) metrics to measure prediction accuracy. The results of this study are expected to contribute significantly to railway companies in optimizing ticket pricing strategies and improving operational efficiency.
Keywords:	Ticket Price Prediction, Dynamic Pricing, XGBoost, Railway, Machine Learning.
DOI:	https://doi.org/10.5281/zenodo.19366465
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	FUSIONDEFECTNET-A CNN VISION TRANSFORMER METHOD WITH ADAPTIVE GATING FOR EXPLAINABLE TEXTILE DEFECT DETECTION
Author:	THANDAVA KRISHNA SAI PANDRAJU, VENKANNA CHANAGONI, P MARY KAMALA KUMARI, MANEESHA VADDURI, SHAIK SALMA BEGUM, V. SUMA AVANI
Abstract:	It is challenging for textile manufacturers to automate inspection due to a pronounced imbalance in the class of samples, with the majority of samples being defect-free, and the complementary nature of local and global defects. In this study, FusionDefectNet is presented as a novel hybrid architecture. For extracting local features, Convolutional Neural Networks (CNNs) are used, and for understanding global context, Vision Transformers (ViTs) are employed. As part of the adaptive gating mechanism, the weights on branch contributions are modified based on the input attributes. The 68.7:1 class imbalance is addressed by integrating class-weighted loss with focal loss (γ=2). To gain insights into decision-making processes, use interpretable AI technologies like Grad-CAM and ViT attention visualizations. According to the TILDA dataset, we were 97.3% accurate and scored 96.9% higher than pure CNN solutions. When α is below 0.3, CNNs are preferred over other search methods for localized imperfections like holes and stains. However, for structural issues at global scale, it favors ViTs when α exceeds 0.7. According to Explainable AI, CNN identifies minor discrepancies, while ViT identifies patterns within the context that are more substantial.
Keywords:	FusionDefectNet, CNN, Grad cam, Explainable AI, Focal Loss, VIT.
DOI:	https://doi.org/10.5281/zenodo.19366478
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	EVALUATING QAOA AND QSVM FOR OPTIMIZATION AND MACHINE LEARNING ON CLASSICAL AND QUANTUM
Author:	Dr. R BOOPATHI, Dr. KALYANAPU SRINIVAS, Dr. N PUSHPALATHA, V V RAMA KRISHNA, V LAKSHMI SAILAJA, Dr. A MUTHUKRISHNAN, SHAIK JILANI BASHA, Dr. R BALAMURUGAN
Abstract:	The ability of quantum computing to transform multiple domains, such as cryptography, optimization, and machine learning, has been recognized. This paper describes how quantum algorithms, such as QAOA, can be applied to solve the TSP and QSVM for MNIST. We want to examine how quantum algorithms compute complex problems and distinguish their performance from classical computing systems. We use quantum processors from IBM and classical computing to tackle simulation and optimization tasks, respectively. The research revealed that QAOA gives close-to-optimal outcomes for the TSP, running much faster than classical algorithms. At the same time, QSVM still performs well in digit classification with only a slight decrease in accuracy. Nevertheless, training quantum neural networks requires more time, showing that hardware should be upgraded. This work points out how quantum computing might play a significant role in helping with optimization and machine learning while explaining the present problems and possibilities for quantum algorithms and hardware. This study indicates that if error correction techniques and hardware are improved, quantum computing can help solve everyday challenges.
Keywords:	Quantum computing, Quantum Approximate Optimization Algorithm (QAOA), Quantum Support Vector Machines (QSVM), Traveling Salesman Problem (TSP), MNIST dataset, Hybrid quantum-classical models
DOI:	https://doi.org/10.5281/zenodo.19366503
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	BIG DATA AND MACHINE LEARNING FRAMEWORK FOR CANCER, FINANCIAL, AND STRESS RISK PREDICTION
Author:	M V B MURALI KRISHNA M, VIJAYA KRISHNA SONTHI, N. SRINIVAS RAO, AVSS SOMASUNDAR, M.SRIKANTH, M CHILAKARAO
Abstract:	The rapid evolution of the big data in the medical, financial, and behavioral track has offered an opportunity to utilize predictive analytics to contribute to the well-being of the whole picture. However, the existing systems tend to work on the prediction of cancer risks, estimation of financial status, and stress analysis independently, which is limited to provide integrated and tailored risk measurement in the future. To address this limitation, the proposed paper will recommend one single Big Data and Machine Learning Framework to forecast the risks of Cancer, Financial, and Stress. It uses a combination of heterogeneous data, medical indicators, financial data, and behavioral data that are connected to stress and executes the supervised machine learning algorithms such as Logistic Regression, Decision Tree, Random Forest, Linear Regression, and Gradient Boosting. The results of the experiment indicate that, as compared to the Logistic Regression or the Random Forest, Decision Tree model predicted the risk of cancer the most accurately with an accuracy of 83%. Gradient Boosting was the lowest in the Mean Squared Error of 5.25 x 10-6 and better than that of Linear Regression, which is 0.15. In addition, stress risk classification was also effective in the determination of the various levels of stress basing on behavioral and physiological indicators. These results confirm the notion that the proposed integrated framework improves predictive accuracy and makes it possible to consider risks in the comprehensive manner. This model provides decision support tool, which is data-oriented to detect early risks of cancer, financial planning, and stress management to improve holistic well-being.
