Scientific Journal of Astana IT University

COMPARISON OF DIFFERENT OBJECT DETECTION TECHNIQUES FOR CENTRAL ASIAN FOOD RECOGNITION

2026-06-01T12:17:36+05:00

Automated food detection is a major component in building reliable dietary monitoring systems, particularly in regions where cuisine-specific datasets are limited. Central Asian food recognition presents unique challenges due to strong intra-class variation, visually similar dishes, multi-item meal compositions, and long-tailed class distributions. To address this problem, we present a comprehensive benchmarking analysis of five modern object detection models across three groups: two-stage (Faster R-CNN), one-stage anchor-based (RetinaNet), one-stage anchor-free (YOLOv12, YOLO26), and transformer-based (RT-DETR). Experiments are conducted on a unified dataset constructed by merging the Global Gastronomic Culinary Dataset (GGCD) and the Food Portion Benchmark (FPB), resulting in 48,228 images across 258 food classes. The dataset includes controlled single-portion photographs and multi-item meals captured across various backgrounds and from different viewing angles. The merged dataset therefore covers both isolated food portions and more complex scenes containing several dishes. We compare the models in terms of detection accuracy, computational requirements, and inference speed, and examine AP by food class across all models. Overall, YOLOv12 achieves the highest accuracy, with an mAP50 of 75.80 and an mAP50-95 of 67.90, while using fewer parameters and less computation than the heavier baselines. RT-DETR reaches an mAP50 of 72.90 but requires more computation, whereas YOLO26 records the shortest single-image inference time on both GPU and CPU. Class-wise analysis further shows high performance on visually distinctive categories, with near-perfect AP for ten top-detected classes. These findings provide a practical baseline for Central Asian food detection and show that future progress depends on improving robustness for long-tail and visually ambiguous classes.

CLOUD-BASED SYSTEM FOR AGRICULTURAL LAND MONITORING USING REMOTE SENSING DATA IN NORTHERN KAZAKHSTAN

2026-03-30T13:47:31+05:00

Satellite remote sensing offers strong opportunities for monitoring agricultural lands, especially in climate-sensitive regions such as Northern Kazakhstan. However, most existing agricultural monitoring platforms rely mainly on optical vegetation indices (e.g., Normalized vegetation index (NDVI) or Enhanced Vegetation Index (EVI)) and provide limited information on soil moisture and surface temperature. As a result, early signs of water and heat stress at the field level are often missed.

This study presents a cloud-based agricultural monitoring system that differs from existing solutions by integrating optical, microwave, and thermal satellite data within a unified framework. The system combines Sentinel-2 vegetation indices, SMAP SM, and land surface temperature (LST) from the MODIS MOD11A2 product, processed using Sentinel Hub and Google Earth Engine. All datasets are harmonized into consistent 10-day composites and aggregated at the field scale.

The Thermo-Moisture Drought Index (TMDI) was calculated by combining LST and soil moisture (SM), enabling improved characterization of drought conditions compared to vegetation indices alone. Integrating TMDI and EVI highlights stress conditions that are not visible in optical data only, demonstrating a clear advantage over traditional NDVI-based platforms.

The developed system provides an operational, cloud-based solution for field-level monitoring, potentially early stress detection, and interactive analysis. Compared to existing systems, it offers improved environmental interpretation by combining multiple satellite sensors, making it a practical decision-support tool for rainfed agriculture in Northern Kazakhstan.

This proposed system architecture is adaptable to other agricultural regions with similar characteristics. Furthermore, the framework can be extended with machine learning algorithms to predict crop yields and automatically assess drought risk.

AI-DRIVEN ATMOSPHERIC EMISSION MONITORING SYSTEM AS A FRAMEWORK FOR SUSTAINABILITY-ORIENTED PROJECT MANAGEMENT

2026-06-25T23:24:20+05:00

This paper presents an artificial intelligence (AI)-driven atmospheric emission monitoring framework for sustainability-oriented project management that integrates predictive analytics, blockchain-based data verification, and digital-twin technologies. Reliable environmental monitoring and short-term forecasting are increasingly required to support regulatory compliance, environmental risk assessment, and evidence-based project decision-making. The objective of this study is to develop and evaluate an integrated framework combining real-time industrial monitoring, neural-network forecasting, and decision-support mechanisms.

