SDU University Bulletin: Natural and Technical Sciences

Permanent URI for this community

https://repository.sdu.edu.kz/handle/123456789/1643

Browse

Now showing 1 - 8 of 8

Open Access
A PROFESSION RECOMMENDER SYSTEM BASED ON DEEP LEARNING AND MACHINE LEARNING APPROACHES
(СДУ хабаршысы - 2023, 2023) Iskalinov F.
Abstract. The issue of uncertain career path choice among modern schoolchildren has become increasingly prominent, resulting in a substantial decrease in the number of university students. This uncertainty has become a major concern as students and their parents are often unfamiliar with the wide range of available professions, particularly those that have emerged in the last decade. A modern solution is proposed in the form of a web application that uses Deep Learning, Machine Learning, and NLP to recommend suitable specialties based on the competencies required for the profession. The system will analyze and extract implicit features through a supervised classification approach, providing a comprehensive solution for profession search in the Kazakhstan market.
Open Access
A Survey on Multimodal Approaches for Lung Disease Diagnosis using Deep Learning
(SDU University, 2025) Zhaniya Medeuova
Lung disorders are a major global health issue. A quick and accurate diagnosis is essential for proper treatment. In order to increase diagnostic accuracy, recent multimodal techniques are gaining popularity. This study carried out a comprehensive analysis of research articles on multimodal approaches that were published between 2020 and 2024 in Scopus and Google Scholar. The results show that there is limited study on the multimodal approach and on a variety of lung disorders such as asthma, TB, pneumonia, and chronic obstructive pulmonary disease. Several studies concentrated mainly on the detection and binary classification of COVID-19. The field has several challenges, including limited datasets, high computing costs, difficulties in integrating multiple modalities, and lack of accessibility of the models. Future studies should look at a wider range of lung diseases, increase the accessibility of datasets, improve fusion methods for merging data from many sources, and create models that are easier to understand and use fewer resources. Resolving these issues will improve patient outcomes by advancing the real-world use of deep learning in medical diagnosis.
Open Access
DETECTING SOCIAL CONFLICTS IN KINDERGARTENS USING DEEPLEARNING AND COMPUTER VISION
(SDU University, 2025) Dina Kengesbay
Early conflict detection in kindergartens plays a significant role in ensuring a harmonious learningatmosphere and in promoting the social growth of young children. While most previous works have onlyaddressed conflict detection through adults, in this paper, we specifically address conflict detection inkindergartens using deep learning, utilizing both spatial and temporal information to improve performance.The application of deep learning and computer vision in automatically detecting and analyzing earlyconflicts among young children is discussed in this paper. Using video footage, we leverage state-of-the-art RNNs and 3D CNNs for high-accuracy detection of conflict instances. Crucial visual cues—facialexpressions, gestures, poses, vocal tone, and movement—are examined for the extraction of tension oraggression signs. The model is evaluated on real kindergarten video data, with promising conflict detectionand classification results. The findings indicate the potential of AI-supported tools in assisting teachers inclass management, child behavior monitoring, early intervention mechanisms, and the fostering of a goodsocial environment.
Open Access
Development of method to analyzefactors of kidney disease by the use of fuzzy logic
(SDU University, 2025) Assel Yembergenova; Azamat Serek; Bauyrzhan Berlikozha
The study introduces a new strategy for the analysis of kidney disease parameters based on fuzzy logic.Fuzzy logic is a more accurate way to categorize clinical parameters than statistical analysis because thereis uncertainty and variability in medical data. The data is comprised of an extensive amount of clinicalparameters including age, blood pressure, specific gravity, albumin, sugar, random blood glucose, bloodurea, serum creatinine, sodium, potassium, hemoglobin, packed cell volume, white blood cell count, andred blood cell count.The methodology utilizes fuzzy logic centroid computation to categorize these parameters into low,medium, and high levels to provide a more dynamic and interpretable assessment of renal health. Fuzzymemberships give the current work the capability to discover intricate interrelationships between clinicalvariables, which may have been otherwise unattainable by conventional mean, median, and standarddeviation-based analyses.The findings confirm that fuzzy logic and conventional statistical methods enhance the comprehensionof kidney disease by incorporating intricate interactions between clinical variables. The method is employedto achieve more accurate prediction and diagnostic models, offering insight to be used in kidney diseaseassessment and medical decisions.
Open Access
Bias and Fairness in Automated Loan Approvals: A Systematic Review of Machine Learning Approaches
(SDU University, 2025) Suraiyo Raziyeva; Meraryslan Meraliyev
