Fusion: Practice and Applications

Data imbalance is a common problem in machine learning, specifically in classification, in which examples in a dominant class outnumber examples in a minority class many times over. Besides, such a problem keeps a model unable to discover meaningful patterns for a minority class —hence, such a problem reduces model performance specifically in terms of Recall and F1-Score.  In current work, activity is performed in overcoming data imbalance problem in sentence classification model of documents of assurance certificate for halal with a combination of over-sampling and under-sampling techniques, namely Adaptive Synthetic (ADASYN) and Tomek Links. Text Classification technique is adopted in classifying sentences regarding assurance of halal in documents of assurance certificate for halal Text Classification; since incorrect classification of such sentences is not preferable, therefore, it is important to make sure no information about halal product is missed out. Over-sampling techniques considered include the SMOTE, Borderline SMOTE, ADASYN, and SMOTENC, and under-sampling techniques include the Random Under-Sampler, Near Miss, and Tomek Links. As comparative result, best performance gain in terms of Accuracy (0.759), F1-Score (0.748), Recall (0.759), and Precision (0.768) is generated with ADASYN. In our use case, ADASYN + Tomek Links is effective; recall is important in case of classification of documents for assurance certificate for halal and therefore, we cannot miss any relevant sentences. The proposed approach remarkably enhances the accuracy level for halal-related sentence identification and can be adopted in the halal product checking systems in industries with a halal feature.

In the context of dynamic and highly diverse IoT (Internet of Things), the nature of threats and the amount of data that needs to be processed by IDSs (Intrusion Detection System) have become much greater and represent considerable problems for modern security systems. This work presents a new method called a Hybrid Deep Learning-Evolutionary Algorithm with an Ensemble Strategy (HDLE-EASE) for improving intrusion detection in IoT systems. Our method combines the spatial feature extraction capability of CNN (Convolutional Neural Networks) and temporal feature extraction of LSTM (Long Short-Term Memory) networks with the optimization feature of GA to optimize model parameters. We further incorporate a composite of boosting-bagging hybrid to enhance the stability and reliability of the intrusion detection mechanism. As privacy and scalability are critical issues in IoT networks, we propose a federated learning approach, allowing for model training on IoT networks while preserving data privacy. Furthermore, the presented approach includes a reinforcement-learning module for the capability of adapting to newly emerge and changing security threats. Initial tests show that the detection accuracy and model optimization capabilities of HDLE-EASE significantly outperform other methods, while its adaptability makes the tool a promising one for developing a holistic solution to protect IoT systems against a wide range of attacks.

As optimization tasks become increasingly complex, particularly in feature selection, there is a growing need for algorithms capable of robustly balancing exploration and exploitation. In this work, we propose the Binary Swordfish Movement Optimization Algorithm (BSMOA), inspired by the synchronized and agile movements of swordfish. BSMOA employs adaptive parameters to navigate high-dimensional search spaces through dynamic exploration, exploitation, and elimination stages. Extensive experiments on benchmark datasets demonstrate that BSMOA outperforms state-of-the-art algorithms, including bHHO, bGWO, and bPSO, regarding average error, feature reduction, and computational efficiency. Key contributions of BSMOA include its improved balance between global and local search and its ability to achieve stable and accurate feature selection. This work has broad implications for applications in machine learning, engineering design, and other optimization domains, providing a reliable tool for tackling challenging binary optimization problems.

Glaucoma is a dangerous eye illness that greatly reduces the sharpness of a person's vision. If not caught early enough, this retinal disorder can damage the optic nerve head (ONH) and cause permanent blindness. Automated glaucoma diagnosis now has tool support thanks to recent advances in deep learning besides the convenience of computing resources. The low reliability of generic convolutional neural networks has prevented their widespread usage in medical procedures, even if deep learning has improved illness diagnosis using medical pictures. While there has been a rise in the use of deep learning for glaucoma classification, very few studies have tested whether or not the models are easy to understand and interpret, which bodes well for their future use. Medical picture feature extraction using Vision Transformers is showcased in this study utilising an EvoTransform: Advanced Evolutionary Algorithm Integration in Transformer Networks named as (EvoTAEA). Combining the powers of Convolutional Neural Networks with Vision Transformers, the suggested EAT Former architecture takes advantage of their data pattern recognition in addition adaptability capabilities. The classification accuracy is enhanced by using the Wild Geese Migration Optimizer (WGMO) to fine-tune the parameters of the proposed feature extraction. The design makes use of new parts, such as the Multi-Scale Region Aggregation, Global and Local Interaction, and Enhanced EA based Transformer blocks with Feed-Forward networks. For dynamically simulating non-standard places, it also presents the Modulated Deformable MSA module. Important components of the Vision Transformer (ViT) model are covered in the study, including patch-based processing, Multi-Head Attention mechanism, and positional context inclusion. In order to give an inductive bias, it presents the Multi-Scale Region Aggregation module, which combines data from several receptive fields. The MSA-based global module is improved by the Global and Local Interaction module, which adds a local path for extracting discriminative local info. An approach to glaucoma diagnosis that integrates ResNet-50, DenseNet-201, and Xception is suggested in the study. Two publicly available datasets, ORIGA and ACRIMA, are used to evaluate the trials. This model can help with the automated diagnosis of glaucoma using fundus pictures.