The empirical forecasting component uses 39,803 synchronized time-stamped observations obtained from an operating industrial monitoring system in Kazakhstan with a 20-minute temporal resolution. The monitored variables include NO, NO₂, SO₂, CO, particulate matter, oxygen concentration, temperature, humidity, pressure, and gas-flow indicators. After quality control, synchronization, outlier removal, and Min-Max normalization, historical sequences of 72 consecutive observations (24 hours) were generated. A Long Short-Term Memory (LSTM) neural network with recurrent layers of 32 and 16 hidden units, a dropout rate of 0.2, and a dense output layer was trained for one-step-ahead forecasting corresponding to the next 20-minute observation. The dataset was divided chronologically into 64% training, 16% validation, and 20% testing subsets to preserve temporal dependencies.

The proposed model achieved an overall MSE of 0.87 and R² of 0.86, while reducing pollutant-specific RMSE by 33.1–46.8% compared with the ARIMA baseline. The monitoring platform was additionally piloted at Promanalit LLP, where telemetry acquisition, preprocessing, visualization, reporting, and blockchain-based microblock verification were successfully validated. The study explicitly distinguishes empirical forecasting and platform validation from conceptual digital-twin and ESG scenarios, which are included only to demonstrate project decision-support capabilities. The proposed framework provides a practical basis for intelligent environmental monitoring, short-term emission forecasting, and sustainability-oriented project management in industrial applications.

FINE-TUNING SMALL LANGUAGE MODELS FOR MENTAL HEALTH TEXT CLASSIFICATION: A PARAMETER-EFFICIENT APPROACH OUTPERFORMING ZERO-SHOT LARGE LANGUAGE MODELS

2026-03-10T09:31:53+05:00

Automated analysis of user-generated text is increasingly used to support scalable mental-health assessment, but deploying large language models for classification remains costly and latency-sensitive. This study evaluates whether parameter-efficient fine-tuning (PEFT) of a small model can provide a better accuracy–efficiency trade-off. We fine-tune SmolLM2-1.7B on two benchmarks: SWMH (5-class mental-health classification) and Dreaddit (binary stress detection). Evaluation follows two protocols: (1) primary full held-out test evaluation for the fine-tuned model, and (2) paired evaluation on identical 200-example subsets for fair cross-model comparison with zero-shot GPT-4o-mini and GPT-4o. On full held-out tests, SmolLM2-1.7B achieves 0.723 accuracy, 0.720 weighted F1, and 0.727 macro F1 on SWMH, and 0.808 accuracy, 0.808 weighted F1, and 0.808 macro F1 on Dreaddit. On paired subsets, SmolLM2-1.7B outperforms both zero-shot GPT baselines on both datasets (SWMH accuracy: 0.705 vs 0.600/0.645; Dreaddit accuracy: 0.755 vs 0.650/0.675). These results indicate that task-specific PEFT can make small models practically deployable and, under controlled paired evaluation, stronger than zero-shot larger models.

COMPARATIVE EVALUATION OF SEMANTIC SEGMENTATION MODELS FOR CRACK DETECTION

2026-06-25T23:05:13+05:00

Cracks on concrete and asphalt surfaces are an early warning that a dam, a bridge, or a road is starting to fail, and missing them can be expensive. Inspection is increasingly done with drones, which produce far more images than an engineer can check by hand, so the analysis has to be automated. In this paper we compare five semantic segmentation networks for crack detection and, unlike most such comparisons, we push all of them through exactly the same pipeline. Four are convolutional (U-Net, UNet++, DeepLabV3+, and a Feature Pyramid Network) and one is a transformer (SegFormer-B2). Every model sees the same data, the same 320 x 320 inputs, the same augmentation, the same Binary Cross-Entropy plus Dice loss, and the same 0.5 threshold, and all are trained with mixed precision on the public Crack500 benchmark and its official train, validation, and test split. We repeat each training five times with different seeds, report the mean and standard deviation on the test set, and check the gap between the two best models with a paired significance test. To see whether the ranking survives outside Crack500, we then run the trained networks, untouched, on a second dataset (DeepCrack). SegFormer-B2 was the most accurate model on Crack500 (IoU 0.595, F1 0.746), and the t-test shows its margin over UNet++ is well beyond chance. It stayed first on DeepCrack too (IoU 0.704, F1 0.826). We also report parameters, FLOPs, inference time, and memory, so the accuracy can be weighed against the cost of running each model.