Keywords:	Big Data Analytics, Machine Learning, Cancer Risk Prediction, Financial Risk Estimation, Stress Risk Analysis
DOI:	https://doi.org/10.5281/zenodo.19366535
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	F4-SERUNET: THEORY-GUIDED TEMPORAL DEEP LEARNING FRAMEWORK FOR MULTI-FACTOR PERFORMANCE PREDICTION IN SERU PRODUCTION SYSTEMS
Author:	MANISH KAUSHIK, DEVENDRA CHOUDHARY , PRIYA DARSHINI
Abstract:	Seru Production Systems entail complex time-evolving interactions between structural flexibility, operational efficiency, human related, and lean process improvement-the result of stochastic task processing, workforce learning, and reconfiguration dynamics. Due to the scarcity of temporally rich datasets and limitations regarding static regression models, accurate predictions of these interacting performance dimensions are rather difficult to make. This research introduces F4-SeruNet, a theory-guided temporal deep learning framework for jointly predicting four interpretable Seru performance factors. A synthetically generated yet realistic factory evolution is mathematically created, generating 180,000 temporal observations from 300 production episodes of 600 discrete time steps each. Time-skewed bottleneck times due to variability-driven congestion effects, learning-induced performance gains, and reconfiguration penalties are all captured in the data; this ensures face validity and consistency in sensitivity. F4-SeruNet integrates temporal modeling of factor-specific encoders and cross-factor attention with soft monotonic constraints that embed Seru production theory. Experimental results are reported to include consistently high goodness-of-fit in all factors, with RMSE values within 0.003747-0.005474 and R² values in excess of 0.9985 for all four factors. Ablation and sensitivity analyses confirm that theory-guided constraints stabilize learning and enforce coherent responses to perturbations, while strong static baselines yield lower one-step errors. F4-SeruNet demonstrates superior temporal consistency, balanced multi-factor behavior, and robustness under rolling-horizon forecasts. These results point to the suitability of F4-SeruNet for scenario analysis and medium-term decision support in dynamic Seru production planning.
Keywords:	Seru Production Systems, Temporal Deep Learning, Theory-Guided Neural Networks, Multi-Factor Performance Prediction, Synthetic Industrial Data
DOI:	https://doi.org/10.5281/zenodo.19366562
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	A MULTIMODAL CONVNEXT AND LLM-POWERED CLINICAL DECISION SUPPORT SYSTEM FOR EARLY AMD CLASSIFICATION AND TREATMENT
Author:	AKILA A , DURGADEVI P
Abstract:	Age-Related Macular Degeneration (AMD) is a leading cause of vision loss in the elderly, making early identification essential to preventing irreparable damage. While timely intervention is critical for preserving vision, traditional diagnostic methods often struggle to balance accuracy with computational speed. The primary research contribution of this work is the development of an Integrated Theranostic Framework that synergizes deep learning with Generative AI to perform simultaneous disease staging and therapeutic guidance, fundamentally solving the computational bottlenecks of unimodal image analysis. The proposed model introduces a multimodal fusion framework, specifically a ConvNeXt, that integrates a dual-stream input of Color Fundus Photography (CFP) and Optical Coherence Tomography (OCT) to transcend the constraints of unimodal classification, specifically targeting the early staging of AMD. By leveraging the advanced feature extraction capabilities of the ConvNeXt architecture, the model robustly synthesizes spatial features from both imaging modalities, optimized using both standard backpropagation and stochastic gradient descent. Furthermore, the new knowledge created by this research is the successful operationalization of a Neuro-Symbolic diagnostic pipeline within ophthalmology. Following the classification of the AMD stage based on fused decision scores, a Large Language Model (LLM) processes the symbolic clinical findings to generate context-aware drug and therapeutic recommendations. Model interpretability is further ensured through Class Activation Mapping (CAM) to visualize decision heatmaps. Implemented in Python, the proposed framework achieves a superior accuracy of 98.65% with a rapid inference time of 13.12 seconds. These results demonstrate that combining ConvNeXt-driven multimodal fusion with LLM-driven insights significantly advances the capability of early AMD detection and establishes a new paradigm for automated, guardrailed patient management
Keywords:	Age-Related Macular Degeneration (AMD), Integrated Theranostics, OCT, CFP, Multi-Modal Fusion; Cascaded Group Attention; Deep Spatial Attention; Large Language Models (LLM); Personalized Therapy.
