Machine learning (ML) in big data frameworks plays a critical role in real-time analytics, decision making, and predictive modeling. Among the most prominent ML libraries for large-scale data processing are Flink-ML, the machine learning extension of Apache Flink, and MLlib, the machine learning library of Apache Spark. This paper provides a comparative analysis of these two frameworks, evaluating their performance, scalability, streaming capabilities, iterative computation efficiency, and ease of integration with external deep learning frameworks. Flink-ML is designed for real-time, event-driven ML applications and provides native support for streaming-based model training and inference. In contrast, Spark MLlib is optimized for batch processing and micro-batch streaming, making it more suitable for traditional machine learning workflows. Experimental results show that training time is nearly identical for both frameworks, with Spark MLlib requiring 4006.4 seconds and Flink-ML 4003.2 seconds, demonstrating comparable efficiency in batch training and streaming-based model updates. Accuracy results show that Flink-ML (74.9%) slightly outperforms Spark MLlib (74.7%), suggesting that continuous learning in Flink-ML may contribute to better generalization. Inference throughput is slightly higher for Spark MLlib (8.4 images/sec) compared to Flink-ML (8.2 images/sec), suggesting that Spark's batch execution provides a slight advantage in processing efficiency. Both frameworks consume the same amount of memory (30.2%), confirming that TensorFlow's deep learning operations dominate resource consumption rather than architectural differences between Spark and Flink. These results highlight the tradeoffs between Flink-ML and Spark MLlib, and guide data scientists and engineers in selecting the appropriate framework based on specific ML workflow requirements and scalability considerations.

• APA 7th style

• Chicago style

• IEEE style

• Vancouver style

This paper reports the experimental results of a study investigating a new eco-composite material made from 100% recycled material. Tensile and density tests were conducted. A numerical model of a one-sided fixed beam was designed by the finite element method and a buckling analysis of this structure was performed. Three different cross-sections and lengths of the beam were tested. The first fundamental buckling mode and the corresponding critical load value were determined. The obtained numerical results were verified by analytical method using Euler's formula, which showed high agreement between the results. The relative error was less than 4%. A higher level of agreement was obtained for longer beams than for shorter ones. The results obtained for the eco-composite were then compared with those reported for other materials with similar properties, namely LDPE, HDPE and PP. Compared to LDPE and HDPE, the eco-composite showed higher stiffness parameters and load resistance, which made the tested structure more rigid and therefore stable for a longer period of time. The analysis of beams with different cross-sections and lengths made it possible to determine the effect of these parameters on the critical load, providing valuable insights for designers. It was observed that a 100% increase in the initial rectangular cross-section of 800mm2 resulted in a 685% increase in the stiffness of the beam. A 100% increase in the initial beam length of 150mm resulted in a 75% decrease in the critical force. The results of this study have confirmed that the new eco-composite material can be effectively used in engineering structures.

• APA 7th style

• Chicago style

• IEEE style

• Vancouver style

Parkinson's disease (PD), a type of neurodegenerative disease, is on the rise globally as the population ages. Today's costly diagnostic techniques for Parkinson's disease often detect the illness after significant brain damage has already occurred. Early detection is essential for improving patient outcomes and potentially slowing the disease's progression. One of the newest advances in artificial intelligence, deep learning (DL), presents new opportunities for the early, non-invasive diagnosis of Parkinson's disease. Fundus imaging, which captures fine-grained images of the retina, is a promising technique for detecting the disease's early symptoms. Changes in the retinal blood vessels and anomalies of the optic disc (OD) have been linked to neurodegeneration. DL models can identify subtle patterns in these fundus images, such as vascular alterations and changes in the optic disc, which have been connected to Parkinson's disease. This approach replaces current diagnostic methods with a scalable and cost-effective solution, increasing access to early detection. This review explores the current state of the art in using DL models with fundus images to detect PD early on, with a focus on significant public datasets, methodologies, and related research. It highlights how DL models could transform PD screening and provides an overview of the advancements and challenges in this emerging field.