A HYBRID SWARM INTELLIGENCE FRAMEWORK FOR ADAPTIVE LEARNING PATH RECOMMENDATION

2026-02-27T18:40:05+05:00

Personalized education requires learning path recommendations that adapt to individual performance while satisfying prerequisite constraints. However, generating optimal learning sequences presents significant computational challenges due to the combinatorial complexity of curriculum graphs, heterogeneous learner profiles, and directed prerequisite dependencies that constrain valid traversal orders. This paper proposes a Hybrid Swarm Intelligence Framework combining Particle Swarm Optimization (PSO), Ant Colony Optimization (ACO), and Artificial Bee Colony (ABC) for adaptive learning path recommendation. The hybrid architecture addresses distinct subproblems: PSO optimizes continuous feature weights and algorithm parameters, ACO constructs prerequisite-feasible paths through pheromone-guided graph traversal, and ABC refines solutions through local neighborhood exploration to escape suboptimal configurations. The curriculum is modeled as a directed weighted graph where nodes represent courses with learner-specific difficulty scores derived from behavioral features. Student behavioral data from 8,783 learners (6,968,707 records) inform learner-specific heuristics. The framework operates as a two-stage process: first, linear regression predicts GPA from behavioral features; second, the swarm intelligence pipeline selects paths maximizing predicted academic success. Experimental evaluation demonstrates a +20.69% improvement in mean path fitness scores compared to the Ridge baseline (p < 0.001, paired Cohen's d_z = 3.48). We compared the framework against Deep Learning baselines (LSTM and GRU) configured as sequence-aware grade regressors. While the optimally tuned GRU emerged as a strong predictor (+11.88% over baseline), it fell significantly short of the Hybrid Swarm. Results suggest that for curriculum sequencing under prerequisite constraints, combinatorial swarm optimization can outperform greedy decoding from neural predictors when training data is limited. By demonstrating that swarm intelligence algorithms can be systematically combined to address prerequisite satisfaction and behavioral personalization, this work offers a scalable, interpretable alternative to opaque deep learning approaches.

INTELLIGENT MODEL FOR PREDICTING METHANE CONCENTRATION IN INDUSTRIAL ENVIRONMENTS BASED ON IoT GAS ANALYZERS AND HYBRID MACHINE LEARNING ALGORITHMS

2026-06-09T09:33:41+05:00

Ensuring industrial safety at enterprises that use combustible and toxic gases requires continuous monitoring of the gas environment and timely identification of potentially hazardous situations. One of the most serious threats in the mining and industrial sectors is the accumulation of methane, which has a high explosion hazard and can lead to technogenic accidents. Traditional monitoring systems based on threshold alarm algorithms allow detection of exceedances of permissible gas concentrations; however, they do not provide the capability to forecast the development of hazardous situations.

This paper proposes an intelligent model for predicting methane concentration based on data obtained from IoT gas analyzers of the SENSOR - Mine 4GN series. The monitoring system integrates a distributed network of gas sensors, wireless data transmission, and machine learning algorithms for time-series analysis of gas concentration measurements.

Experimental data were obtained under laboratory conditions over a ten-day period using certified calibration gas mixtures simulating methane release events under different temperature conditions. For time-series analysis, a hybrid machine learning model was applied, combining a Long Short-Term Memory (LSTM) recurrent neural network and the ensemble algorithm XGBoost.

The obtained results demonstrate that the proposed model improves the accuracy of methane concentration dynamics prediction and enables early detection of hazardous trends in the gas environment. The use of predictive analytics methods combined with Internet of Things (IoT) gas analyzers significantly increases the effectiveness of industrial safety systems and reduces the risk of accidents at industrial facilities. The proposed methodology can be applied as a foundation for the development of intelligent predictive monitoring systems for mining, energy, and other hazardous industrial environments.