DOI:	https://doi.org/10.5281/zenodo.19366586
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	INTEGRATION OF ADAPTIVE EDUCATIONAL PLATFORMS BASED ON ARTIFICIAL INTELLIGENCE FOR FOREIGN LANGUAGE TEACHING IN NON-LINGUISTIC HIGHER EDUCATIONAL INSTITUTIONS
Author:	ALEVTYNA MINIAILOVA, IRYNA HOLUBIEVA, YEVHENIIA KOSTYK, SNIZHANA KUTSYN, OKSANA SYVYK
Abstract:	The study assesses the effectiveness of adaptive educational platforms based on artificial intelligence (AI) in teaching foreign languages in non-linguistic higher education institutions (HEIs). The relevance of the issue is determined by the growing need for personalized language trajectories that can combine prediction accuracy with pedagogical interpretability and stability of algorithms. The paper compares three groups of platforms – Machine Learning (ML)-based, Deep Learning (DL)-based, and Hybrid Cognitive-oriented. They were tested based on the data of success, Learning Management System (LMS) activity, and standardized language testing (n = 524). To generalize the results, the integral Language Training Effectiveness Index (LTEI) was used, which combines the metrics F0.5-score, Receiver Operating Characteristic (ROC)- Area Under Curve (AUC), Gini index, and pedagogical interpretability assessment (Ped-Value). The methodology included stratified validation, analysis of the stability of the weighting coefficients (ΔW-index), and testing the statistical significance of the differences (analysis of variance (ANOVA), paired t-tests). The results showed that the hybrid systems provided the highest balance (LTEI = 0.85–0.87; Ped-Value = 0.85; Kappa = 0.79), combining accuracy and understandable interpretation for teachers, despite the higher computational costs. Deep models achieved maximum accuracy (F0.5 ≈ 0.81; AUC ≈ 0.87), but lost in stability (ΔW-index = 0.134) and transparency of the results. Classical ML algorithms provided the fastest adaptation time (≈0.84 s/epoch) and the lowest weight fluctuations (ΔW-index = 0.091), but their LTEI indicators remained at the level of 0.72–0.76. Statistically significant differences (p < 0.01) confirmed the superiority of hybrid solutions, while the choice between ML and DL platforms should depend on the resources and tasks of the educational process. The academic novelty of the study is the comprehensive comparison of different architectures of adaptive platforms according to technical and pedagogical criteria, as well as taking into account the user experience (UX)/user interface (UI) aspects of educational web interfaces.
Keywords:	Adaptive Platforms, Artificial Intelligence, Machine Learning, Deep Learning, Hybrid Cognitive-Oriented, LTEI, ΔW-Index, Interpretability, Stability of Forecasts, Teaching Foreign Languages, UX (User Experience), Human–Computer Interaction.
DOI:	https://doi.org/10.5281/zenodo.19366610
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	FORMING SOCIAL BUSINESS OPPORTUNITY BELIEFS IN A DEVELOPING ECONOMY: A CASE STUDY OF INDONESIAN DIGITAL ENTREPRENEURS
Author:	REYNARD J.N MAKARAWUNG, AGUNG STEFANUS KEMBAU, JULIUS SUTRISNO, LILIS SUSILAWATY4, TANNIA
Abstract:	The "opportunity actualization" perspective has become a dominant framework for understanding how entrepreneurs identify and pursue social business opportunities. However, existing models are predominantly Western-centric, overlooking the distinct institutional and cultural realities of developing economies. This study addresses this theoretical gap by investigating how digital entrepreneurs in Indonesia form social business opportunity beliefs in an environment characterized by institutional voids. Adopting a qualitative multiple case study design, we conducted in-depth interviews with Indonesian digital impact entrepreneurs and analyzed the data using the Gioia methodology. The findings reveal that opportunity formation in this context is not merely an individual cognitive act but a socially negotiated process driven by "Communal Alertness" rather than individual dissatisfaction. Crucially, the study identifies a novel mechanism of "Digital-Relational Substitution," where entrepreneurs leverage digital platforms and informal networks to substitute for the lack of formal institutional validation common in the Global North. We argue that in high-context cultures, opportunity beliefs are validated through relational capital and digital micro-experiments rather than formal market data. These insights contribute to "recalibrating" entrepreneurship research for the Global South and offer practical guidance for policymakers to foster trust-based entrepreneurial ecosystems.