Artificial intelligence (AI) is increasingly transforming credit approval processes, enabling financial institutions to assess risk more efficiently and at greater scale. As these systems become more embedded in lending decisions, concerns around fairness, bias, and accountability have grown significantly. Many of these concerns stem from the use of historical data, proxy variables, and model optimization choices that can unintentionally reinforce existing social and economic inequalities. This work presents a systematic overview of the types and sources of bias in AI - driven loan approval systems and critically examines how machine learning techniques attempt to address them. It also highlights emerging solutions, including explainable AI, federated learning, human-in-the-loop frameworks, and intersectional fairness approaches. Despite ongoing advancements, unresolved challenges remain - particularly the need for dynamic fairness monitoring and for addressing intersectional biases affecting individuals from multiple marginalized groups. To bridge these gaps, the paper emphasizes the importance of interdisciplinary collaboration among AI developers, regulatory bodies, and social scientists. It advocates embedding fairness as a core design principle in the development and deployment of future AI systems. Overall, this study contributes to the growing effort to develop more transparent, inclusive, and socially responsible financial technologies.
Open Access
Forecasting Student Academic Performance Using Machine Learning
(SDU University, 2025) Danial Polat; Azamat Serek
Educational data mining depends on accurate student academic outcome forecasting to detect studentswho need help early and receive specific support. Traditional linear models have been used extensively yetthey fail to detect the intricate non-linear patterns which exist in student achievement data. The evaluationof machine learning algorithms and their features for student outcome prediction in Portuguese secondaryeducation remains insufficient because of missing systematic assessments. The research investigates howLinear Regression and Random Forest and K-Nearest Neighbors perform when predicting Portugueselanguage grades from 649 student records containing 30 demographic and social and academic attributes.The evaluation of model performance used three established metrics which included Mean Squared Error(MSE) and R-Squared (R²) and Mean Absolute Error (MAE). The results showed Linear Regressionproduced the most accurate predictions through its lowest MSE (9.00) and MAE (2.30) values but its weakR² value (0.01) indicated poor explanatory power. The error rates of Random Forest matched those of LinearRegression (MSE = 9.48 and MAE = 2.34) yet its negative R² (-0.04) indicated poor generalization becauseof irrelevant features and suboptimal hyperparameters. The KNN model showed the worst results (MSE =11.10 and MAE = 2.57 and R² = -0.21) because it failed to detect important patterns without additionaloptimization. The results show that educational prediction tasks require both optimal feature selectionand parameter adjustment for successful results. The research shows that linear models perform betterthan complex methods in specific situations yet optimized non-linear models demonstrate superior abilityto understand student achievement complexity. The research provides essential guidelines for developingbetter feature engineering and machine learning approaches to predict educational results
Open Access
Development of sensor systems for floodwater monitoring and alerting
(SDU University, 2025) Adilbek Sarsenov; Lyazzat Ilipbayeva; Ulzhalgas Seidaliyeva
This study addresses the systematic prediction of river water levels in Kazakhstan via hy-drological computations, which are essential for forecasting water-related events and formulatingplans for sustainable water resource management. Particular focus is placed on the significance ofprompt and efficient monitoring of river dynamics to alleviate natural disasters such as floods andmudflows, especially in high-risk places like Almaty, situated in geologically unstable mountainouslandscapes. The research focuses the potential of intelligent sensor-based monitoring systems thatcan gather real-time data on water levels, precipitation, soil moisture, and various environmentalconditions. Systems integrated with artificial intelligence and data analysis can substantiallyaugment decision-making processes, facilitate early warning mechanisms, and boost the precisionof forecasts. This method ultimately protects natural ecosystems and local communities from thedetrimental effects of hydrological hazards
Open Access
Modeling and Forecasting DigitalCurrency Volatility with GARCH(1,1)
(SDU University, 2025) Bizhigit Sagidolla; Maral Zholaman; Meruert Bilyalova; Ayagoz Sagidolla
The burgeoning field of digital currencies presents unique challenges for predictive modeling due totheir inherent volatility and market dynamics distinct from traditional financial assets.We study the use of the GARCH(1,1) model to characterize and forecast the conditional volatility ofdaily Bitcoin returns. Using standard OHLCV data, we estimate a parsimonious GARCH(1,1) specificationand produce one-step-ahead volatility forecasts. We discuss model assumptions, stability conditions, andpractical considerations for risk metrics (e.g., VaR). The aim is to document a transparent, reproduciblepipeline rather than to compare exhaustively against alternative models. Results illustrate how a standardGARCH(1,1) specification can provide interpretable volatility estimates for Bitcoin, serving as a transparentbaseline rather than a novel predictive breakthrough.