Browse Items (3095 total)
Sort by:
-
Ensemble Hybrid LSTM Architectures for Robust Multi-Currency Forex Forecasting
The analysis of financial time series presents a longlasting obstacle regarding currency exchange rate forecasting because volatility and nonlinearity and non-stationarity characterize currency markets. The research presents an ensemble forecasting system which combines various deep learning and hybrid predictive models such as LSTM and GRU-LSTM and CNN-LSTM and Attention-LSTM and XGBoost-LSTM for scalable integration. The ensemble methodology follows a dynamic weighted averaging technique which bases its priority on assigning weights through the reciprocal calculation of Mean Squared Errors from individual models to identify accurate forecasters. A representative study based on the EUR/USD exchange rate took place as part of extensive evaluations that spanned various currency pairs. The standalone XGBoost-LSTM model proved most effective in terms of MSE and R2 values at 0.000088 and 0.9778 respectively. The ensemble model proved to be highly robust and generalizable through its outcomes which produced an MSE of 0.000142 along with MAE of 0.009204 and R2 of 0.9643. The ensemble approach stands as an effective and reliable method to increase both stability and predictive power of forex forecasting systems. The conceptual structure offers sound potential applications for algorithmic trading as well as financial risk management and multi-currency strategic decision-making systems. 2025 IEEE. -
Enhancing Malware Detection Through Hybrid Deep Learning Techniques
The detection of malware needs superior methods than basic signature detection because it remains vital to cybersecurity. This research examines malware classification through the deep learning approach by analyzing Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU) and develops a new BiGRU + CNN hybrid model. The main purpose is to achieve better detection performance through reduced numbers of false alarms. The research employs executable file feature data while implementing preprocessing methods together with fivefold cross-validation validation to establish strong model reliability. Experimental findings show CNN along with LSTM and GRU attains excellent recall values yet produces elevated erroneous positive predictions. The proposed BiGRU + CNN model delivers superiority over single-model architecture as it reaches 96.06% accuracy alongside 96.13% precision and 99.92% recall and 97.99% F1-score. The obtained results show that this integration has better malware detection capabilities thereby demonstrating its potential for cybersecurity applications. 2025 IEEE. -
A Novel Approach to Packet Dropping and Malicious Attack Detection using Ensemble Techniques
Packet-dropping attacks interrupt data transfer while damaging security protocols, which create a threat to wireless Sensor Networks and Mobile Ad Hoc networks. This paper examines packet-dropping detection methods as well as security attack identification since these threats represent significant risks to networks such as Wireless sensor networks and Mobile Ad Hoc Networks. The research paper utilized a dataset from Kaggle for network traffic analysis, which classified packets through their behaviors as either abnormal or normal. The detection employed a stacking classifier with logistic regression as the meta-classifier and Support Vector Machine, Gradient Boosting, and K-Nearest Neighbour as its main constituents. The analysis model showed high detection rates for packet-dropping incidents, reaching 93.5%, and for malicious attacks, reaching 98.2%, based on the experimental test results. The obtained data shows that stacking models show stable reliability levels above traditional approaches. Ensemble learning proves effective for discovering cyber threats through results that reduce the number of incorrect detections. The stacking classifier functions as a dependable framework for developing security measures required to protect computer networks from modern-day threats. 2025 IEEE. -
An Explainable AI-Driven Deep Learning Algorithm for Heart Disease Detection in Healthcare
The application of preprocessed Kaggle data serves as a subject of analysis to investigate heart attack prediction capabilities through machine learning models. The research examines performance outcomes of five algorithms which consist of K-Nearest Neighbors (KNN), Random Forest, Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost) and Convolutional Neural Networks (CNN). Random Forest together with XGBoost proved as the most accurate machine learning models when used for cardiovascular risk assessment. The researchers built a hybrid structure of CNN and SVM because it improved both data classification and feature extraction processes for better prediction outcomes. The training and evaluation process of models encountered difficulties because of overfitting along with high computational expenses and problems regarding optimal hyperparameter settings. The research stresses that explainable AI (XAI) methods should integrate into systems to enhance model interpretability and achieve trust from clinical professionals. Future initiatives seek real-time patient monitoring and innovative interpretability systems for heart attack prediction to enable person-specific diagnoses and optimal clinical choices in medical fields. 2025 IEEE. -
An Algorithm for Cybersecurity Threats Detection in the Internet of Things using Deep Learning Approach