Keywords:	Social Entrepreneurship, Opportunity Actualization, Digital Entrepreneurship, Developing Economy, Gioia Methodology.
DOI:	https://doi.org/10.5281/zenodo.19366642
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	A FUSION-BASED SENTIMENT CLASSIFIER FOR HINGLISH USING CONTEXTUAL ENCODERS AND COMMONSENSE INFERENCE
Author:	KISHORE KUMAR P. V, HARI JYOTHULA
Abstract:	Sentiment analysis of code-mixed text like Hinglish is still a tough problem. People who use social media often mix Hindi and English. The grammar isn't very formal and isn't always the same. People often show how they feel without saying it directly. A lot of the sentiment analysis models that are already out there depend mostly on the text's context. These models have a hard time figuring out how people feel when they show it through situations or cultural cues. This limit causes incorrect classification. This issue happens more often in content that is neutral or has a negative tone. The goal of this study is to make it easier to classify feelings in Hinglish text. The work focuses on dealing with emotional expressions that aren't clear. A framework for classifying sentiments based on fusion is suggested. The framework brings together contextual embeddings and external commonsense knowledge. Contextual embeddings record how words are used in the text. Common sense knowledge helps us figure out how people feel about everyday things. The Hinglish SentiMix dataset, which has about 20,000 tweets that are marked as positive, negative, or neutral, is used to test the proposed model. The results indicate that the suggested method attains a weighted F1-score of 74.1%. This score is better than those of well-known transformer-based baseline models. The results show that adding commonsense reasoning makes it easier to understand sentiment. The improvement is big for feelings that are not directly stated. The study shows how useful it is to combine linguistic context with outside knowledge. This method works well for analyzing multilingual social media in the real world.
Keywords:	Hinglish Sentiment Analysis, Code-Mixed Text, Contextual Embeddings, Commonsense Knowledge, Fusion Model.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	KRILL HERD OPTIMIZED ZONE ROUTING PROTOCOL (KHO-ZRP) FOR ENERGY-EFFICIENT ULTRA-DYNAMIC FLYING AD HOC NETWORKS
Author:	GUNAVATHY H, Dr.RAVINDRANATH P.V
Abstract:	Flying Ad Hoc Networks operate under extreme mobility, fluctuating node density, frequent link disruptions, and strict energy constraints, which jointly degrade routing stability and communication reliability. Existing FANET routing approaches address these challenges in isolation through learning-based adaptation, clustering, or bio-inspired heuristics, leaving coordination among zoning, relay selection, and energy regulation insufficiently explored. This work introduces a Krill Herd Optimized Zone Routing Protocol that embeds collective swarm intelligence directly into zonal routing control. New knowledge is created through an optimization-governed routing framework where krill herd dynamics regulate zone radius adaptation, relay selection, hop progression, and spatial stability within a unified routing architecture. The contribution advances routing-system design rather than proposing a new optimizer, enabling stable and energy-aware communication under ultra-dynamic aerial conditions. NS3-based evaluation across varying node densities demonstrates reduced delay, lower packet loss, improved packet delivery, higher throughput, and controlled energy consumption compared to existing protocols. Results confirm the effectiveness of optimization-driven zonal coordination for sustaining quality of service in highly dynamic FANET deployments.
Keywords:	Zone Routing Protocol, Krill Herd Optimization, UAV, FANET, QoS)
DOI:	https://doi.org/10.5281/zenodo.19366952
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	A DATA-DRIVEN HYBRID FRAMEWORK FOR TELEVISION AND MEDIA ETHICS: SCIENTOMETRICS AND DISCOURSE NETWORK ANALYTICS
Author:	JIMI NAROTAMA MAHAMERUAJI, ATWAR BAJARI, DADANG RAHMAT HIDAYAT, ACENG ABDULLAH
Abstract:	This study integrates bibliometric mapping and Discourse Network Analysis (DNA) to examine television and media ethics across scholarship and online news discourse. Bibliometrically, Scopus-indexed publications (2000–2024; n = 238) were analyzed using ScientoPy to compute AGR, ADY, PDLY, and the h-index. Results indicate a marked rise in output after 2018, with recurrent themes around media ethics and television discourse, and a prominent cluster related to reality television, alongside emerging lines of inquiry such as artificial intelligence, indigenization, and televangelism. Complementing this macro-level view, DNA was applied to Indonesian online news coverage in 2025 (Google News retrieval via Linkclump; keyword “Etika Televisi”; n = 69). Network modeling (DNA 3.0.11) and visualization (Visone 2.28.1) show a centralized discourse structure in which KPI and the Press Council (Dewan Pers) function as anchor nodes, while issue connectivity concentrates on General News and Broadcasting Ethics. The individual–issue network further suggests cross-issue linkage through intermediary nodes (e.g., journalists). Limitations include reliance on a single indexing database and potential platform/source bias in Google News retrieval; future research should expand databases, use multi-platform media corpora, and test robustness with longitudinal and metric-based network validation.