HIERARCHICAL DECISION MODELLING OF CLOUD ENTERPRISE DATA STORAGE SYSTEMS USING AHP

2026-02-06T18:11:51+05:00

Evaluating cloud enterprise data storage systems is mission-critical for ensuring reliability, security and compliance, interoperability, scalability, performance, vendor support, and cost-effectiveness in modern organizations where petabytes of data are stored. However, choosing the appropriate storage solution is complex, as enterprises must consider numerous strategic, technological, and operational dimensions. This article addresses these challenges by presenting a multiperspective assessment of cloud enterprise storage systems supported by multi-criteria decision making (MCDM) tool, such as Analytic Hierarchy Process (AHP). A three-tier hierarchy was constructed that organizes nine criteria into strategic, technological, and operational dimensions. Using this structure, three leading cloud storage platforms – Amazon S3, Microsoft Azure, and Google Cloud Storage – were evaluated. The opinions of twelve senior experts from National Information Technologies Joint-Stock Company (NITEC JSC) in Kazakhstan, all with experience in corporate cloud infrastructure, were collected using pairwise comparison questionnaires and summarized using the geometric mean. Local and global weights were obtained by calculating the priority based on eigenvectors with a consistency coefficient check and sensitivity analysis to confirm the reliability of the model. The results show that the criteria related to continuity and reliability prevail in decision-making, while pure productivity and manageability have a comparatively lower weights. Amazon S3 had the highest overall priority, followed by Microsoft Azure and Google Cloud Storage. This study contributes not only a practical solution for a multiperspective decision model but also a validated methodological blueprint that helps enterprises adapt continuously as technologies evolve. Future work will explore how hierarchical decision modeling can facilitate transparent and repeatable selection of enterprise cloud storage providers and can be expanded to include additional criteria such as cost and security in organization-specific applications.

REPRODUCIBLE EVALUATION OF ONTOLOGY-GUIDED HYBRID EXTRACTIVE QUESTION ANSWERING FOR KAZAKH HISTORY

2026-05-28T18:16:51+05:00

Reproducibility and comparability remain important methodological concerns in domain-specific question answering, particularly in low-resource educational settings. This study examines the impact of ontology-guided context filtering on extractive question answering in the Kazakh language under controlled and reproducible conditions. A benchmark corpus on the History of Kazakhstan was constructed from 11 school textbooks for grades 5–10 and canonically indexed using deterministic section identifiers. Fixed train–test splits, standardized chunking, and unified preprocessing and evaluation scripts ensured consistency across experiments.

Three question answering systems were evaluated under identical constraints: a neural extractive baseline operating on unfiltered context, an ontology-only system relying on structured knowledge representations, and a hybrid system applying ontology-guided context filtering prior to neural answer extraction. Filtering was conducted under an inference setting that excludes access to gold document or section identifiers, preventing the use of privileged metadata.

The hybrid system demonstrates modest improvements over the neural-only baseline. Exact Match accuracy increases from 68.87% to 70.45%, while the token-level F1 score improves from 80.02% to 81.68%. Ranking-based retrieval metrics also show small gains, with Mean Reciprocal Rank increasing from 0.870 to 0.892 and Recall@10 from 95.1% to 96.9%. Inference latency is reduced from 20.01 ms to 6.83 ms per question. Paired statistical testing confirms that the improvement in F1 is statistically significant, whereas improvements in Exact Match are less conclusive.

Ablation experiments suggest that these differences are not attributable to context length reduction alone, but are associated with the interaction between ontology-guided context selection and neural reader training. Overall, ontology-guided filtering can provide incremental benefits for extractive question answering in a domain-specific, low-resource educational setting under transparent and reproducible experimental conditions.

COMPARISON OF YOLOV8N AND YOLOV12N MODELS FOR SCALABLE REAL TIME VIDEO MONITORING OF CONSTRUCTION SITES

2026-05-05T16:09:32+05:00

This study presents a comparative analysis of the object detection models YOLOv8n and YOLOv12n in their Nano configurations for real-time construction site monitoring tasks. Key model characteristics were evaluated, including detection accuracy, recall, inference speed, and computational resource consumption across various server configurations. Particular attention was given to the use of the Slicing Aided Hyper Inference (SAHI) method to improve the detection quality of small objects, as well as the impact of scene complexity on system performance.

YOLOv12n achieved significantly higher accuracy and recall than YOLOv8n. In addition, the integration of the SAHI method further enhanced small object detection performance. Parallel stream processing showed that servers equipped with Tesla L4 GPUs offered the best trade-off between inference speed and stability under multi-threaded workloads. The inclusion of data augmentation techniques simulating various weather conditions enhanced model robustness, especially for YOLOv8n, in noisy and complex scenes.

The findings underscore the clear advantage of YOLOv12n for applications requiring high recall in challenging environments, while YOLOv8n demonstrated strong efficiency on resource-constrained systems. The proposed approaches can be utilized in the development of intelligent monitoring systems for construction sites, industrial facilities, and logistics hubs, including automated access control and real-time employee activity tracking.