The E-Government Development Index (EGDI) represents the performance and reality of e-government. The importance of maintaining and planning for the enhancement of such an index enables the policymakers to understand, process, and develop the right plans and strategies for it. In this paper, the Auto Regressive Integrated Moving Average (ARIMA) has been utilized to build predictive models. The time-series data collected from the UN survey versions for the years 2003, 2005, 2008, 2010, 2012, 2014, 2016, 2018, 2020, 2022, and 2024 for the countries of Iraq and KSA. The necessary data maintenance was implemented, then analyzed, covering the inspection of their temporal behavior. Afterwards, two individual data sets were created for both countries under study, containing 253 months. The optimal values ​​for the ARIMA models were determined by implementing the data transformation, including the autocorrelation function (ACF) and partial autocorrelation function (PACF). 80% of the dataset is used for training, and 20% is used for testing. The data residuals analyzed by ACF, PACF, and the Ljung-Box test were performed for the residuals independence check. Nine metrics were utilized for model evaluation and ruthlessness. By using ARIMA models, the e-government performance (EGDI) has been predicted for the next five years for Iraq and KSA. The ARIMA models for both Iraq and KSA showed high performance, where the RMSE value for the Iraq model was (0.0054) and the MAE value was (0.0031) compared to the RMSE value (0.0481) and the MAE value (0.0093) for the KSA model. The Iraq arima model has better quality of the prediction in absolute terms. On the other hand, the ARIMA model for KSA was better in terms of predicted trends with an accuracy of 98.44% compared to 97.39% for the Iraq model.

The sustainable development of automobile enterprises in Uzbekistan is a topic that deserves our attention and analysis. In this study, we focus on utilizing fusion analysis to identify and understand the factors that play a crucial role in the sustainable growth of these enterprises. By examining and fusion, multiple dimensions, such as economic, environmental, social and technological factors, we aim to provide valuable insights and recommendations for promoting sustainable within the automobile industry in Uzbekistan. Through fusion analysis, we examine how economic factors, such as market demand, production efficiency, and financial viability, influence the sustainable development of the country and automobile enterprises.

Cuneiform script, among the earliest writing systems, poses a distinct challenge for classification because of its complex symbols and varied linguistic contexts. This study investigates the use of Convolutional Neural Network (CNN) architectures for the classification of cuneiform symbols. The preprocessing includes resizing the cuneiform images to a uniform dimension and categorizing them into training, validation, and testing sets. A modified CNN model has been introduced. The CNN model demonstrates a lower parameter count in comparison to other deep learning models, which frequently necessitate significant storage capacity. The results from the CLI dataset indicate that the proposed CNN model reached an impressive accuracy of 99.55%, This study enhances computational approaches for the analysis of ancient scripts and underscores the significance of utilizing deep learning techniques within the fields of historical linguistics and digital humanities.

This study aimed to identify the most relevant factors influencing the effectiveness of self-assessment and external evaluation processes in higher education in Ecuador. Through an analytical approach, the DEMATEL method integrated with neutrosophic logic was employed to evaluate interactions, prioritize these factors, and enhance information fusion in decision-making. The methodology allowed for the incorporation of inherent uncertainty and subjectivity in evaluation, generating a more adaptive and robust model for integrating multiple sources of information. The results revealed that key factors included the clarity of quality indicators, institutional commitment to continuous improvement, training of evaluators, and institutional infrastructure. Furthermore, the study highlighted that the fusion of internal and external evaluation data is crucial for a comprehensive quality assessment. The most influential factors within the system were identified as the impact of evaluation results on decision-making and infrastructure quality. Findings indicate that improving educational quality in Ecuador requires strengthening data integration mechanisms, ensuring coherence between self-assessment and external evaluation, and optimizing the interaction between different quality assurance processes. It is recommended to enhance information fusion strategies in quality assurance policies to improve the efficiency and accuracy of evaluation processes in higher education.