• APA 7th style

• Chicago style

• IEEE style

• Vancouver style

The article's objective is to conduct theoretical research of modern digital solutions for risk management in business ecosystems and develop an intelligent digital tool for risk management under sustainable development conditions. A comprehensive analysis of modern technologies that include artificial intelligence, big data, and blockchain, is conducted, their role in improving risk management efficiency is determined. The research methodology combines both theoretical methods (systems analysis, comparative analysis, SWOT analysis) and empirical methods (statistical analysis, machine learning methods, and experimental research). The main categories of risks are systematized, and possibilities for their optimization through digital solutions are explored. The impact of digital technologies on achieving sustainable development goals is analyzed, particularly in aspects of efficient resource usage, social integration, and innovation development. Key challenges of digital transformation in risk management are identified, including cybersecurity issues and regulatory requirements compliance. The practical application of machine learning methods for predicting employee attrition is examined, demonstrating the potential of digital solutions in solving specific business challenges. A prediction system that uses various machine learning algorithms was developed and tested. A comparative analysis of the effectiveness of various machine learning algorithms for the prediction task was conducted. When selecting the optimal classifier, both standard quality metrics and probability distribution analysis for identifying risk groups were taken into account. A modular system structure is proposed, and practical recommendations for implementing digital solutions in business ecosystems are provided.

• APA 7th style

• Chicago style

• IEEE style

• Vancouver style

This study aims to evaluate and compare five algorithms in diabetes detection, namely Flower Pollination Neural Network (FPNN), Particle Swarm Optimization Neural Network (PSONN), Bat Artificial Neural Network (BANN), Stochastic Gradient Descent (SGD), and Quadratic Interpolation Flower Pollination Neural Network (QIFPNN). These algorithms were tested on a diabetes risk dataset divided into training, validation, and testing subsets. The evaluation was based on three main aspects: accuracy, F1 score, and training time. Experimental results showed that QIFPNN outperformed others with an average accuracy of 97.90% and an F1 score of 98.30%, although it required the longest training time (4107.89 seconds). FPNN and BANN achieved competitive accuracy (97.34% and 97.43%) and F1 scores (97.84% and 97.91%), while SGD offered a favorable trade-off with accuracy of 96.87%, F1 score of 97.42%, and the shortest training time (584.50 seconds). PSONN performed less well with an average accuracy of 89.26% and an F1 score of 91.45%. These results indicate that QIFPNN can be relied upon as an effective diabetes risk detection model with superior predictive performance. Although the training time of QIFPNN is longer due to its sophisticated optimization process, this is only a concern during model development, as the final trained model can be efficiently used for real-time prediction in practical applications.

• APA 7th style

• Chicago style

• IEEE style

• Vancouver style

The rapid development of digital technologies has transformed healthcare systems around the world, and telemedicine has become the primary solution to problems related to the availability and quality of medical care. This study examines the adoption of telemedicine in five Central Asian countries - Kazakhstan, Kyrgyzstan, Uzbekistan, Tajikistan, and Turkmenistan - by modeling the relationship between key medical, demographic, and technological factors and the number of telemedicine users. To identify the factors that contribute to telemedicine adoption, a dataset of epidemiological, demographic, and digital infrastructure indicators was analyzed. For the analysis, data from the National Statistical Office of the Republic of Kazakhstan (2014-2024) were used. To predict the number of telemedicine users, an artificial neural network (ANN) was used, which has a shallow network structure with four input neurons representing the main predictors and one output neuron for potential telemedicine users. The predictive model showed excellent accuracy, as evidenced by a very strong correlation between predicted and observed values (R = 0.99245). In addition, the reliability of the model is confirmed by its low error rates, with a mean squared error (MSE) of 0.007 and a root mean squared error (RMSE) of 0.0839. These findings underscore the transformative potential of telemedicine to address health challenges in Central Asia, while providing valuable insights into the epidemiological, demographic, and technological drivers that can guide targeted policy initiatives and strategic investments in digital infrastructure.

• APA 7th style

• Chicago style

• IEEE style

• Vancouver style

The Internet of Things (IoT) is a new technology that builds on the old Internet. A network connects all objects using technologies such as Radio Frequency Identification (RFID), sensors, GPS, or Machine-to-Machine (M2M) communication. The development of IoT has been negatively impacted by security concerns, which has led to a significant increase in research interest. However, very few methods look at the security of IoT from the attacker's point of view. As of today, penetration testing is a common way to check the security of traditional internet or systems. It usually takes a lot of time and money. In this paper, we look at the security problems of the Internet of Things (IoT) and suggest a way to test for them. This way is based on a combination of the belief-desire intention (BDI) model and machine learning. The results of the experiments showed that they were very good at detecting and defending against cyberattacks on IoT devices. The proposed BDI-based recall method provided 85% of the results. The 90% precision suggests that the measurements are very accurate. The F1-score was 87.4%, and the accuracy was 95%. The proposed BDI is of exceptional quality in every part of the penetration-testing model.  Therefore, it is possible to create a system that can detect and defend against cyberattacks based on the proposed BDI model.