A BREAKPOINT MODEL FOR DATA VOLUME AND NETWORK CAPACITY IN INTERNET OF THINGS

2026-06-03T15:04:37+05:00

The migration from fifth generation (5G) to sixth generation (6G) mobile networks is driven by the Massive Internet of Things (IoT), which targets device densities of up to ten million units per square kilometer. Two scaling limits govern such systems: a data-volume limit, beyond which batch analytics must give way to distributed streaming, and a network-capacity limit, beyond which the radio access network can no longer carry the aggregate uplink. These limits have so far been treated independently, and no empirical anchor exists for Central Asia. This study makes three contributions. First, it derives a closed-form, dimensionless separation ratio that couples the two breakpoints through shared system parameters and determines which limit binds first. Second, the model is instantiated using 145 regulated 5G New Radio measurements from the commercial Tele2 network in Astana, Kazakhstan, across five load scenarios spanning 13 to 100,000 emulated devices. Third, the closed-form structure of the ratio suggests the splitting is a general trait of uplink-bound IoT domains, though this study empirically instantiates only the vertical-farming case. The instantiated model yields a separation ratio of about thirty-two, with breakpoints at twenty-eight nodes for network capacity and eight hundred eighty for data volume. The measured round-trip time of 124.8 milliseconds and the measured uplink of 0.19 megabits per second miss the International Mobile Telecommunications-2030 (IMT-2030) targets by factors of 1248 and about 5.3 × 10⁵, respectively. Because the network-capacity limit binds first, the analytics regime required for dense deployment is unreachable on 5G. For the measured single-operator deployment, 6G is a precondition rather than an enhancement; generalization to other operators and sites requires further measurement.

MULTI-AGENT EVACUATION SIMULATION WITH FIRE AND SMOKE DYNAMICS: A REPRODUCIBLE BASELINE SINGLE-FLOOR STUDY

2026-06-12T11:01:55+05:00

Building evacuation during fire emergencies remains a critical challenge for urban safety, particularly in rapidly developing cities with complex high-rise architecture. This study presents a novel multi-agent evacuation simulation framework implemented using the Mesa agent-based modeling platform, specifically tailored for building safety analysis in Astana, Kazakhstan. The model integrates dynamic fire and smoke propagation with heterogeneous agent behaviors, simulating four distinct evacuee types: healthy adults, elderly/children, panicking individuals, and leaders. Through comprehensive simulation experiments, we analyzed evacuation patterns, exit utilization, agent performance differences, and spatial density dynamics under fire conditions. Results demonstrate that evacuation completion occurs within 180 simulation steps, with significant performance variations across agent types. Leaders achieved the fastest evacuation times (mean: 85 steps), while elderly agents required substantially longer (mean: 142 steps). Fire and smoke spread analysis revealed exponential growth patterns, with smoke cells expanding from 5 to over 120 cells within 100 steps. Exit utilization analysis showed balanced distribution across three exits (32%, 35%, 33%), indicating effective pathfinding algorithms. This research addresses a critical gap in regional fire safety literature, as no prior studies have specifically examined multi-agent evacuation dynamics in the Astana context. The findings provide actionable insights for building design, emergency planning, and evacuation protocol optimization in Central Asian urban environments characterized by extreme climatic conditions and modern high-rise architecture.

EVENT-DRIVEN SERVERLESS ARCHITECTURE FOR SMART CITY E-PARTICIPATION PLATFORMS UNDER DIGITAL SOVEREIGNTY CONSTRAINTS

2026-03-02T13:51:26+05:00

With the digital transformation of urban governance and the development of the "smart city" concept, the role of digital civic participation platforms for engaging the public in decision-making processes is increasing. However, most existing e-participation systems are based on monolithic or containerized architectures, which limits their scalability, resilience to peak loads, and the ability to integrate intelligent analytics. This article proposes an event-driven, serverless architecture for a digital civic participation platform designed for use in urban ecosystems while adhering to digital sovereignty requirements. The architecture is based on a formal Citizen Participation Event Model (CPM), which views user actions as a stream of events processed in a distributed computing environment. The architecture natively integrates artificial intelligence modules for biometric user verification, behavioral anomaly detection, and text content analysis. Particular attention is paid to personal data protection and compliance with national information localization requirements using a hybrid data management model that separates personal data from anonymized analytical events. An experimental evaluation of the proposed solution was conducted under simulated conditions of typical urban civic engagement scenarios. The results demonstrate stable system operation under loads of up to 10,000 concurrent users, predictable latency dynamics, and a reduced total cost of ownership compared to a containerized architecture based on Kubernetes. It is concluded that event-driven serverless architectures can serve as the technological foundation for next-generation digital civic engagement platforms that combine scalability, intelligent data processing, and compliance with digital sovereignty requirements.