Keywords:	Scientometrics, ScientoPy, Discourse Network Analysis (DNA), Network Analytics, Television and Media Ethics
DOI:	https://doi.org/10.5281/zenodo.19366973
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	A REAL TIME HYBRID QUANTUM CLASSICAL DIAGNOSTIC FRAMEWORK FOR EARLY DISEASE DETECTION
Author:	GOWRIPUSHPA GEDDAM, NARASIMHA RAO THOTA, PRANUSHA DOGGA, SELVA MALAR N4, RAVURI LALITHA, MYLAVARAPU KALYAN RAM, M CHAITANYA KUMARI, SATHISHKUMAR SHANMUGAM
Abstract:	Early detection of clinical deterioration is crucial for better patient outcomes, and current diagnostic models struggle to model the complex interactions observed in nonlinear patterns in multimodal medical data while meeting the performance requirements for real-time operation. This paper proposes a Real-Time Hybrid Quantum-Classical Diagnostic Framework (HQCDF) to improve early disease detection by combining a Quantum Variational Biomarker Embedding module (QVBE), a Classical Temporal-Clinical Learning module (CTCLM), and a novel Quantum-Classical Diagnostic Fusion mechanism (QCDF). A semi-synthetic multimodal dataset, Med-EarlyQ, physiological time series, lab biomarkers, and clinical scores was built to evaluate. The hybrid model has been trained using End-to-end differentiable quantum-classical optimization via the parameter-shift rule. Results showed that HQCDF obtained 96% accuracy, 0.93 F1-score, and 0.95 AUROC to outperform state-of-the-art classical deep learning models and existing hybrid approaches by 6~15% while controlling the real-time inference latency to 18 ms. Analysis revealed that the QCDF module contributed to increased robustness against quantum noise and missing clinical information, and to QE's increased sensitivity to subtle trends in disease in its early stages. The results show that the proposed hybrid framework is a promising evolution, addressing the call for next-generation, real-time, and resource-efficient diagnostic systems with great potential for early intervention and clinical decision-support impact.
Keywords:	Hybrid Quantum–Classical Diagnosis, Early Disease Detection, Quantum Variational Circuits, Real-Time Clinical Analytics, Multimodal Biomarker Modeling, Quantum–Classical Fusion.
DOI:	https://doi.org/10.5281/zenodo.19367026
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	HCRHE-NET: HIGH-CONFIDENCE RESIDUAL HYBRID ENSEMBLE NETWORK FOR BREAST CANCER DETECTION FROM TCGA-BRCA DNA METHYLATION DATA
Author:	HEMALATHA D , N. GOMATHI
Abstract:	Due to noise, redundancy and uncertainty with respect to predictive confidence, proper classification of high-dimensional biological data remains a significant challenge. To overcome these shortcomings, this paper proposes a High-Confidence Residual Hybrid Ensemble (HCRHE) Network, which is a composite of residual learning, deep neural modelling, and confidence-conscious decision fusion to classify diseases with high confidence. The Cancer Genome Atlas (TCGA) contains large-scale data on DNA methylation that are susceptible to overfitting and unstable forecasts by using conventional deep learning models to evaluate the proposed methodology. To learn latent patterns due to reconstruction, HC-RHE architecture consists of a primary-prediction base multilayer perceptron (MLP), together with a residual learning path that is constructed using an autoencoder and residual MLP. A new machine called a confidence-based fusion technique allows the dynamical weighting of the base and residual prediction in terms of model certainty and makes adaptive decision-making. Also, forecasts with high confidence margins are retained and a high-confidence filtering process used, which maximizes reliability with minimal coverage loss. An accuracy-optimized threshold selection strategy is also provided in order to enhance the performance of classification further. Vast comparative experiments are conducted with the state-of-the-art deep learning baselines, including CNN, autoencoder-based classifiers, Dense DropConnect, residual CNN frameworks, and Basic MLP (Adam and SGD). The proposed HC-RHE has a much better accuracy at 98.7 compared to all other methods. These results prove success of confidence-aware residual fusion as they reflect continuous improvement over the best baseline CNN model. The proposed framework is, all in all, exceptionally promising to the field of clinical decision-support systems and gives a credible, intuitive, and high-confidence classification paradigm of high-dimensional biomedical data.
