In domestic wastewater treatment plants, wastewater purification is typically carried out through biological processes based on bacteria. For the microorganisms in these systems to survive and function effectively, certain environmental and chemical parameters must be maintained within optimal ranges. These parameters can be monitored at minute, hourly, or daily intervals. However, the laboratory determination of Biochemical Oxygen Demand (BOD), a critical parameter for assessing bacterial activity in advanced biological treatment plants, typically takes 3 to 5 days to obtain results, significantly longer than for other parameters. In this study, the aim was to estimate BOD values using process parameters that can be measured at minute or hourly intervals in a wastewater treatment plant. To develop the predictive model, several Machine Learning algorithms, including Artificial Neural Network, Multiple Linear Regression, Gradient Boosting Regressor, XGBoost, and Random Forest, were employed. The performance of these models was evaluated using the Coefficient of Determination, Root Mean Square Error and Mean Absolute Error metrics. According to the results, the Multiple Linear Regression algorithm achieved the highest predictive performance. The model achieved R² = 0.66 with direct imputation and R² = 0.91 with model-based imputation. These findings demonstrate that machine learning algorithms provide an effective alternative for estimating the BOD parameter, particularly when laboratory analyses require extended processing times.
Biometric authentication has become increasingly important in security and identity management applications. This study proposes a multi-stage biometric system capable of performing both person verification and gender detection using single-hand images. The proposed framework integrates multiple feature extraction and classification strategies within a unified experimental setup. Three feature extraction methods (Zernike moments (ZM), wavelet transform (WT), and Hough transform (HT)) are employed in single and hybrid combinations, yielding five distinct feature extraction configurations. The classification stage includes five classical machine learning models, a triple hybrid ensemble model, and one deep learning model, including Decision Trees (DT), Gradient Boosting (GB), Gaussian Naive Bayes (GNB), Random Forest (RF), Support Vector Classifier (SVC), a triple hybrid model (DT+GB+GNB), and a Deep Neural Network (DNN), enabling a comprehensive comparative analysis. Experimental results show that the SVC and DNN models achieved the highest performance, yielding person verification accuracy of 97.86% and gender detection accuracy of 98.00%, respectively. The findings of this study demonstrate the feasibility of a unified hand-based biometric framework and provide valuable insights for the design of robust and scalable biometric authentication systems.
A new model based on the combination of artificial neural networks and genetic algorithm for anemia prediction problems is presented. The goal of the algorithm to estimate anemia types by using data collected from laboratories based on blood variables for 560 subjects. ANN is used to classify the types of anemia and then use the GA to enhance the performance of the ANN to increase the classification accuracy. This process is carried out by enhancing the weights of neurons for the ANN to be more suitable for recognizing anemia types. The results confirm that the performance of the combined algorithm with a hidden layer is the most appropriate architecture among the networks with a different number of hidden layers. Results showed that the hybrid algorithm provides a highly precise result for the classification of anemia types, where the algorithm has an accuracy of 95.1% and the baseline algorithm has an accuracy of 91.8%, then the proposed algorithm may be considered to be clinically significant. Those results confirm that predicting anemia types is of great importance in health-related education and applications. Based on the biophysical factors, this study is provided a better understanding of the importance of predicting anemia.
This study investigates grape leaf disease classification using both hand-crafted feature extraction methods and deep learning models. Principal Component Analysis (PCA) and Local Binary Patterns (LBP) are employed as hand-crafted approaches, while Convolutional Neural Networks, including VGG16, ResNet-50, AlexNet, GoogLeNet, MobileNetV2, and EfficientNet-B0, are utilized as deep learning methods. Experiments are conducted on two publicly available datasets, Grape400 and PlantVillage. The results demonstrate that deep learning models consistently outperform hand-crafted methods across both datasets. VGG16 achieves the highest accuracy of 98.75% on Grape400 and 99.51% on PlantVillage, while MobileNetV2 and EfficientNet-B0 also show competitive performance with accuracies exceeding 97%. In contrast, PCA and LBP yield notably lower accuracies, particularly on the Grape400 dataset. These findings confirm the effectiveness of transfer learning-based deep learning architectures for robust and accurate grape leaf disease classification, highlighting their potential application in precision agriculture.