We perform research to develop a combined deep learning algorithm that enhances security threat detection within the Internet of Things networks. The resource variations across IoT devices create obstacles for Traditional Intrusion Detection Systems (IDSs) regarding their scalability and adaptability elements. This study explores the application of Bidirectional Recurrent Neural Networks and Long Short-Term Memory networks, which are trained on Traffic data records from NSL-KDD, a widely recognized benchmark dataset. It's a secondary dataset. This dataset is preprocessed and features are engineered to be optimized for sequential pattern recognition and handling of long-term dependency. Experimental results validate the achievement of a cross-validation accuracy of 93.40%, F1 is 91.62% and precision is 90.42%, which is greater than the individual models, such as CNN, BiRNN, or LSTM. The stacking Models Bi-RNN sequential learning and LSTM dependency retention makes the system perform better at threat classification along with elevated detection accuracy for IoT-related security issues like DoS, Probe, R2L, and U2R. The consistent performance of the model through this validation split provides evidence that the system can effectively handle IoT cybersecurity threats. 2025 IEEE. -
Building Smarter Systems with Advanced Computational Techniques
The biological data analysis is a key approach that uses the genetic, transcriptomics, proteomics, metabolomics, or clinical data to discover diseases. Diabetes and leukemia are two independent medical disorders, but research has found that people with type 2 diabetes have a 20% higher chance of developing blood malignancies such as acute leukemia, showing a link between the two. Early identification of these disorders by studying biological datasets is critical for providing prognostic information. However, the class imbalance and high dimensionality problems in Machine Learning (ML)based techniques have often degraded effective analysis of clinical and genomic datasets for disease detection. This paper focuses on developing an efficient clinical decision support system using advanced metaheuristic and ML algorithms to solve class imbalance and high dimensionality problems. The first stage of the proposed approach utilizes an optional data augmentation and another pre-processing method for outlier detection and removal using Modified Z-Score (MZS) based on the Median Absolute Deviation (MAD) metric. Then, the optimal features/genes are selected using a hybrid Firefly Pearson's Correlation Coefficient (FPCC)-based Feature/Gene Selection method to reduce the higher feature dimensionality problem. Once the features/genes are selected, the proposed Ladybug Beetle Optimized Universum Learning-based Twin Boosted Adaptive Support Vector Machine (LBO-ULTBASVM) classifier detects the disease with reduced model complexity and error rates. LBO-ULTBASVM is developed by improving the Twin Support Vector Machine (TSVM) classifier by integrating the Universum Learning, Ladybug Beetle Optimization (LBO), and XGBoost for solving the class imbalance problem, reducing training time and improving disease accuracy. Experiments are conducted using PIMA Indians Diabetes and GSE9476 Leukemia datasets and the outcomes indicated that the LBO-ULTBASVM-based model increases the diabetes and leukemia detection accuracy with reduced model complexity and processing time. 2025 IEEE. -
Blockchain Technology for Mitigating Copyright Infringement on OTT Platforms through the PRISMA Model
Copyright infringement has been a rampant issue on OTT (Over-the-Top) platforms and post-COVID-19 pandemic. Continuous infringement, illegal streaming, and piracy have contributed to 87 billion INR (Indian rupees) in the piracy economy through OTT platforms. Copyright and its protection regime need upscaling through Blockchain technology as a highly monetized and important asset. This paper identifies key gaps in the ongoing research on applying Blockchain technology to mitigate copyright infringement on OTT platforms through the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) model. It creates a techno-legal framework to enhance this application and make it more compliant with the existing copyright laws. The study further showcases the absence of legal implications in the existing studies surrounding Blockchain technology, making them a base for the proposed framework. The proposed framework can be further used to create a robust copyright portfolio for OTT platforms while testing their relevant advancements in Blockchain technology in line with legal and commercial necessities. 2025 IEEE. -
Artificial Intelligence in Detecting and Mitigating Online Child Sexual Abuse: Approaches and Solutions
Previous research papers have discussed whether Artificial Intelligence (AI) -based tools like Chat Bots, Law-U Model, and Sweetie. 20 have the potential to mitigate online child sexual abuse. The literature review indicates that AI tools promise good intervention and prevention strategies for several tech-based companies like Google and Microsoft. However, there is a lack of systematic study on AI tools' potential uses, limitations, and legal risks. This paper conducts a systematic literature review to explore the uses and limitations of AI-based interventions in combating online child sexual abuse. It explores the legal and ethical risks of deploying such technological innovations from the viewpoint of data protection, privacy, and security. The authors use the PRISMA technique and thematically answer the research questions. Data are collected from reliable sources such as Statista and the World Health Organisation. The findings of this paper highlight the potential uses of AI for law agencies, forensic experts, victims, and technology companies. The research reports the absence of a sufficient legal framework for the governance and accountability of AI tools. The findings further indicate the need for clarification in the law regarding the legal status of AI tools like Sweetie 2.0. Lastly, this paper offers a framework for harmonizing AI usage with human rights standards. 2025 IEEE. -