Keywords:	High-Confidence Learning, Residual Hybrid Ensemble, Deep Neural Networks, Autoencoder, DNA Methylation, Biomedical Classification, TCGA, Confidence-Aware Fusion, Cancer Prediction
DOI:	https://doi.org/10.5281/zenodo.19367048
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	SCARLET MACAW-INSPIRED DEEP BELIEF NETWORK FOR EARLY, ACCURATE, AND INTERPRETABLE PREDICTION OF GESTATIONAL DIABETES MELLITUS
Author:	D.SHOBANA, V VINODHINI
Abstract:	Gestational Diabetes Mellitus (GDM) is a prevalent pregnancy-related condition that contributes significantly to maternal and neonatal health complications, such as pre-eclampsia, neonatal hypoglycemia, and long-term metabolic disorders. Existing diagnostic methods, such as the Oral Glucose Tolerance Test (OGTT), are invasive, time-consuming, and not easily scalable for early detection. This research introduces the Scarlet Macaw-Inspired Deep Belief Network (SM-DBN), a deep learning-based model designed to predict GDM risk early in pregnancy. By incorporating bio-inspired optimization techniques such as adaptive foraging and territorial behavior, SM-DBN addresses challenges like class imbalance, missing data, and the dynamic nature of pregnancy-related risks. The model achieved a classification accuracy of 81.277%, showing significant promise in early-stage GDM detection. SM-DBN integrates temporal pattern learning to capture trimester-specific risk variations, and explainable decision support ensures transparency for clinicians. The model’s ability to handle incomplete and noisy clinical data, adapt dynamically to evolving risks, and offer interpretable predictions makes it a highly effective tool. This approach provides a scalable, interpretable, and non-invasive solution for GDM risk assessment, reducing the reliance on traditional diagnostic tests and enhancing maternal-fetal health outcomes across diverse healthcare settings.
Keywords:	Healthcare, Gestational Diabetes Mellitus, Prediction, Bio-Inspired Optimization, Deep Belief Network, Diabetes.
DOI:	https://doi.org/10.5281/zenodo.19367062
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	DEEP REINFORCEMENT LEARNING FOR PROACTIVE CYBERSECURITY THREAT DETECTION
Author:	DR. NAIM SHAIKH, DR. VIVEK VEERAIAH, DR. A.PANKAJAM, DR. TARUN DALAL, DR. MAMATHA G, DR. G. NAGESWARA RAO, DR. VINOD MOTIRAM RATHOD, DR. TRIPTI SHARMA
Abstract:	The proliferation of interconnected ecosystems, encompassing cloud infrastructures, IoT networks, and 5G platforms, has facilitated the execution of cyberattacks. Consequently, systems are increasingly susceptible to intricate, adaptive attacks. Reactive security measures, such as signature-based IDS and conventional machine learning models, are ineffective until an attack has already occurred. This deficiency stems from their inability to predict and mitigate threats characterised by aggressive evasion, cognitive wander, and evolving assault methodologies. Furthermore, the expansion of digitally linked systems has exacerbated vulnerabilities to sophisticated cyberattacks. Traditional cybersecurity protocols typically identify threats only post-incident. In the field of cybersecurity, a shift towards proactive and adaptive approaches is necessary due to AI's limitations, even if AI enhances pattern recognition. In contrast to conventional reactive methods, this research demonstrates the potential of DRL to build a proactive system for danger identification that can adapt in real-time to new threats. To tackle these issues, we provide DRL-PRoTECT, a new proactive cybersecurity approach that combines deep reinforcement learning with existing methods. The system is able to autonomously detect and mitigate threats in real-time thanks to its hierarchical DRL decision engine, predictive anomaly scoring, and self-supervised representation learning. Results on enterprise-scale systems, NSL-KDD, and UNSW-NB15 show that DRL-PRoTECT outperforms traditional IDS, ML/DL benchmarks, and virtual testbeds. With an F1 score of 94.5%, a false positive rate of 2.8%, and a recall rate of 93.7%, the framework accomplished its goals. The technology also reduced the time needed to identify threats by half. Its ability to adapt allowed it to keep working well despite changing priorities, new types of attacks, and attempts to bypass it. Analysts found that including a human-in-the-loop orchestrator made it easier and less demanding to stay alert. This led to better understanding, compliance, and trust. The results suggest that DRL-PRoTECT could help move cybersecurity defences from a detection-focused approach to a more proactive, self-sufficient, and resilient one. In response to changing threats, this article presents a proactive and scalable cybersecurity model that automatically shifts from detection to defense.