This paper presents a novel ensemble reinforcement learning framework for multi-asset portfolio management, referred to as the Confidence-Weighted Dynamic Ensemble (CWDE). The proposed model integrates five state-of-the-art actor–critic algorithms—PPO, A2C, DDPG, TD3, and SAC—under a dynamic aggregation mechanism that adjusts model weights based on entropy-derived confidence and historical performance. Using a diversified dataset spanning equities, bonds, commodities, and real estate ETFs, CWDE is benchmarked against its constituent DRL agents. Experimental results demonstrate that CWDE outperforms all baselines, achieving the highest risk-adjusted returns and the lowest drawdowns. Statistical analysis confirms the ensemble’s robustness and adaptability to market volatility. The findings highlight CWDE’s potential to serve as a scalable and interpretable framework for trading intelligence. The study concludes with a discussion of computational and practical limitations and outlines future directions for integrating explainability, macroeconomic features, and real-time deployment.
The dynamic and resource-constrained nature of Internet of Things (IoT) environments demands software systems that are both adaptive and transparent. Traditional static architectures fall short in addressing these challenges, particularly in scenarios requiring real-time decision-making and energy efficiency. This paper presents a modular and explainable self-adaptive software architecture that leverages Deep Q-Learning (DQN) for intelligent adaptation and integrates hybrid explainability mechanisms using SHAP and LIME. The proposed architecture supports runtime decision-making through a layered design, enabling context-aware sensing, battery-aware adaptation, and interpretable actuation. A custom reward function optimizes energy consumption, latency, and service quality, while explainability modules provide insight into both global learning behavior and individual runtime decisions. Experimental evaluation in a simulated smart home environment demonstrates the effectiveness and transparency of the system, highlighting its potential for deployment in constrained and safety-critical IoT settings.
Trajectory prediction remains a significant operation in mobile communications. In 5G and Beyond (B5G) networks, next-cell prediction for User Equipment (UE) becomes increasingly critical amid the exponential network complexity driven by unprecedented subscriber growth. Markov Chains are selected for their simplicity, interpretability, low computational demands, and proven effectiveness in modeling sequential mobility patterns, which makes them ideal for real-time predictions, despite the existence of more complex alternatives. Paralleling the growing interest in metaheuristic (MH) algorithms for parameter optimization, this paper employs the Whale Optimization Algorithm (WOA) to select the optimal Markov Chain order for each UE trajectory, thereby enhancing next-location prediction. Compared to the traditional fixed-order Markov Chain, the proposed method boosts average prediction accuracy by 20%.
In today’s digital age, malware constantly evolves and becomes more sophisticated. Traditional malware identification techniques are not designed to address the threats posed by evolving next-generation malware. The threats include, but are not limited to, system damage, data theft, privacy breach, financial loss or disruption of operations. Deep learning techniques can be used to detect and classify this new generation of malware. Geometric deep learning (GDL) methods leverage graph neural networks (GNNs) and are recognized for their enhanced representation learning and superior generalization capabilities compared to conventional Deep learning (DL) approaches. Experiments in this study assess the effectiveness of GDL algorithms for malware identification. Convolutional Neural Network - Long Short-Term Memory (CNN-LSTM) networks are contrasted with three GNN models: Graph Convolutional Network (GCN), Graph Attention Network (GAT), and GraphSAGE Network (GraphSAGE). The findings demonstrate that two out of three GDL models, GCN and GraphSAGE, except for GAT, outperform with a significant gain under various conditions that are proven in experiments. The research demonstrates the superior performance of GDL techniques over traditional DL for effective next-generation malware identification.
Machine learning classifiers deployed in high-stakes domains like healthcare and finance face the dual challenges of class imbalance and asymmetric misclassification costs, which are poorly addressed by traditional evaluation metrics. The primary purpose of this study is to address this critical gap by developing and validating the Cost-Weighted Harmonic (CWH) score, a novel, bounded performance metric that unifies precision, recall, and specificity within a normalized harmonic mean, explicitly weighted by a user-defined cost ratio, for high-stakes imbalanced classification. Unlike cost-agnostic metrics (e.g., F1, HMRS) or unbounded cost-aware scores (e.g., C-score), CWH is interpretable, stable, and aligns evaluation with domain-specific risk priorities. It is integrated with threshold optimization and validated across healthcare, cybersecurity, and financial datasets, demonstrating superior stability and up to 69% performance improvement against C-score in life-critical scenarios without excessive false positives. CWH effectively bridges the gap between statistical evaluation and operational decision-making, offering practitioners a reliable tool for model selection that aligns with domain-specific risk priorities.