Cooperative Social Entrepreneurship Among Rural Artisans: A YouTube-Based NLP Analysis of Environmental Practices and Livelihood Outcome
This paper presents a novel data-driven study of rural artisans in Karnataka by leveraging YouTube video transcripts and natural language processing (NLP) to examine how cooperative social entrepreneurship (CSE) relates to environmental practices and livelihood outcomes. CSE initiatives in India typically rely on primary surveys to understand how artisanal groups adopt eco-friendly practices and how this affects their livelihoods, but such data are costly to collect and difficult to scale. We investigate whether publicly available video narratives can serve as a scalable secondary data source for studying CSE among rural artisans. We compile a corpus of YouTube videos on banana-fibre craft, Anegundi/Hampi artisan collectives, and Karnataka handicrafts. Audio is transcribed using an automatic speech recognition pipeline, and the resulting bilingual/multilingual text (English-Hindi-Kannada) is processed with a rule-based NLP tagger to identify three constructs central to our CSE perspective: (i) artisan and community references (CSE signals), (ii) environmental practices (e.g., "banana waste to fibre,""eco-friendly,""sustainable"), and (iii) livelihood/product mentions (e.g., baskets, mats, runners) as observable proxies for livelihood outcomes. On top of this, we apply text mining techniques topic modeling, sentiment analysis, and supervised classification with fine-tuned transformer models (BERT) to classify transcript segments (e.g., environmental focus vs livelihood focus) and extract key thematic topics. Experimental results show that our BERT-based classifier achieves over 90% accuracy, substantially outperforming traditional baselines such as TF-IDF+SVM and LSTM. The videos frequently encode both CSE signals and explicit environmental practices, and a non-trivial subset articulates marketable products, suggesting that platform narratives can partially capture the CSE-environment-outcome chain without questionnaires. However, explicit statements about market, seasonality, or constraint variables remain sparse, revealing limitations of video-based secondary data. The study contributes methodologically by integrating digital media analytics into rural development research, offers complexity and performance analysis of the employed algorithms, and stresses reproducibility through transparent documentation of data sources, model architectures, training configurations, and evaluation metrics. 2025 IEEE. -
Beyond Transcripts: A Learner-Centred Review for Closing the Graduate Skills Gap
The majority of the university graduates leave their courses with high grades, but they usually do not have the necessary skills needed in the working environments including teamwork, problem-solving and digital skills. This disconnect between higher education training and the needs of the industry is what is referred to as the graduate skills gap. The article consists of a literature review from 2020-2025 to explore how learner-centred pedagogy can be used to reduce this gap. The results have shown that project-based learning, real-life assessment, internship and micro-credential equip students better than conventional exams. Employers prefer technical and soft skills to academic performance, but most universities are facing problems with stiff curricula and lack of faculty training. This review proposes the incorporation of practice projects, industry partnership, and online skill records to fill the gap. These are some of the strategies that can be used to equip the students with the competencies needed in the current dynamically changing labour market. 2025 IEEE. -
Deep Learning for Mental Health: Attention-Driven Multilayer CNN for Audio Depression Detection
Depressive Disorder is a common mental health problem that affects millions of people around the world. This study proposes a Self-attention based Multi-layer Convolutional Neural Network (CNN) model to perform enhanced depression detection from audio modality. The model employs a diverse array of filters, kernel sizes, and pooling strategies across multiple CNN layers to capture local features, while the attention mechanism prioritizes emotionally salient parts of the speech signal, such as regions of low energy and lengthened pauses by assigning higher weights. Measured against the RAVDESS and TESS emotional speech datasets, the method attains an F1 score of 0.81, an accuracy of 83% and ROC-AUC of 0.96 when using attention, beating the baseline CNN model, F1 score of 0.77 and 83% accuracy without attention. The results demonstrate the effectiveness of attention-enhanced architectures in detecting depressive cues from speech and support the feasibility of developing real-world, speech-based mental health screening tools. 2025 IEEE. -
Hybrid Semantic Evaluation of Student Answers Using Rule Matching and BERT Embeddings
Accurate evaluation of student answers in online and traditional assessments is critical in education. In recent years, various text similarity-based methods have been proposed. However, there are certain challenges, such as the semantic and structural understanding of text. Thus, this paper uses the BERT model to present a hybrid evaluation framework that combines rule-based similarity techniques with deep semantic knowledge. The rule-based component utilizes predefined linguistic and domain-specific rules to ensure interpretability. At the same time, BERT-based similarity captures the semantic similarity and the paraphrased answers. Experimentation has been carried out on benchmark datasets, with the proposed hybrid model and human experts on manual evaluation. The performance comparison demonstrates that the hybrid model has performed significantly better than the traditional machine learning approaches in terms of accuracy and fairness of scoring. The proposed hybrid model is also compatible with deployment in educational platforms as it provides suitable feedback to learners. 2025 IEEE. -
