Welcome to SDU Institutional Repository
Communities
Select a community to browse its collections.
Recent Submissions
Using Visualization Methods in Teaching 6th Grade Students to Solve Word Problems
(SDU University, 2026) Kyrgyzbay A.
This study focuses at the efficiency of employing visualization tools to solve mathematical word problems for sixth graders. The study topic is focused on students' difficulty converting content of word problem into mathematical expressions (modeling) and their tendency to depend solely on "keywords" in the content. To overcome this issue, an one-group pretest and posttest quantitative research methodology was implemented. The study included 30 sixth-grade students recruited by convenience sampling. A pretest to identify the beginning level was used to gather data, followed by a one-month pedagogical intervention to teach visualization tools such as modeling and schematic diagrams, and finally a posttest. To assess the research findings, all 300 replies (10 questions⋅30 students) were examined using a specialized structural modeling descriptor ranging from 0 to 3. The t-test for dependent samples revealed a substantial increase in scores between the pretest (M=14.37) and posttest (M=15.87) (𝑝 < 0,001). The calculated Cohen’s d effect size of 0.311 further confirms a small-to-medium practical impact of the visualization intervention on the students' modeling and word problem-solving abilities. Quantitative study revealed that most of students effectively advanced from a low level of competence to graph word problems (0-1 points) to a level of proper structural modeling (2 points). Finally, this study statistically verified that systematic training of visualization tools minimizes cognitive burden while considerably improving 6th grade students' ability to properly model and solve mathematical word problems.
Impact of Smart Technologies on Teaching Sustainable Chemistry: A Focus on Biohumus Production
(SDU University, 2026) Abdybeluva A.
This paper investigates how smart Internet of Things (IoT) sensors might help in teaching sustainable chemistry through a biohumus production project. The research attempts to answer the following questions: how do smart technologies increase the effectiveness of sustainable chemistry teaching; to what extent do digital tools enhance students' understanding of biohumus production; and which technologies are the most effective in enhancing green chemistry knowledge and skills? The study employed a mixed-methods design. The quantitative phase involved a quasi-experimental pre-test/post-test design with 72 third-year undergraduate chemistry students at SDU University, Kaskelen, Kazakhstan (n=36 experimental; n=36 control). The experiment group applied an IoT intelligent soil detector during lab sessions; the control group applied conventional measuring tools. The qualitative phase consisted of an online survey of 15 practicing chemistry teachers, analysed using thematic analysis. Results illustrated that the learning gains of the experimental group were significantly higher than that of the control group, U = 306, p < .001, r = .529. Posttest accuracy was 81.9% for the experimental group and 60.8% for the control group. Qualitative findings revealed four themes: real-time data visualization as IoT's core teaching value; natural complementarity between biohumus and smart technology; structural barriers to adoption; and a recommendation to start small. This Central Asian study offers empirical evidence from a context underrepresented in international literature and provides a replicable instructional model for technology-enhanced sustainable chemistry teaching. A preliminary version was presented at ICEDU 2026 in Bali, Indonesia.
Optimizing Task Assignment in Project Management Using Recommendation Algorithms and Improving Team Efficiency
(SDU University, 2026) Turysbay D.
Allocating software development tasks to the right team member is one of the most consequential decisions a project manager makes daily. This thesis presents the design, implementation, and empirical evaluation of a full-stack project management web application augmented by a machine learning recommendation engine that ranks developers for a given task. The problem is modelled as learning-to-rank: given a task description (skills, position, grade, deadline, effort, priority) and the team’s current state (skills, workload, performance history), the system produces a ranked list of candidate developers with interpretable explanations. Four strategies are compared under a single protocol on a held-out test set: (i) a transparent heuristic baseline combining seven priority-weighted component scores; (ii) an embedding approach using sentence-transformer cosine similarity; (iii) a LightGBM LambdaRank learning-to-rank model over thirty engineered features; and (iv) an LLM-augmented pipeline using a local Mistral model via Ollama for skill extraction and fit scoring. Seven additional classical algorithms provide a wider comparison. The system is implemented as a containerised microarchitecture: a FastAPI/PostgreSQL backend, a React (Vite) frontend with roleaware navigation, a sidecar Ollama service, and a GitLab CI pipeline deploying to Railway. Evaluation uses a synthetic dataset of 1,000 (task, developer) pairs over one hundred tasks and twenty developers, with a task-grouped train–test split, scored by NDCG@K, MAP, MRR and Precision@1. The tuned baseline is strong (NDCG@5= 0.847, MAP= 0.794); embeddings give a small consistent gain (MAP= 0.799); LambdaRank slightly underperforms on synthetic data (MAP= 0.717); and the LLM pipeline matches embeddings while adding feature-enrichment and explanation value.