Among all oncology disorders, brain tumors are among the most complex and lethal, necessitating rapid and accurate differential diagnosis. Traditional manual examination of Magnetic Resonance Imaging (MRI) scans is time-consuming and subject to significant variability among radiologists. This study evaluates five deep learning architectures — InceptionV3, ResNet50, VGG16, InceptionResNetV2, and EfficientNetV2L — for the automated classification of brain tumors into glioma, meningioma, and pituitary tumor types using the Figshare Brain Tumor Classification dataset. The dataset was split into 60% training, 20% validation and 20% testing sets with class weights addressing class imbalance. Models were initialized with ImageNet pre-trained weights, with custom layers (GlobalAveragePooling2D, dense layers with 1024–4096 units, and dropout) added for feature extraction and classification. They were trained for 30 epochs using the Adam optimizer. The experimental results show that the InceptionV3 model achieved an accuracy of 92.10% and also demonstrated better Grad-CAM performance for the effective localization of tumor regions than the other four algorithms. These results highlight the potential of deep learning, particularly InceptionV3, to enhance diagnostic accuracy and efficiency in brain tumor classification. Future research should focus on refining model architectures and optimizing computations for real-time clinical applications.
Online review platforms have become crucial decision-making tools in the hospitality industry, where automated sentiment analysis and rating prediction offer valuable insights for both businesses and consumers. This study investigates the performance of transformer-based language models for predicting hotel review ratings and examines the impact of oversampling techniques on model accuracy. We introduce a novel dataset of 68,785 English hotel reviews from TripAdvisor (2014-2023) in Turkey. Four transformer models, i.e., BERT, DistilBERT, RoBERTa, and DeBERTa, were systematically compared using multiple perspectives. Results show DeBERTa achieves the highest performance among all evaluated models. Random oversampling (ROS) significantly improved classification performance, with F1-scores increasing from 62% to 81% and accuracy from 76% to over 82% across all models. The oversampling approach effectively addressed class imbalance while preserving semantic information, enabling better distinction between rating categories. Through quantitative and qualitative analysis, including the embedding of visualization and SHAP-based interpretability studies, we demonstrate that transformer models effectively capture sentiment patterns. However, they remain sensitive to mixed sentiments and linguistic subtleties. This work contributes a novel dataset, a systematic comparison of four transformer models, and empirical evidence of oversampling effectiveness in sentiment analysis.
Deep Neural Networks (DNNs) have attained remarkable prediction outcomes in image classification tasks, leading to significant progress in computer vision applications. However, the presence of adversarial examples has emerged as a critical challenge to the robustness and efficiency of deep learning-based image classifiers. Adversarial examples are specially designed perturbations applied to input images to deceive the models into generating inaccurate predictions while appearing indistinguishable to human observers. In this paper, we present a defense mechanism, namely Defensive Distillation with Gaussian Blurring (DDGB), that improves the robustness of deep learning models towards adversarial attacks. First, two models, a teacher and a student model, were utilized to train and validate the presented approach. The teacher model is trained and then leveraged to determine softened probabilities, which are later utilized to train the student model. Second, a feature-squeezing technique based on Gaussian blurring is applied to the adversarial examples generated from the distilled student model as a form of defense mechanism to make the adversarial perturbations less effective. The obtained findings demonstrate that the proposed approach is effective in improving the performance, achieving classification accuracies of 87.61% and 87.48% using the Fast Gradient Sign Method (FGSM) and Basic Iterative Method (BIM) attacks, respectively, based on the CIFAR-10 dataset. In summary, the presented approach achieves a 70.66% reduction in computations for the student model, allowing the model to be deployed on devices with limited resources and provide improved prediction accuracy towards adversarial attacks.