Development of a Material Database and an Automated Selection Algorithm for a Surgical Robotic Manipulator

2026-06-12T15:06:18+05:00

Material selection for surgical robotic manipulators remains a challenging engineering problem because candidate materials must simultaneously satisfy mechanical, biomedical, environmental, manufacturing, and economic requirements. This challenge is particularly important in robotic systems for knee arthroplasty, where structural reliability and positioning accuracy strongly depend on material properties. This study presents the development of a structured component-oriented material database and an automated material selection algorithm for a collaborative surgical manipulator intended for knee joint endoprosthesis procedures. The database includes metallic materials, engineering polymers, composites, and ceramics together with their mechanical, environmental, biomedical, manufacturing, and economic characteristics. Based on this database, a multi-stage decision-support framework was developed that integrates component-specific material retrieval, hard-constraint filtering, criterion normalization, adaptive weighting, TOPSIS-based ranking, and post-ranking feasibility verification. Unlike conventional approaches that evaluate all materials simultaneously, the proposed framework retrieves only materials relevant to the selected manipulator component, reducing the search space and improving decision consistency. A numerical case study was performed for the load-bearing structure. The algorithm retrieved 21 candidate materials with complete data, of which 6 satisfied the mandatory engineering and biomedical constraints. These materials were subsequently ranked using the TOPSIS method, with AISI 316 identified as the most suitable material because of its balanced combination of mechanical strength, corrosion resistance, sterilization compatibility, and durability. Comparative evaluation using the SAW method produced the same three highest-ranked materials, while sensitivity analysis demonstrated that the ranking remained stable under moderate variations in weighting coefficients. The proposed framework provides a transparent, reproducible, and practically applicable approach to material selection for surgical robotic manipulators and establishes a foundation for future integration with CAD/CAE environments, intelligent decision-support systems, and machine-learning-assisted engineering design.

IDENTIFYING USER BEHAVIORAL PATTERNS IN WEB BROWSER ACTIVITY USING THE K-NEAREST NEIGHBORS ALGORITHM

2026-02-26T11:26:09+05:00

The rapid growth of internet usage has significantly increased the volume of digital traces created by users, making web browsers a critical source of evidence in digital forensics. Web browsers store extensive user-related data, including browsing history, search queries, visited domains, cookies, and cache files, which collectively reflect behavioral patterns and decision-making processes. However, existing forensic tools primarily focus on manually extracting and visualizing browser artifacts, providing limited capabilities for automated behavioral analysis. This study addresses this shortcoming by proposing an automated method for identifying and analyzing user behavioral patterns based on web browser activity using machine learning techniques. The proposed approach is based on extracting and structuring browser history data from personal computers, followed by similarity-based classification using the k-nearest neighbors algorithm. Behavioral characteristics are extracted from recurring search queries, domain frequency, and content categories of visited websites, enabling the creation of a structured behavioral dataset. To justify the model selection, the k-nearest neighbors classifier was evaluated alongside logistic regression and Gaussian naive Bayes classifiers using standard performance metrics, including accuracy, precision, recall, F1-score, and area under the curve. Experimental results show that the k-nearest neighbors model achieves the highest recall and F1-score, indicating superior performance in identifying clustered and nonlinear patterns of user behavior compared to the baseline and probabilistic models. The results confirm that similarity-based classification is particularly effective for browser activity analysis, where user actions tend to form localized groups in high-dimensional feature spaces. The proposed method enables the transformation of raw browser data into interpretable user behavior profiles that integrate interests, behavioral indicators, and visit frequency statistics. This improves the practical applicability of web browser analysis by supporting structured behavioral assessment rather than raw data analysis. Overall, the study demonstrates the feasibility and effectiveness of integrating machine learning-based behavioral analysis into web browser forensics workflows and lays the foundation for future research incorporating temporal features, larger datasets, and hybrid learning approaches.

