Journal of Problems in Computer Science and Information Technologies

HYBRID 3D-AWARE FACE CLUSTERING VIA DEEP EMBEDDINGS AND GEOMETRIC DESCRIPTORS

Leila Rzayeva — 2026-03-26

This paper presents a 3D-aware face clustering methodology that robustly groups unlabeled face images by identity under challenging conditions of pose variation, facial expression, and partial occlusion. The proposed approach integrates 2D deep embeddings with 3D geometric features extracted from reconstructed facial meshes, leveraging both photometric and structural information. Preprocessing includes grayscale normalization, landmark-based alignment, and contrast enhancement. 3D face models are generated using a 3D Morphable Model (3DMM) and optionally refined through neural rendering to improve shape fidelity. From these reconstructions, we extract interpretable 3D descriptors-PCA shape coefficients, geodesic distances, and curvature histograms - that complement embeddings from ArcFace and FaceNet. Clustering is performed using a two-stage hybrid algorithm: DBSCAN for outlier removal followed by K-Means++ with a fused distance metric combining cosine and Mahalanobis distances. Experimental results demonstrate that the proposed method significantly outperforms 2D-only and 3D-only baselines in terms of Silhouette Score, Adjusted Rand Index (ARI), and Purity. The findings confirm that fusing 2D and 3D modalities yields semantically consistent and pose-invariant identity clusters, establishing a strong foundation for face analysis in unconstrained environments.

RESNET EMBEDDING-BASED PIPELINE FOR TRANSPARENT DIAGNOSIS OF PULMONARY EMPHYSEMA ON LOW-DOSE CT

Talshyn Sarsembayeva — 2026-03-26

This study presents a methodology for the automated detection and quantification of pulmonary emphysema from low-dose chest computed tomography (CT) scans. As a morphological subtype of chronic obstructive pulmonary disease (COPD), emphysema can be accurately assessed on CT imaging. Our approach utilizes a pre-trained ResNet152 model to extract high-dimensional feature embeddings (2048 dimensions) from mid-lung patches. Patients were automatically categorized based on the percentage of low attenuation areas (LAA%) below –950 Hounsfield units (HU), a standard measure for emphysema severity. The extracted feature embeddings were subsequently analyzed using statistical methods and logistic regression to identify key discriminative and interpretable features. A logistic regression model, trained on the top 20 most salient features, achieved a high level of performance, with an Area Under the Curve (AUC) of 0.94 and an Average Precision (AP) of 0.87 on a balanced dataset of 90 subjects. Furthermore, the selected features exhibited a strong correlation with LAA%, demonstrating their utility for regression-based severity assessment.

The findings confirm the viability of using pre-trained deep embeddings for transparent and reproducible emphysema screening. This method avoids the need for extensive end-to-end model retraining, making it highly adaptable for integration into existing clinical CT analysis workflows.

Keywords: Emphysema, low-dose CT, deep learning, ResNet, feature embeddings, explainable AI, COPD.

ALGORITHMIC APPROACH TO OPTIMAL RESOURCE CONTROL IN AN OPEN ECONOMY MODEL

Kamshat Tussupova — 2026-03-26

The dynamic development of the global economy and the impact of external factors create challenges for ensuring sustainable growth, making effective resource management an urgent task. This paper addresses the problem of optimal resource management in an open economy model with constraints on labor distribution, investment, and foreign trade balances. To solve this problem, a numerical implementation algorithm based on a modified Lagrange multipliers method is proposed. The developed algorithm enables obtaining a numerical solution to the optimal control problem (OCP) and makes the model applicable for practical use. The conducted computational experiments confirmed the effectiveness of the approach: the constructed state and control trajectories demonstrate the achievement of a stable equilibrium state of the system while satisfying all imposed constraints. The practical value of the study lies in the fact that the algorithm provides the possibility of implementing numerical solutions to OCPs and using them for analyzing and forecasting economic development under resource limitations.