In wireless sensor networks (WSNs), energy consumption is a critical concern, and replacing batteries at the node level is often costly and impractical. Network lifespan extension not only improves operational capacity but also minimizes maintenance overhead. As the network size increases, overall network management and scalability become challenging, especially when handling a large amount of sensor data. To address these issues, we proposed an approach known as Measurement Sum Master (MSM) clustering, which leverages particle swarm optimization to detect healthier inter-cluster routing directions. This approach introduces a master node strategy to select the average cluster size for maintaining scalability according to the dimensional factor. The proposed approach is designed to analyze such issues in clustering based on the concept of particle swarm methodology for identifying the best-suited inter-cluster routing, that is, master-based sum-measured optimization offers improvements over existing methods by maintaining the fitness of the cluster head in terms of energy and packet delivery. It is formulated with an algorithm to identify the best cluster size to handle the dimension. It is finally compared with the existing protocol in terms of energy consumption to analyze the performance of the protocols. The suggested approach introduces a new methodology of clustering for WSNs to improve efficiency and enhance operational longevity. The MSM-PSO protocol was used for the analysis, and the results were compared in terms of lifespan and scalability. The power efficiency increased by approximately 8–9% and the packet transmission cost was reduced by approximately 5% along with the improved fitness measurement. A comparative analysis with existing protocols was also presented to demonstrate that the MSM clustering approach significantly minimizes energy consumption, improves data delivery reliability, and increases network lifetime. This novel clustering approach offers an energy-efficient and scalable solution for enhancing the reliability of WSNs.
The precise identification and segmentation are indispensable for effective therapy planning and complete surgi cal management in medical imaging. However, conventional segmentation methods recurrently encounter chal lenges due to the intrinsic complexity of anatomical structures, variability in organ structure, and inconsistency across diverse imaging modalities. This study employs an optimized hybrid deep learning approach that inte grates ResNet50 with an attention-augmented UNet architecture to improve segmentation accuracy and organ localization in medical imaging. The ResNet50 model uses an encoder that focuses on deep feature extraction, whereas the UNet serves as the decoder and is enhanced with an attention mechanism. They are both integrated to enhance the model’s efficiency by capturing both the global framework and local spatial details through a hy brid structure and skip connections, which improve segmentation performance. The proposed model was trained on a large-scale multi-organ dataset of high-resolution MRI to ensure robustness. In addition, the dataset was augmented and regularized to stabilize the model. The major performance metrics, namely Intersection over Union (IoU) and Dice coefficient, indicate that the proposed scheme achieves a high segmentation precision of up to 98.41%. These outcomes indicate that the model is highly feasible for deployment in clinical workflows to improve multi-organ detection with precise performance.
In today’s technological era, the pervasive presence of technology has led to exponential growth in data generation. The tourism industry, a major contributor to this data flood, generates large volumes of data, including comments, photos, and location-sharing on social media. Online tourism agencies collect metadata, including hotel views, clicks, and visitor comments. This metadata enables these agencies to predict click estimates and cost-per-click (CPC) for hotels, aiding in the development of effective bid strategies. This study presents a model for estimating CPC using big data analytics, leveraging metadata from online tourism agency dashboards. The key findings show that the gradient-boosted tree algorithm outperforms the Random Forest algorithm in predicting CPC with greater accuracy. The proposed model improves bid strategies and offers a significant advantage by leveraging extensive, diverse data. This research contributes to the field by demonstrating how advanced machine learning techniques can optimize marketing strategies within the tourism industry.
The Internet of Things (IoT) is a communication paradigm and a network of interconnected heterogeneous devices. It generates large-scale, fast-changing, and differently formatted data, a phenomenon referred to as Big Data, which has become widely used with the rapid growth of IoT applications, especially since the end of the first decade of the 2000s. The Internet of Things (IoT) is a communication paradigm and a network of interconnected heterogeneous devices. It generates large-scale, fast-changing, and differently formatted data, which is referred to as Big Data and has become widely used with the rapid growth of IoT applications, especially since the end of the first decade of the 2000s. A critical characteristic of this data is its real-time nature, which—coupled with its volume, velocity, and variety—demands advanced real-time analytics to extract value. While existing reviews have explored the broader intersection of IoT and Big Data Analytics, this paper provides a systematic, PRISMA-guided review focusing specifically on real-time analytics in IoT devices. By restricting the scope to real-time streaming data, the study provides a detailed synthesis that complements existing broader surveys. Applying the PRISMA methodology, we selected and analyzed 33 relevant articles published between 2018 and 2024. Our analysis reveals that scalable architectures are crucial for real-time Big Data analytics and that integrating machine learning techniques is fundamental to enabling intelligent decision-making. The study also underscores the necessity of appropriate data processing tools. The primary challenge in this area is minimizing latency, followed by data heterogeneity, scalability, and resource constraints. Furthermore, the survey identifies the critical role of real-time analytics in areas such as healthcare and smart cities. It discusses integrating fog and edge computing into future architectures to address latency and resource constraints.