DEEP LEARNING FOR LYMPH NODE DETECTION IN CT IMAGING: A BENCHMARKING STUDY OF OBJECT DETECTION FRAMEWORKS

2026-06-21T12:06:26+05:00

Accurate detection of lymph nodes on computed tomography (CT) scans is essential for cancer staging and treatment planning, yet manual identification on CT is time-consuming and subject to considerable inter-observer variability, motivating the development of reliable automated methods. Automating this task remains challenging because many nodes are very small, contrast with surrounding tissue is frequently low, and target instances constitute only a minuscule fraction of all voxels, an extreme class-imbalance problem. In this study, we evaluate five state-of-the-art object detection architectures (Faster R-CNN, YOLOv8-m, MULAN, YOLOv11-m, and RT-DETR-L) under standardized training and evaluation protocols on two datasets with different anatomical compositions: a heterogeneous multi-region dataset (Ver1) and a mediastinal-focused dataset (Ver2). Both datasets were processed using a memory-efficient 2.5D pipeline that feeds three consecutive CT slices as input, capturing volumetric context while avoiding the computational cost of full 3D models. On Ver2, Faster R-CNN and MULAN achieved the highest mAP50 of 0.641, while RT-DETR-L led in recall (0.646) and F1-score (0.630). Every model improved its mAP50 on the anatomically focused dataset. Because the same shift appeared across five structurally different detectors, we attribute it to the detection task itself rather than to any particular architectural design. Paired testing on Ver2 showed that MULAN significantly outperformed both YOLO variants in mAP50, with Faster R-CNN following the same tendency. On Ver1, no pair of models differed significantly, so performance converges once anatomical diversity increases. Taken together, the comparison gives clinicians practical grounds for choosing between speed, precision, and sensitivity, and supplies reference baselines for later work on automated lymph node detection in CT.

EXPLORATORY DATA ANALYSIS AS A FOUNDATION FOR TRUSTWORTHY CLINICAL MACHINE LEARNING

2026-05-05T16:23:59+05:00

Exploratory data analysis is frequently treated as a preliminary or purely descriptive step in clinical machine learning studies. However, many methodological failures of clinical prediction models arise not from algorithm selection, but from unrecognized properties of the data-generating process, including temporal non-stationarity, non-random missingness, and systematic documentation effects. These issues are particularly critical in early clinical triage settings, where risk assessment must be performed at hospital admission using limited and heterogeneous clinical information. In this study, we examine exploratory data analysis as a dedicated methodological stage that precedes and constrains downstream machine learning design in early triage–level mortality risk prediction. Using a large-scale national registry of hospitalized patients with confirmed COVID-19 infection, we conduct a structured exploratory analysis of admission-level clinical data, focusing on temporal drift in outcome prevalence and patient case-mix, feature-level missingness mechanisms, documentation bias in symptom recording, and structural redundancy among clinical indicators. The results demonstrate pronounced temporal non-stationarity across pandemic phases, indicating that commonly used random data splitting strategies may introduce methodological bias even prior to model development. Missingness patterns are shown to be structured and outcome-dependent rather than random, reflecting both clinical severity and organizational processes. In addition, several apparent univariate associations between symptoms and outcomes are attributable to documentation bias rather than underlying clinical effects. These findings provide concrete diagnostic evidence that exploratory analysis directly constrains valid choices related to validation design, handling and representation of missing information, feature specification, and interpretation boundaries. The findings suggest that exploratory data analysis should be regarded as a fundamental methodological stage supporting reliable and trustworthy clinical machine learning in early-stage decision-making settings.

GRADIENT-VELOCITY VECTOR LYAPUNOV ANALYSIS OF APERIODIC ROBUST STABILITY IN CYBER-DEFENSE ASSET DYNAMICS

2026-06-16T12:16:20+05:00

An enterprise security posture changes over time as exploits are disclosed, patches are applied, and assets age out of vendor support. Empirical incident studies indicate that many breaches follow not from a single point of failure but from a gradual decline of the posture over weeks and months. In this paper the population of protected assets in an information-technology infrastructure is modeled as a multidimensional, continuous-time linear dynamical system and analyzed for stability by the gradient-velocity vector Lyapunov function method. Robust aperiodic stability conditions are derived as an explicit system of linear inequalities in the patching, aging and recovery rates. These conditions are sufficient rather than necessary, so the divergent regime is confirmed by a direct spectral check of the model matrix. For the configuration studied it shows a single real positive eigenvalue, and exposure diverges monotonically and aperiodically; this loss of stability is interpreted as a structural, endogenous mechanism rather than an isolated random event. The same framework yields a linear feedback patch-management controller that places the closed-loop dynamics in the aperiodic robust stability class, with closed-form gains. A numerical experiment on a five-cohort asset model and a fourth-order companion-form plant shows that this controller reduces the residual-risk norm by about four orders of magnitude over twelve weeks, whereas the uncontrolled system grows by more than five. The method turns the intuition that “patching faster helps” into a quantitative structural criterion evaluable, in principle, from an enterprise’s own asset telemetry without a statistical baseline. All coefficients were assigned rather than identified from operational data and are reported in full, so the results hold within the proposed model only; empirical validation on operational data remains essential future work.