COMPARATIVE ANALYSIS OF PHYSICS-INFORMED AND CONVENTIONAL LSTM AND RNN MODELS FOR TEMPERATURE FORECASTING USING ERA5 REANALYSIS DATA

Zharasbek Baishemirov — 2026-03-26

Climate change is one of the most serious modern problems affecting the Earth's atmosphere, as it causes a range of harmful effects worldwide. Due to the uneven nature of climate data, forecasting climate change is a challenging task today. Many previous studies in climate and machine learning have used recurrent neural networks (RNNs) and long short-term memory (LSTM) models to predict climate trends. Although these models are effective at identifying long-term trends in data, they often fail to satisfy physical laws such as energy conservation, mass balance, and thermodynamic principles. In this research, the aim was to develop physics-informed RNN and physics-informed LSTM models and compare their effectiveness with standard RNN and LSTM models. The study utilized data on air temperature at 2 meters above the surface for the cities of Astana, Almaty, and Shymkent for model training, validation, and testing. According to the results, physics-informed models achieved the lowest root mean square errors in Almaty (3.52 °C) and Shymkent (3.80 °C). RNN and LSTM models performed better in Astana (RMSE = 5.44 and 5.47 °C), where seasonal changes are relatively abrupt. These findings demonstrate that PINNs give more physically consistent forecasts for moderate climates, while conventional recurrent models remain more effective for locales with highly variable conditions.

DEVELOPMENT OF THE RETRIEVAL-AUGMENTED GENERATION (RAG) SYSTEM FOR THE KAZAKH LANGUAGE USING HYBRID INFORMATION METHODS

Nurlykhan Kalzhanov — 2026-03-26

Abstract. This study presents the development and experimental evaluation of the Retrieval-Augmented Generation (RAG) system for the Kazakh language with an emphasis on comparative analysis of information retrieval methods. The main purpose of the work was to test the hypothesis of the superiority of a hybrid approach combining statistical (BM25) and semantic (vector search) methods over individual approaches. Based on the corpus of legal documents of the Republic of Kazakhstan, 270 experiments were conducted using three data extraction methods combined with six modern large language models (LLM). The results demonstrate that the hybrid method achieves the highest accuracy (82.2%), statistically significantly surpassing vector search by 3.3% (p < 0.01) and BM25 by 5.5% (p < 0.001). All three methods showed accuracy above 75%, which confirms the high efficiency of RAG systems for the Kazakh language. The analysis also revealed that hybrid search provides the greatest stability of results when working with different language models. This study makes a significant contribution to the development of RAG systems for languages with limited resources, offering an empirically based methodology to improve the accuracy and reliability of response generation.

THREE-DIMENSIONAL FRACTAL GEOMETRY MODELING AND DIGITAL HOLOGRAPHY BASED ON R-FUNCTIONS

Saida Tastanova — 2026-03-26

This paper is devoted to modern research in the field of digital modeling of complex-shaped geometric objects and the determination of their optical properties, which currently represent one of the most relevant challenges in contemporary science. In particular, the problem of realistic representation of three-dimensional objects with fractal geometry and their holographic reconstruction in a full 3D format is of significant scientific and practical importance for such fields as industry, medicine, engineering, architecture, materials science, virtual reality (VR), and digital art.

Fractal structures possess a number of unique properties, including self-similarity, unlimited detail, and high spatial complexity, which makes them an effective mathematical basis for modeling natural objects such as plants, vascular systems, bone tissues, crystalline structures, and surface reliefs. At the same time, the geometric representation of fractal forms using classical methods is challenging, and their mathematical modeling requires the application of high-precision and formally rigorous techniques.

At present, the mathematical description of fractal objects is often based on statistical, stochastic, or iterative algorithms. However, such approaches are generally characterized by insufficient analytical rigor and smoothness, blurred boundaries, and the lack of a holistic spatial description. In this regard, there arises a need to develop methods for modeling complex fractal forms based on strict analytical expressions, in particular using the R-functions apparatus.

An additional challenging task is the reconstruction of holographic images of the modeled fractal objects, which requires high-precision optical modeling. The application of holographic technologies based on the principles of interference and diffraction makes it possible to adequately reproduce the spatial and structural features of fractal objects in a digital environment.

HADOOP, MULTI-AGENT SYSTEMS AND MACHINE LEARNING: EXPLORING SCALABILITY, FAULT TOLERANCE AND WORKLOAD DISTRIBUTION BEHAVIORS

Bolatzhan Kumalakov — 2026-03-26

Paper reports experimental results comparing several machine learning techniques performance when distributing massively parallel computation to a set of interconnected machines. Computational resources are intentionally heterogenous to simulate real ad-hoc network environment and provide realistic setting test results. Namely Round Robin, Q-Learning and Least Loaded algorithms-based solutions are examined for their scalability, fault tolerance and workload distribution behaviors. The novelty of the paper is a set of empirical set of coefficients and bottlenecks for each implementation that is free of infrastructure specifics or error and exemption handling tools for future considerations by engineering professionals and scholars.

