Journal of Problems in Computer Science and Information Technologies

DEVELOPMENT OF A GENETIC ALGORITHM FOR OPTIMIZING CONVOLUTIONAL NEURAL NETWORKS IN ORDER TO IMPROVE THE ACCURACY OF OBJECT DETECTION IN DIFFICULT LIGHTING AND BACKGROUND CONDITIONS

Mukhtar Zhassuzak — 2025-06-27

This article addresses the challenge of improving object detection accuracy in video data captured under low-light conditions. Modern video detection systems—particularly in areas such as security, autonomous systems, and medicine—often suffer from reduced accuracy due to poor lighting. The proposed method is based on the integration of the YOLOv5 object detection model with a variety of image processing filters (including CLAHE, gamma correction, histogram equalization, Gaussian blur, bilateral filtering, the Non-Local Means algorithm, Gray-World and Max-RGB balancing schemes, as well as Retinex and MSRCR methods) and genetic algorithms. This approach enhances both the reliability of detection and computational efficiency. Experimental evaluations demonstrate that the proposed system achieves significantly higher object detection accuracy in low-light data compared to traditional methods.

DIGITAL FOOTPRINTS: CLUSTERING BROWSER HISTORY FOR USER PROFILING USING MACHINE LEARNING

Marzhan Idrissova — 2025-06-27

This study explores the use of unsupervised machine learning techniques to analyze historical web activity, segment users, and detect anomalies for user profiling. By applying hierarchical clustering and Gaussian Mixture Models, we identified distinct browsing behaviors, categorizing users into four to five groups, including general browsing, social media engagement, high-bandwidth consumption, and automated system processes. For anomaly detection, One-Class SVM and Isolation Forest were employed to flag deviations from expected behavior. The results indicate that approximately 5% of sessions were classified as anomalous by SVM, while Isolation Forest highlighted outliers associated with extended session durations and potentially high-risk application usage. These findings underscore the effectiveness of machine learning in distinguishing user behavior through digital footprints while identifying potential security threats or atypical usage patterns. The study demonstrates that unsupervised learning can serve as a valuable tool for user profiling and behavioral analysis, with implications for cybersecurity, network monitoring, and online behavior modeling. Integrating clustering with anomaly detection provides a scalable approach for uncovering usage trends and deviations in web traffic. Future research should expand dataset coverage and incorporate adaptive models to enhance classification accuracy and responsiveness to evolving web behaviors

COMPARATIVE ANALYSIS OF CNN MODELS FOR DETECTING CARDIOVASCULAR DISEASES

Alfarabi Mazhit — 2025-06-27

In modern medical practice, cardiovascular diseases (CVDs) remain among the leading causes of mortality and morbidity worldwide. Electrocardiography (ECG) continues to be one of the essential non-invasive methods for diagnosing cardiac rhythm disturbances and other myocardial pathologies. However, traditional ECG analysis relies heavily on specialist interpretation, which can be subjective and prone to human error. With advances in machine learning and deep learning, there is now an opportunity to automate and standardize the diagnostic process. In this study, convolutional neural networks (CNNs)—specifically the ResNet50 and VGG16 architectures—are applied to classify ECG images. Hyperparameter tuning is conducted to optimize batch size and learning rate. In addition, a prototype web service is presented, implemented with React for the frontend and Django REST Framework for the backend, that allows real-time, automated ECG classification. This approach has the potential to ease the workload of cardiologists and enhance the objectivity of diagnosis in clinical environments.