Measuring Critical Thinking Skills with the R-BiLSTM-C Model using a Logical Approach
Critical thinking is essential for making informed judgements; it necessitates careful evaluation of pertinent evidence and the application of reasoning. While some individuals possess a more constrained perspective on critical thinking, this elucidation encompasses the predominant views held by the majority. Note-taking, formulating enquiries, and designing experiments exemplify practical actions that may be applied to other creative pursuits, rendering it a valuable skill across diverse domains. This work employed a systematic approach for data preparation, model training, and feature extraction. The primary phase in training unified R-BiLSTM-C models was identified as feature extraction. Standardisation and normalisation, two crucial preprocessing techniques, were employed to ensure uniform and dependable data handling. Furthermore, as the quantity of dependent observations escalated, the study evaluated the efficacy of Reduced Kernel PCA. The proposed solution achieved a 96.82% accuracy rate, surpassing advanced techniques such as CNN and BiLSTM. The findings indicate that the systematic approach enhances the model's performance. A systematic approach is essential for enhancing precision, and analytical reasoning skills are crucial for developing effective machine learning models. The study reinforces the significance of critical thinking in effective decision-making and problem resolution. 2025 IEEE. -
An Intelligent Method for Fraud Detection in Digital Payments based on SVR with GC-RF Approach
The use of automated algorithms to detect fraud on electronic payment networks is challenging. Digital payment systems and their users are vulnerable to cybercriminals who take advantage of security holes or users' negligence to steal passwords, perpetrate fraud, launder money, and carry out other malicious acts. Conventional methods of fraud detection are challenging to execute because of the difficulties of acquiring massive volumes of manually annotated data. It is tough to notice new trends because fraudsters are often changing their techniques. Feature extraction, model training, and data preprocessing were the main areas of emphasis in this systematic research. Data pretreatment encompassed tasks such as acquiring training sample data, cleaning, converting, integrating, and altering the data. Feature extraction is the backbone of SVR-GC-RF model training; it takes all the data in a dataset and turns it into features. The suggested method outperformed SVR and RF in terms of accuracy by 95.23 percent. The importance of hierarchical fraud detection in online payment systems is highlighted in this paper. Through the use of effective feature extraction and model training, the study enhances fraud detection. Methods for detecting fraud need to change if they are to keep up with the criminals. 2025 IEEE. -
Factors Affecting Predicting Teacher Evaluation in Higher Educational Institutions Among Faculty Members based on SA-BiLSTM
To keep up with good teaching standards and pedagogical improvements, it is vital to predict assessment of higher education teachers. Based on principles of pragmatism, psychology, and pedagogy, this research should contribute to the development of diagnostic procedures and training standards that may be used in many educational settings. Because of the three-tiered speciality training architecture used to prepare teachers in master's degree programs in education in Ukraine, institutions there require a differentiated approach. Model training, feature extraction, and preprocessing are all parts of the proposed methodology. While translating, cleaning, reducing, and normalising data is part of data preparation, feature selection uses correlation and mutual information criteria to establish the significance of variables. Following feature selection using information gain, the SA-BiLSTM model begins training. Different institutional features were associated with unique approaches to teacher preparation, according to a comparison of active empathy and the growth of pedagogical reflection. The SA-BiLSTM model outperforms SA and BiLSTM when it comes to predicting teacher ratings. The results show how important it is to have different levels of individualised strategies for preparing teachers. Institutions can improve training and instruction with the support of the predictive approach, which better evaluates teachers. 2025 IEEE. -
Quantum-Inspired Genetic Algorithms for Secure and Scalable Cloud-Based Decision Support Systems
Cloud-based DSS are critical for data-intensive decision-making but tremendously challenged by issues of scalability, security, and the optimization of resources. In general, optimization approaches such as GA, PSO, and ACO treat the problems of allocation and security enhancement of cloud resources very inadequately. Hence, the present work addresses developing a QIGA-based optimization framework for the performance optimization of cloud DSS. At its core, it utilizes quantum-like principles, such as superposition and probabilistic search, for resource optimization with respect to stability, security, and rapid convergence. Therefore, this underpinning framework comprises resource optimization, anomalous detection, and quantum-independent encryption through QIGA, which enhances the data security along with computational efficiency. The experimental results depict the performance efficiency of QIGA since its execution time, CPU, memory utilization requirements, and energy consumption are less while the task completion rate is higher and the security vulnerabilities are reduced in comparison to traditional optimization techniques. QIGA-based anomaly detection improves accuracy at the expense of response time, while its quantum-inspired encryption provides the best cryptographic security. Therefore, these results verify that QIGA is an efficient secured methodology for scalable cloud-controlled DSS, hence being a potential candidate for decision-making optimization in highly dynamic cloud environments. 2025 IEEE. -