TEACHER-CURATED AI FOR COMMUNICATIVE CHINESE ACQUISITION IN UNIVERSITIES OF KAZAKHSTAN

2026-02-25T15:34:53+05:00

As China's global economic and political influence continues to expand, the demand for Chinese language proficiency in Kazakhstan has reached an all-time high, yet students frequently struggle with the mastery of the logographic writing system, which often remains isolated from communicative practice. This study investigates the challenge of bridging the gap between mechanical character acquisition and functional communication among first-year students at Kazakh Ablai Khan University, set against the strategic national mandate for digital transformation and the integration of artificial intelligence into the education system. To address systemic orthographic errors and a lack of semantic awareness, the research evaluated a teacher-curated artificial intelligence framework designed to shift pedagogy from rote memorization to active semantic use. The methodology employed a quasi-experimental, mixed-methods design involving 29 students during the 2025–2026 academic year, beginning with the analysis of a diagnostic corpus of over 120 homework assignments to categorize baseline errors. This was followed by an eight-week intervention where the experimental group utilized specialized artificial intelligence-assisted "Free Chat" sessions for real-time scaffolding. Results from an independent samples t-test revealed a significant performance divergence, with the experimental group increasing character accuracy from 68.50% to 84.20%, compared to a control group mean of 71.80%. Qualitative analysis of interaction logs and interviews demonstrated that the artificial intelligence successfully mitigated cognitive load by deconstructing complex characters into manageable semantic narratives, though students noted a "digital-analog gap" regarding handwriting recall. The study concludes that teacher-curated artificial intelligence serves as a powerful cognitive bridge that corrects fossilized structural errors and fosters communicative risk-taking. These findings provide a scalable blueprint for integrating human-centric digital tools in university language departments, emphasizing that curated artificial intelligence enhances rather than replaces the cognitive processes essential for autonomous linguistic competence.

ENHANCING EDUCATIONAL COMPETENCY FRAMEWORKS IN KAZAKHSTAN THROUGH A HYBRID BLOOM’S TAXONOMY AND GENERATIVE AI APPROACH WITH PCA-BASED VALIDATION

2026-02-23T17:53:24+05:00

Competency-based education frameworks require clear differentiation of learning outcomes across qualification levels to support effective curriculum design and ensure alignment with dynamic labor market requirements. However, national professional standards frequently contain overlapping, ambiguous, or weakly differentiated competency statements, which complicates the systematic development of educational programs. This study proposes a hybrid methodology that integrates the hierarchical framework of Bloom’s Taxonomy with generative artificial intelligence to improve the structural differentiation of competency statements within Kazakhstan’s professional standards.

A comprehensive dataset of 595 professional standards containing 8,736 competency records was compiled and analyzed using natural language processing techniques. The textual competencies were vectorized utilizing TF-IDF and examined through Principal Component Analysis (PCA) to assess structural differentiation across qualification levels. To address existing gaps and enrich the dataset, 4,024 additional competencies for advanced qualification levels 6-8 were generated using GPT-4. This process was guided by Bloom’s Taxonomy action verbs to accurately reflect progressive cognitive complexity and depth of knowledge.

Comparative analysis conducted before and after data augmentation showed a moderate increase in explained variance in the first principal components, rising from approximately 4.5% to 7.5%. This shift suggests improved structural differentiation within the augmented competency dataset. Additionally, classification experiments using multiple machine-learning models demonstrated notably enhanced performance metrics after augmentation, indicating greater separability of competency statements across the targeted qualification levels.

The findings suggest that combining established pedagogical taxonomies with generative AI can efficiently support the systematic refinement of competency frameworks. Rather than replacing expert curriculum designers, generative AI functions as a powerful analytical support tool for identifying structural gaps and improving competency differentiation within large collections of educational standards.