KAZAKH TRADITIONAL FOOD IMAGE CLASSIFICATION USING CNNS

Iliyas Makhatbek — 2025-06-27

Obesity is becoming an increasingly serious global health issue with severe consequences. Effective nutrition management is crucial in combating this epidemic. In Kazakhstan, traditional foods are a central part of the culture, yet comprehensive data and tools for analyzing dietary habits are lacking. Leveraging advances in computer vision, we developed a convolutional neural network (CNN) based approach to automatically classify images of traditional Kazakh dishes. We compiled a new dataset of 9,577 images across 22 categories of Kazakh foods and used it to train and evaluate several CNN models. The best-performing model (a fine-tuned DenseNet121) achieved a top-1 classification accuracy of 95%. These results highlight the potential of AI-based food recognition for dietary monitoring, nutritional assessment, and cultural preservation. Furthermore, the trained model was deployed in a Telegram chatbot to enable real-time food identification through image uploads, demonstrating a practical application of the system.

SHORT-TERM BUS PASSENGER DEMAND FORECASTING USING MACHINE LEARNING: A CASE STUDY OF ROUTE №50 IN ASTANA

Aruzhan Amanova — 2025-06-27

This study introduced a machine learning-based approach for short-term forecasting of passenger traffic in the Astana city bus system, in particular, focusing on Route №50. Сonsidering how rapidly cities are developing and increasing transport problems, accurate forecasting of bus demand is an important step towards optimizing resource allocation and improving the quality of service and passenger satisfaction. The research combines data from several sources - information about passenger traffic from a transport company, 15-minute traffic figures and weather conditions, and offers a predictive model that was developed using CatBoostRegressor. The data was collected over one week in December 2024 and covered 9,819 passenger traffic records with a total of 22,111 boardings. According to the results of the study, the model showed high performance with RMSE values of 2,920 and 2,516 for directions from A to B and from B to A, respectively, accurately reflecting the structure of demand at different times and in different places. Also, an analysis of the importance of features showed that factors such as the location of the stop, time of day and traffic congestion are the most significant factors affecting bus demand. The results serve as the foundation for dynamic bus allocation and timetable optimization in difficult urban conditions characterized by harsh winter conditions and traffic congestion. This research addresses critical gaps in the literature by developing a resource-efficient forecasting solution adaptable to evolving urban environments with limited historical data sets. This study offers resource-efficient solutions for forecasting bus demand, adaptable to evolving urban environments with limited historical data sets like Astana, and addresses gaps in the literature.

NUMERICAL MODELING OF THE FILTRATION PROCESS IN DYNAMICALLY INTERCONNECTED MULTILAYER GAS FIELDS

Abdugani Nematov — 2025-06-27

This article is devoted to modeling of the gas filtration process in a dynamically interconnected multilayer porous medium. The article is devoted to modeling the process of gas filtration in a dynamically interconnected multilayer porous medium. In the article, the process of gas filtration in a heterogeneous three layer porous medium with low-permeability intermediate layers and the dynamic interaction between the layers are described by a mathematical model based on a system of differential equations of parabolic type. This mathematical model is numerically simulated using finite difference methods, i.e. explicit and implicit schemes. Since the resulting system of finite difference equations is nonlinear with respect to the pressure function, a quasilinear method was used. The dynamics of the pressure function over time was analyzed for time intervals of 360, 720 and 1080 days, and during this period the pressure distribution in the layers, the rate of pressure drop around the well and the dynamics of interlayer interaction were studied. The calculation results are presented in numerical and graphical form, which accurately reflect how the interlayer movement of the gas flow occurs. Using graphical analysis, the time step limit was determined, ensuring the stability of the computational process in the explicit scheme: a stable calculation is carried out only with a dimensionless time step Δt ≤ 1.7e-4. Also, the calculations carried out using the implicit scheme showed that this method has more stable stability compared to the explicit scheme. The results show that with a large permeability coefficient of the formation, the pressure distribution accelerates, and in the wells the pressure drop slows down. At the same time, in directly connected multilayer porous media, the permeability coefficient of the layers plays an important role. Based on the obtained results, it is possible to carry out calculations for various parameters to improve the efficiency of gas field development. It is also possible to analyze and forecast oil and gas deposits using software created on the basis of numerical models and algorithms developed in the article.