The rapid growth of the mobile gaming industry has intensified competition, prompting developers to adopt advanced methods to improve engagement and retention. This paper presents an AI-driven platform for personalized mobile game development that applies Machine Learning Operations (MLOps) principles to tailor game content based on individual user behavior. Built on the MLflow (an open-source platform for managing the ML lifecycle) framework, the system ensures scalability, reproducibility, and efficient model lifecycle management. It integrates algorithms such as XGBoost, Random Forest, K-means, and DBSCAN for tasks including churn prediction, ad optimization, and adaptive difficulty adjustment. Multi-dimensional user data—covering session metrics, gameplay patterns, demographics, and monetization events—is processed through automated backend services. MLflow enables experiment tracking, model versioning, and artifact management across all development stages. Empirical results show strong predictive performance, with Random Forest achieving an F1-score of 0.8679 in churn prediction. The proposed MLOps-centric framework advances data-driven game development and establishes a replicable model for AI-powered personalization.
Healthcare expenditures constitute a significant portion of national budgets in many countries, and striking a balance between cost-effectiveness and service quality is becoming increasingly important. In this context, financial analyses play a critical role in strategic decision-making processes for healthcare organizations. In particular, forecasting hospital turnover guides in many areas, such as budgeting, resource management, and personnel planning. This study aims to develop revenue forecasting models using the hospital’s historical turnover data to gain a better understanding of its financial status and to provide a more solid foundation for future financial decision-making in the healthcare sector. The present work reports a comparative time-series analysis of deep learning techniques (Recurrent Neural Networks with the Gated Recurrent Units (GRU), Long Short-Term Memory (LSTM) cells, and Neural Basis Expansion Analysis Time Series (N-Beats)) and statistical techniques (ARIMA and SARIMA) and the Naïve algorithm to forecast revenue using the hospital’s historical turnover data. The findings reveal that the SARIMA model provides high accuracy by successfully analyzing seasonal and trend components. In contrast, the N-BEATS model exhibits the best overall performance with the lowest error rate, highlighting its strong ability to capture complex temporal dependencies and outperforming both traditional recurrent architectures and conventional statistical methods. The models developed as a result of the study are evaluated as contributing to the financial planning processes of healthcare organizations and supporting strategic decision-making mechanisms.
This study introduces RetinaNeXt, a deep learning model based on the ConvNeXt-Base architecture for grading the severity of diabetic retinopathy (DR) into five classes. The goal is to achieve robust, interpretable, and cross-dataset generalization using a scalable convolutional backbone. RetinaNeXt integrates transfer learning from pre-trained ConvNeXt, class balancing via Focal Loss, and extensive data augmentation. It is trained on the EyePACS dataset and externally validated on APTOS and IDRiD datasets. Grad-CAM visualizations are used to interpret the model’s focus on pathological regions. RetinaNeXt achieved 95.26% accuracy on EyePACS, 94.85% on APTOS, and 90.29% on IDRiD, demonstrating strong cross-domain generalization. Class-wise analysis revealed consistent performance across all ICDR grades, with improved sensitivity in minority classes, attributed to Focal Loss. The proposed RetinaNeXt framework combines architectural efficiency, interpretability, and generalization, making it suitable for real-world DR screening applications. Future work will explore lightweight deployment and integration with federated learning for privacy-preserving detection of DR.
This research proposes a framework to assess the sustainability performance of businesses within the textile industry. By integrating KPI’s created in line with the most common sustainability standards in the textile sector, the three core pillars of sustainability, social, environmental, and economic elements, are incorporated into an evaluation model constructed using multi-criteria decision-making techniques. The impact of the sector on sustainability was reflected in the structure of the developed model, and standards compatible with the Global Reporting Initiative, which are globally recognized, have been addressed. In determining the KPI’s, the most commonly used standards in the sector were examined, and performance indicators covering these standards were derived from insights shared by sustainability specialists who are experts in their fields. The weights of the criteria for each dimension were calculated with the AHP technique, based on evaluations by sustainability and analytical professionals with domain expertise and industry-specific experience. These weighted indicators were assessed using TOPSIS and Grey Relational Analysis, with real data obtained from a medium-sized textile company operating in Izmir. The company’s sustainability performance scores were then calculated. The framework offers a systematic and straightforward approach, enabling decision-makers to identify areas for improvement and enhance the effectiveness of sustainability initiatives.