Melanoma Skin cancer is a serious type of cancer affecting people globally in order to improve survival rates, it is crucial to detect the infection at an early stage. Old Traditional methods for cancer detection make use of biopsies, which were time-consuming and involved complex procedures, which delayed diagnosis. However, accurate diagnosis is challenging due its complex imaging techniques. With the advancements in technology, particularly in deep learning techniques like CNN, have significantly improved the accuracy and efficiency of melanoma skin cancer detection. This research paper presents a Novel Hybrid deep learning architecture that combines Convolution Neural Networks (CNNs) and Vision Transformers (ViT) for automated classification of skin lesions into binary categories: Malignant (cancerous) and Benign (Non-cancerous). The proposed model influences CNN's superior ability to extract local features alongside ViT's capability to extract global features. This hybrid architecture was trained and evaluated on ISIC 2020 challenging Dataset of dermatological images representing excellent performance with an accuracy of 94%, with a precision of 91%, recall (sensitivity) of 90%, and an F1 score of 91% after 25 epochs.  The model's robustness is further authorized through confusion matrix analysis, which forms a strong classification capability across various melanoma presentations. The proposed hybrid approach offers a more efficient and less complex approach in the automatic detection and identification of melanoma skin cancer, thus increasing the chances of successful early intervention and improving patient outcomes, thus making it suitable for Clinical use and sets a foundation for future developments in automated skin cancer detection systems. In comparison to other advanced networks, this model displays superior performance.

The examination of commercial bank deposits together with their influencing factors relies on econometric analyses in this paper. The econometric model for commercial bank deposit base factors used a multiple linear regression (LS) method because the data came from time series that included multiple variables. The research used 74 economic indicators spanning an eight-year period and collected those indicators in monthly intervals. The dependent variable was the deposit volume (y), while the independent variables were the inflation rate (x1), the minimum wage (x2), the number of individuals using digital banking services (x3), the average interest rate on term deposits (x4), and the per capita GDP (x5). Our analysis, based on data from the Central Bank of the Republic of Uzbekistan, indicates that the selected independent variables are significantly related to the growth of the deposit base. The implementation of multiple linear regression (LS) answered Gauss-Markov assumption tests successfully while the Durbin-Watson test and Shapiro-Wilk test along with the Breusch-Pagan test evaluated the statistical import of the obtained results. The key findings indicate that a 1% increase in the inflation rate leads to a 1.06% decrease in the deposit volume; a 1% increase in the minimum wage results in a 0.32% increase in the deposit volume; a 1% increase in the number of individuals using digital banking services leads to a 0.59% increase in the deposit volume; a 1% increase in the average interest rate on term deposits results in a 0.81% increase in the deposit volume; and a 1% increase in per capita GDP causes a 0.79% increase in the deposit volume. Banks should concentrate their efforts on fighting inflation while developing their digital systems because these strategies build a better deposit base, which boosts interbank rivalry and supports economic stability.

Wind energy is one of the fastest-growing sustainable, clean, and renewable sources, attracting significant attention and investment from many countries. However, given the substantial capital investment required for wind power plants, understanding the proposed plants’ performance becomes critical before implementation. This assessment is most effectively conducted using refined wind power predictability models and precise wind velocity data. Accurate wind forecasts are essential for informed decision-making and effective wind energy utilization. In this study, three advanced Machine Learning (ML) regression methods were applied to the TNWind dataset to predict the power output of wind turbines. The dataset variables included date and time (measured at 10-minute intervals), low-voltage active power (in kW), wind speed (in m/s), the theoretical wind power curve (in kWh), and wind direction. To predict wind power output, six supervised ML models were trained, including Random Forest Regressor (RF), Extreme Gradient Boosting Regressor (XGB), Gradient Boosting Regressor (GB), Support Vector Machine Regressor (SVR), K-Neighbors Regressor (KN), and Linear Regressor. The analysis revealed that the Random Forest model outperformed the others, achieving exceptional performance metrics: an R2 value of 0.97, an MAE of 0.17 and an MSE of 0.07. The analysis to identify the outcomes for wind power generation from machine learning proves that renewable energies are more capable and are a lucrative investment.