Student Performance Prediction Using Machine Learning
(SDU University, 2026) Syzdykov D.
The early identification of academically at-risk students remains one of the most pressing problems in contemporary higher education. As universities accumulate large amounts of administrative, demographic and behavioural data through their Virtual Learning Environments, machine learning has emerged as a powerful instrument for transforming raw educational data into actionable early-warning signals. This dissertation develops, evaluates and explains a comprehensive machine learning pipeline for predicting student academic performance on the Open University Learning Analytics Dataset (OULAD), a publicly available benchmark containing detailed records for 32,593 learners across 22 module-presentations. The work is framed around four research questions concerning the most predictive student features, the most suitable learning algorithm for the task, the relative contribution of individual pipeline components and the practical interpretability of the resulting models. To answer these questions, the dissertation proposes an end-to-end methodology that combines multi-source data integration, domain-driven feature engineering, hybrid feature selection, synthetic minority over-sampling, Bayesian hyperparameter optimisation using Optuna and post-hoc model interpretation through SHapley Additive exPlanations (SHAP). Four representative classifiers are compared on the same task: a regularised Logistic Regression baseline, Random Forest, XGBoost and LightGBм. Experimental results demonstrate that gradient-boosted tree ensembles, and in particular an Optuna-tuned XGBoost model, consistently outperform the alternative baselines on the held-out test set, reaching a test area under the receiver operating characteristic curve of 0.97 and a test F1 score of 0.90 with a moderate overfitting gap. Engagement-related features-including the number of active days on the Virtual Learning Environment, total clicks and an engineered score_consistency indicator-are identified as the most influential predictors by both global feature-importance methods and SHAP analysis. An ablation study reveals that feature engineering, class-imbalance handling and hyperparameter tuning each contribute incrementally to the final performance and that their combination yields the best balance of accuracy, generalisation and computational efficiency. The dissertation also discusses educational, ethical and deployment-related implications of the proposed pipeline. SHAP-based explanations are shown to produce individualised, actionable insights that can be conveyed to educators and policymakers without requiring a deep understanding of the underlying algorithms. Limitations associated with the use of a single benchmark dataset, with binary outcome modelling and with the absence of causal claims are explicitly acknowledged, and directions for future research are outlined, including temporal modelling, fairness-aware learning and integration with existing learning analytics dashboards. Overall, the dissertation contributes a reproducible methodological template that bridges predictive accuracy and humaninterpretable insight in the context of student performance prediction.
Coordination and Communication Protocols for Multi-Agent Financial AI Systems in Automated Trading
(SDU University, 2026) Sabidolda Sh.
Machine-learning models now dominate daily-frequency financial forecasting but remain di!cult to audit, and the recent turn toward multi-agent architectures has not resolved the underlying problem: existing LLM-coordinated frameworks couple specialised agents through stochastic natural-language inference, provide no formal guarantees of inter-agent message integrity, and retrain on single-observation performance dips that conflate transient noise with persistent drift. This dissertation proposes the Deterministic Multi-Agent Coordination and Control Protocol (DMACCP), a five-agent pipeline in which every message is carried by a typed envelope with a SHA-256 integrity checksum on a file-backed bus; an orchestrator evaluates five validation gates between stages, admits retraining only when a degradation signal persists across K = 2 consecutive observations, and uses SHAP feature attributions as a cross-run control signal that feeds back into adaptive feature masking. The protocol is evaluated on eight assets spanning five market segments over 2019–2024, against a no-coordination baseline using identical data and models. Walk-forward accuracy lies within noise bounds of the baseline (mean ” = →0.22 percentage points, overlapping 95% bootstrap intervals); the bu#ered mechanism reduces unnecessary retraining events by 66.7% on the ablation configuration; a macroeconomic context block dominates the SHAP attributions of seven of eight assets, validating the protected feature set. The contribution is an engineering proposal with formal reproducibility and integrity properties—obtained at under one percent of the pipeline’s runtime—rather than a performance-extension result, and is positioned as a designed answer to the audit and reproducibility requirements of institutional financial machine-learning deployment.