• APA 7th style

• Chicago style

• IEEE style

• Vancouver style

By gaining new insights into the gene expression of individual patient profiles, clinicians and researchers can identify patterns, biomarkers and therapies. In addition, accurate classification enables the development of predictive models for prognosis and treatment response, facilitating personalized medicine approaches. Determining the optimal model for classification remains a time-consuming, nondeterministic, polynomial-time hard problem. However, the available large amount of gene expression data is too much for the traditional data analysis approaches. Therefore, a two-phase ensemble deep learning approach can be considered as a reliable framework for the root-level investigation of genomic data. In this experimental model, a gene extraction approach, a Kernel-Applied Fisher Score (KFScore) method is presented to select the prominent genomes, and a Sine-Cosine Ensemble Monarch Butterfly algorithm (SC-MBO) optimized CNN (Convolutional Neural Network) strategy is implemented for genomic data classification. Here, the SC-MBO ensemble approach is used to obtain the optimal value of hyperparameters in CNN. The effectiveness of the presented model is estimated by accuracy% of classification, number of extracted prominent genomic features, sensitivity, specificity and ROC (Receiver Operating Characteristic) curve. The effectiveness of the proposed methods is successfully tested on GSE13159, GSE15061, GSE13204, breast cancer and ovarian cancer gene expression dataset with 91.6%, 90.22%, 91.9%, 97.93% and 99.6% accuracy. The proposed model is also compared with other existing models. According to the experimental evaluation, the proposed strategy is accurate, reliable and robust. Consequently, the presented method can be treated as a trustworthy basis for disease risk prediction.

• APA 7th style

• Chicago style

• IEEE style

• Vancouver style

This paper investigates the effectiveness of selected tools for sentiment analysis, focusing on both dedicated software libraries (NLTK, Pattern, TextBlob) and large language models (ChatGPT and Gemini). The evaluation was conducted in two stages: sentiment analysis of 30 synthetic opinions of varying linguistic complexity, and analysis of 5 sets of real user reviews collected from the web. The results show that large language models - although not explicitly designed for sentiment analysis - achieved the highest accuracy, with ChatGPT consistently producing the lowest deviation from human ratings. In contrast, software libraries showed greater variation, especially in the presence of complex linguistic structures. These findings highlight the potential of large language models in sentiment analysis tasks and underscore their robustness in interpreting nuanced language.

• APA 7th style

• Chicago style

• IEEE style

• Vancouver style

Deep fakes pose a significant threat to medical imaging. These deep fakes appear very similar to real diagnostic scans and are often difficult to distinguish from real medical images. This paper discusses how deepfakes are created and highlights their potential for research and education, as well as risks such as misdiagnosis and data manipulation. We also review various deepfake detection techniques, ranging from traditional image forensics to advanced deep learning models, and highlight the strengths and weaknesses of these approaches for detecting sophisticated deepfakes. We also discuss the ethical issues of deepfakes in healthcare, such as patient privacy, data security, informed consent, algorithmic bias, and the potential loss of trust in medical systems. In addition, we present an experimental study that evaluates how well different deep learning models detect deepfakes in a lung CT scan dataset, demonstrating both the potential and limitations of current detection methods. Finally, we outline future research directions, including real-time detection, explicable AI, enhanced cybersecurity, and strengthened ethical guidelines. This review is a valuable resource for researchers, clinicians, and policymakers interested in exploring AI medical imaging and ethics in the age of deepfakes.

• APA 7th style

• Chicago style

• IEEE style

• Vancouver style

In the rapidly evolving retail industry, data-driven decision making is critical to maintaining competitive advantage and operational efficiency. This paper explores the diverse applications of Microsoft Power BI (MPBI) in retail, highlighting its impact on real-time data management, sales analysis, inventory optimization, customer insights, and supply chain performance. By synthesizing findings from recent studies and presenting empirical data from case studies, we demonstrate how Power BI's advanced analytics and visualization capabilities can transform raw data into actionable insights. Our research underscores the importance of integrating disparate data sources into a unified platform, facilitating comprehensive data analysis, and fostering a culture of data literacy across retail organizations. We also discuss the challenges and best practices for implementing Power BI across retail functions, highlighting its role in driving innovation and adapting to emerging market trends. The results of this study provide practical insights for retailers seeking to leverage data analytics for strategic decision-making and operational excellence.

• APA 7th style

• Chicago style

• IEEE style

• Vancouver style