Network intrusions are becoming more common, resulting in significant privacy violations, financial losses, and the illegal transfer of sensitive information. Numerous intrusion strategies pose a threat to data, computer resources, and networks. While hackers may focus on obtaining trade secrets, private information, or confidential data that can then be disclosed for illegal purposes, each type of intrusion aims to achieve a distinct objective. False attack detection by security systems and changing threat environments create challenges such as delayed identification of true attacks and long-term financial harm. This paper presents a novel hybrid optimization algorithm-assisted deep learning model for accurately identifying intrusion types and enhancing network security. Initially, input information is composed from openly obtainable datasets. The input data is cleaned, normalized, and standardized to produce accurate results. An improved synthetic minority oversampling technique (ISMOTE) for data balance reduces the method's overfitting problem. Then, the Chaotic Bat Artificial Bee Colony optimization algorithm (CBABCOA) is used to identify critical features and reduce feature dimensionality issues. HSVM-XGBoost (Hybrid Kernel Support Vector Machine-Extreme Gradient Boosting) is used for intrusion detection and classification. The Chaotic Binary Horse Optimization Algorithm (CBHOA) is used for hyper parameter tuning. This method makes use of the CIC UNSW-NB15 Augmented dataset, the CICIDS 2019 data set, and the NSL-KDD information set. The proposed method achieves better than the other method.

One of the significant challenges in modern healthcare is the early and accurate detection of health anomalies, especially in the case of life-threatening diseases such as breast cancer. This paper investigates the comparative efficacy of ML models and statistical methods for the classification of breast tumors as benign or malignant using the Breast Cancer Wisconsin (Diagnostic) Dataset. The dataset, comprising various tumor cell attributes, was preprocessed with Principal Component Analysis (PCA) to enhance model training efficiency. The first 11 principal components retained 95% of the total variance, ensuring minimal information loss while reducing dimensionality. We compared the performance of several machine learning algorithms, including Logistic Regression, Support Vector Machines (SVM), Artificial Neural Networks (ANN), Decision Trees (DT), Random Forests (RF), Naïve Bayes (NB), and K-Nearest Neighbors (KNN). Among them, Logistic Regression, SVM, and ANN achieved near-perfect classification accuracy with balanced precision-recall metrics, where the accuracy rates were all more than 98%. XGBoost and Random Forest were also very impressive as advanced models, while simple models like Decision Trees and Naïve Bayes proved to be less potent and were unable to manage class imbalances and complex data patterns. Our main findings are essentially reflective of the transformative role machine learning would play in healthcare; for instance, enhancing the accuracy of diagnosis, optimizing clinical workflow, and promoting decision-making. These insights are made actionable for practitioners in healthcare to promote the adoption of reliable ML solutions for breast cancer detection. In the future, real-time data integration, external validation, and hybrid modeling approaches must be considered to further enhance the practical utility of these findings.

In autonomous navigation, the ability to detect 3D objects from a Bird’s-Eye View (BEV) perspective is essential. Nevertheless, many obstacles remain before LiDAR and camera data can be effectively combined. We propose CL-FusionBEV, a novel framework for sensor fusion that enhances Three-dimensional object recognition in the BEV domain. This method structures LiDAR point clouds for improved spatial feature extraction while converting camera data into BEV format via an implicit learning technique. An implicit fusion network and a multi-modal cross-attention mechanism facilitate seamless sensor interaction, ensuring comprehensive feature integration. Additionally, a self-attention mechanism of BEV enhances broad-scale reasoning and data extraction, improving the detection of occluded and distant objects. By efficiently synchronising data from several sensors, the suggested method improves feature uniformity and resolves spatial inconsistencies. It further leverages adaptive feature selection to enhance robustness against sensor noise and varying conditions. We evaluate CL-FusionBEV on the nuScenes dataset, achieving achieved a 73.3% mAP and a 75.5% NDS on the nuScenes benchmark, with vehicle and pedestrian detection accuracies of 89% and 90.7%, respectively. Our model demonstrates superior robustness in challenging conditions such as low visibility and dense urban environments. CL-FusionBEV maintains high efficiency with real-time inference, making it suitable for deployment in autonomous systems. Extensive experiments show our strategy routinely beats cutting-edge techniques, especially in detecting small and distant objects. By addressing key sensor fusion challenges in the BEV domain, CL-FusionBEV offers a notable advancement in Three-dimensional object recognition, ensuring high accuracy, efficiency, and reliability for real-world driving scenarios.

A strong sign language recognition system can break down the barriers that separate hearing and speaking members of society from speechless members. A novel fast recognition system with low computational cost for digital American Sign Language (ASL) is introduced in this research. Different image processing techniques are used to optimize and extract the shape of the hand fingers in each sign. The feature extraction stage includes a determination of the optimal threshold based on statistical bases and then recognizing the gap area in the zero sign and calculating the heights of each finger in the other digits. The classification stage depends on the gap area in the zero signs and the number of opened fingers in the other signs as well as the sequence in which the opened fingers appear for those that have the same number of opened fingers. The conducted test results showed the system’s high capability to classify all the digits; where both the precision and F-score percentages of the proposed model reached the desired optimal value (100%).