Keywords:	DRL; Proactive Cybersecurity; Threat Detection; IDS; Anomaly Detection; Attacks; Concept Drift; Federated Learning; Blockchain Security.
DOI:	https://doi.org/10.5281/zenodo.19367074
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	LEGAL PROTECTION IN ALGORITHMIC CONSUMER MARKETS: A SYSTEMIC APPROACH TO PLATFORM LIABILITY AND DIGITAL CONTRACTS
Author:	OLEKSANDR DONCHENKO, OLENA OLSHANSKA, IHOR OKUNIEV, OLEXANDER VERHOLIAS, SERHII VASYLYNA
Abstract:	The relevance of this study stems from the transformational changes in consumer legal relations in the context of digitalization and the development of the platform economy, which are accompanied by an increase in the structural vulnerability of consumers and a decrease in the effectiveness of traditional legal protection mechanisms. The purpose of the study is to substantiate directions for improving legal mechanisms for consumer protection, and the subject of the study is defined as consumer legal relations arising in the digital and platform environment. The theoretical basis of the study is the provisions of modern consumer law doctrine, the concept of structural asymmetry in digital markets, behavioral approaches, and the development of European regulatory law, with the simultaneous application of general scientific and specific legal methods of analysis, synthesis, comparison, and generalization. The study demonstrated that digital platforms play an active role in shaping consumer choice, leading to a transformation of the right to informed choice, fair contract terms, and effective legal remedies. It has been demonstrated that personalized commercial practices and algorithmic ranking mechanisms undermine consumer autonomy and require preventive legal regulation. The feasibility of a functional approach to the responsibility of digital platforms and the development of collective mechanisms for consumer protection was substantiated. The practical significance of the results lies in their potential use for improving national legislation, harmonizing it with European Union legislation, and forming effective state policy for consumer protection.
Keywords:	Consumer Protection, Digitalization, Platform Economy, Algorithmic Influence, Consumer Law, Collective Protection, Collective Consumer Interests, EU Alignment, Invalidity, Types of Invalidity, Consumer Interests; EU Alignment, Invalidity, Types of Invalidity
DOI:	https://doi.org/10.5281/zenodo.19367092
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	MO-GRPC AN OPTIMIZED FRAMEWORK FOR ENHANCING PERFORMANCE, SCALABILITY, AND RELIABLITY IN WEB SERVICES
Author:	J.GNANABHARATHI , K.VADIVAZHAGAN
Abstract:	The MO-gRPC framework has been proposed to enhance the performance, scalability, and reliability of web services in dynamic and resource-constrained network environments. Web services often face challenges such as high latency, inefficient resource utilization, and reduced throughput under increasing workloads, which impact their overall efficiency and adaptability. The purpose of MO-gRPC is to address these challenges by incorporating optimized mechanisms that improve communication precision, adaptive load balancing, and fault resilience.Through rigorous evaluation, MO-gRPC has demonstrated significant improvements in throughput, packet delivery reliability, and load balancing efficiency compared to conventional approaches. The proposed framework plays a pivotal role in web services by ensuring efficient data transmission, dynamic resource allocation, and robust performance under varying network conditions.MO-gRPC fills critical gaps by offering a scalable, resource-efficient solution that reduces delays and maximizes bandwidth utilization, making it a promising approach for optimizing web service frameworks in diverse and evolving network scenarios.
Keywords:	Mantis Optimization - Web Services Optimization - Multi-Channel Communication - Adaptive Load Balancing - Fault Tolerance - Resource Allocation Efficiency.