Crop Disease and Pest Management in Agriculture via UAV Remote Sensing and Advanced Machine Learning Models
Pests and diseases greatly reduce crop quality and yield; therefore, IA relies on effective pest and disease control. UAVs have become a crucial remote sensing (RS) tool for agricultural process monitoring and management. This study will examine major advances in this field using bibliometric methodologies including author co-occurrence and keyword co-contribution studies. The suggested technique involves preprocessing, feature extraction, and model training. Data quality improves with preprocessing. UAV images are used for feature extraction, focusing on canopy structure and height. PPO is trained the prediction model. Compared to ultramodern GANs and LSTM networks, the recommended model wins. The model consistently outperforms competitors with 91.17 percent accuracy. The study suggests employing UAVs in smart farming to reduce pests and diseases. The suggested model's accuracy and reliability improve crop quality and production by solving agricultural monitoring and management problems. 2025 IEEE. -
Enhancing Human Resource Management Practices in Marketing Companies Using Dual Graph Attention Networks
Marketing organization features and strategy implementation have been studied for over 30 years. These include organizational structure, culture, leadership, and processes. HR regulations can motivate marketing professionals to support group and individual goals when correctly implemented, but this part of HR has gotten little attention. Model preparation, feature extraction, and training comprise the suggestive technique. It reviewed data quality, evaluated dataset structure, and described data types during pre-processing. Principal component analysis (PCA) ranked and evaluated decision-making units to reduce dimension. Model training used MGGAN. In comparison to GAN and CNN, the proposed model performed well. With an average accuracy rate of 94.36%, it surpassed earlier approaches and captured all dataset peculiarities. MGGAN modeling can increase predictive performance, and marketing organizations should integrate HR regulations, according to this study. This study opens up new organizational analysis and strategy execution methods. 2025 IEEE. -
User Behavior Prediction with Deep Learning: An Evaluation of CNN, LSTM, RNN, and Hybrid Models
In this work, with a novel Hybrid CNN LSTM model, we present an in depth comparison and evaluation of various deep learning models (CNN, LSTM, RNN) on predicting user behaviour. To compare our proposed models we conducted detailed experimentation and rigorous performance comparisons using aforementioned metrics: accuracy, precision, recall, F1-score. The findings show that our RNN model performed incredibly well, reporting an accuracy of 99.86%. These results illustrate the power of the model in generalizing, and in capturing sequential dependencies in the user behaviour. On the other hand, the CNN model also performed very robustly, achieving 97.86% accuracy showing that it has powerful ability to extract spatial features. However, the LSTM model got decent results, with an accuracy of 87.14%. However, the Hybrid CNN-LSTM model's accuracy was for 80.29% which was up to the Hybrid CNN as mentioned above, and hence had lagged behind and has scope for improvement. Our proposed approach has provable advantages over existing methods, most notably, by using the RNN model outperforming the existing method. The RNN is able to detect sequential patterns more effectively than traditional techniques and show its ability to learn deeply to better predict user behaviour. Finally, we also hope that this investigation highlights two important directions for future RNN research: first, the promise of RNNs, and second, promising avenues for future research, such as refining hybrid architectures to yield improved performance. 2025 IEEE. -
A Computing Assisted Test Method in Healthcare Industry Using Artificial Intelligence
New and improved methods of diagnosis are needed because breast cancer is still the leading cancer-related killer worldwide. Updates to the methods used to categorize breast cancer have emerged because of recent advances in DL and ML. This research goal in conducting this research is to bring together existing breast cancer diagnostic and classification methods that make use of deep learning and machine learning techniques. Early detection and accurate prediction of breast cancer are crucial for improving outcomes and reducing the impact of this disease. The creation of prediction models and tools to assess risk has become an important field of research as it can assist healthcare workers in identifying individuals who are more likely to acquire breast cancer and adapting screening and preventative programs accordingly. This introduction provides an overview of breast cancer prediction, highlighting the importance of the topic and the motivation behind predictive models. It sets the stage for a more in-depth exploration of the subject and the various technologies, methods and factors involved in breast cancer prediction. 2025 IEEE.