NONLINEAR DIMENSIONALITY REDUCTION FOR LOOKALIKE AUDIENCE DETECTION USING MANIFOLD LEARNING AND AUTOENCODER-BASED REPRESENTATIONS

Il’murat Tokhtakhunov — 2026-03-26

Identifying users with similar behavioral characteristics is a critical task in modern targeted advertising and customer analytics systems. High-dimensional tabular datasets describing user activity often contain complex nonlinear relationships that cannot be effectively captured by traditional linear dimensionality reduction techniques. This study investigates representation learning approaches for constructing scalable look-alike audience detection systems using large-scale telecommunications data. Classical dimensionality reduction techniques, including Principal Component Analysis (PCA) and t-distributed Stochastic Neighbor Embedding (t-SNE), are first analyzed as baseline methods for exploring the structure of high-dimensional data. While PCA performs linear projections that preserve global variance and t-SNE reveals local neighborhood structures through nonlinear embedding, these methods are primarily designed for visualization and exploratory analysis and do not provide scalable parametric mappings for new data samples. To address these limitations, a representation learning framework based on autoencoders is proposed for generating compact latent embeddings of users. The model is trained on a large-scale anonymized telecommunications dataset containing behavioral, demographic, device-related, and service usage attributes. Embeddings are learned for multiple feature entities and concatenated into a unified user representation that integrates heterogeneous behavioral information. User similarity is then computed using cosine similarity in the latent space, enabling efficient identification of look-alike audiences. The proposed system is evaluated using clustering metrics and multiple independent validation tasks with external target variables to ensure unbiased performance estimation. Experimental results demonstrate that autoencoder-based embeddings produce a more structured latent space and improve both similarity-based retrieval and downstream classification performance compared to classical dimensionality reduction techniques. The findings highlight the effectiveness of deep representation learning for high-dimensional tabular data in real-world recommendation and targeted advertising systems.

INTEGRATING MACHINE LEARNING WITH OPEN-SOURCE 5G SA TESTBEDS FOR PERFORMANCE ANALYSIS AND KPI TIME SERIES MODELING

Zhenis Otarbay — 2026-03-26

Open source 5G Standalone (SA) testbeds provide cost-effective environments for research and teaching, yet most existing implementations focus primarily on functional validation rather than leveraging machine learning for advanced network analytics. This study presents a comprehensive framework integrating SARIMAX, LSTM, and Transformer models with a fully operational 5G SA testbed combining Open5GS, srsRAN, MongoDB, and ZeroMQ-based RF emulation. The primary objective is to demonstrate predictive analytics capabilities for 5G network performance forecasting using real testbed-generated Key Performance Indicator (KPI) data. A comparative forecasting analysis was conducted using the three models trained on KPI datasets augmented through CTGAN synthetic data generation. Experimental validation confirmed reliable end-to-end 5G operation with synchronized configuration across PLMN, TAC, DNN, and security parameters. Under controlled single-UE, RF-free conditions, the testbed achieved ultra-low latency (1.34 ms RTT), near-gigabit throughput (847 Mbps downlink, 823 Mbps uplink), and rapid PDU session establishment (0.22 s). Performance profiling identified the User Plane Function (UPF) and database interactions as primary scaling bottlenecks. The machine learning evaluation revealed that while SARIMAX provides a reliable statistical baseline, neural network models achieve substantially higher forecasting accuracy for network KPIs. These results demonstrate the extensibility of open source 5G testbeds toward intelligent network management and predictive analytics applications.

CONCEPTUAL MODEL OF THE QADAM DIGITAL PLATFORM AS A UNIFIED DIGITAL ECOSYSTEM FOR SMES

Bauyrzhan Abilda — 2026-03-26

The digital transformation of small and medium-sized businesses (SMEs) is a key factor in sustainable economic growth, increased competitiveness, and the integration of national economies into global value chains. In the developing platform economy, digital platforms act not only as technological solutions but also as institutional coordination mechanisms, facilitating interactions between businesses, government, financial institutions, and consumers.

This article examines the conceptual model of the QADAM digital platform, focused on SMEs, as a unified digital ecosystem. The platform integrates a B2B portal, an AI assistant for businesses, a digital commercial platform, P2P interactions, and a set of business tools, creating an end-to-end chain of digital services—from business registration and support to management and commercial decision-making. The objective of this study is to develop and validate the conceptual model of the QADAM digital platform, including a business model, decision-making model, and monetization model, as well as to analyze its potential for SME development.