DOI:	https://doi.org/10.5281/zenodo.19367115
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	SUSTAINABLE AI FRAMEWORK FOR FAULT DETECTION IN 6G-INTEGRATED INDUSTRY 4.0 DATA ECOSYSTEMS
Author:	RALLA SURESH, PANTHANGI VENKATESWARA RAO, K. VENKATA SUBBA REDDY, Z. SUNITHA BAI, PARASA KONDALA RAO, P. LAKSHMI PRASANNA, B. VARAPRASAD RAO
Abstract:	The advent of 6G and Industry 4.0 technologies has revolutionized industrial automation, connectivity, and data processing. With the growing complexity of heterogeneous data environments in these domains, detecting faults in real-time has become increasingly challenging. This paper proposes a sustainable deep learning framework that integrates advanced neural networks with resource-efficient processing techniques for fault detection in 6G-enabled Industry 4.0 environments. The framework leverages data from various sources, including IoT devices, sensors, and industrial machines, ensuring high accuracy, scalability, and energy efficiency. A hybrid deep learning model combining Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks is employed to capture both spatial and temporal data patterns. The framework is designed to optimize resource allocation while maintaining fault detection performance. Simulation results demonstrate the efficacy of the proposed approach, highlighting its potential to enhance fault management in smart industrial systems.
Keywords:	6G, Industry 4.0, Deep learning, Fault detection, Convolutional Neural Networks, Long Short-Term Memory, Sustainability, IoT
DOI:	https://doi.org/10.5281/zenodo.19367146
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	SECURING THE FUTURE OF WSNS: A HYBRID FEDERATED AND DEEP LEARNING APPROACH TO FAULT DETECTION AND THREAT MITIGATION
Author:	DR. R. SARAVANAKUMAR, DR. B. NARMADA, DR. KEERTHI KETHINENI, VENKATA BALA ANNAPURNA P, DR. RAKSHITHA KIRAN P, SHAIK JILANI BASHA, DR. N. SATHEESH
Abstract:	Intelligent computing is getting more and more close to the concept along with Wireless Sensor Networks (WSNs) operated for fault detection and system reliability assurance. However, more traditional centrally located deep learning (DL) models still have limitations in terms of scalability, suffer from high communication overhead and may have privacy issues. To overcome these problems, this work proposes a Hybrid Federated Deep Learning (HFDL) model, which merges federated learning (FL) and distributed DL models to enable secure, energy-efficient, and reliable fault detection in large-scale WSNs. This approach was evaluated through a simulated setup consisting of 500 sensors nodes along with various fault conditions (data loss, node failure, communication errors). At the edge nodes, the model comprises CNN and LSTM-based local models, and the global model is updated through a federated virtual aggregation of local updates, thereby no raw data sharing is involved. They were compared with common machine learning baselines such as Support Vector Machine (SVM), k -Nearest Neighbors (kNN), standalone DL only, and FL only. The results indicate that HFDL is going to have an average fault detection rate of 96.8, a reduction in latency by 22, and energy consumption of 18 less than the existing methods. Such outcomes are evidence that the model under consideration elevates not only the computational efficiency level but also the data privacy standards and can be implemented in the next-generation intelligent sensor instances.
Keywords:	Wireless Sensor Networks (WSNs), Federated Learning, Deep Learning, Fault Detection, Anomaly Detection, Security.
DOI:	https://doi.org/10.5281/zenodo.19367160
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text

Title:	AI CHATBOTS IN CONVERSATIONAL COMMERCE: INVESTIGATING THE ROLE OF PERCEIVED SERENDIPITY AND SIMILARITY IN CUSTOMER SATISFACTION WITHIN INDIAN E-RETAIL
Author:	DR. P ARCHANA, DR. U BHOJANNA, DR. G V M SHARMA, DR. MANJUNATH NAGABHUSHAN
Abstract:	AI-driven chatbots play an increasingly important role in e-retail, yet prior research has mainly focused on their functional performance, such as speed and accuracy, with limited attention to the emotional and experiential mechanisms through which they influence consumer behavior. This study examines how AI chatbot interactions generate perceived serendipity and perceived similarity and how these perceptions enhance customer satisfaction in India’s online cosmetics retail sector. A quantitative survey was conducted with 588 Flipkart users who interacted with chatbots while browsing or purchasing cosmetic products. Guided by Social Presence Theory and Affective Response Theory, the conceptual model was tested using Structural Equation Modelling (SEM) through SmartPLS 4.0. The results show that chatbot service quality significantly improves perceived serendipity and perceived similarity, and these constructs fully mediate the relationship between service quality and customer satisfaction. The findings indicate that satisfaction in conversational commerce is not only a cognitive evaluation of utility but also an emotional outcome shaped by unexpected discovery and human-like interaction, highlighting the importance of affective design in AI-enabled retail systems.
Keywords:	AI Chatbots, Perceived Serendipity, Perceived Similarity, Customer Satisfaction, Conversational Commerce, Flipkart, E-Retail
DOI:	https://doi.org/10.5281/zenodo.19367183
Source:	Journal of Theoretical and Applied Information Technology 31^st March 2026 -- Vol. 104. No. 6-- 2026
Full Text