Sinkron : jurnal dan penelitian teknik informatika

Hybrid Multilayer Architecture Integrating Suricata, Wazuh, and Cyber Threat Intelligence for Drive-by-Download Malvertising Detection

2025-12-04T15:13:39+00:00

Malvertising has emerged as a serious cybersecurity threat, leveraging legitimate advertising networks to deliver malware through drive-by-download techniques without requiring user interaction. Existing standalone network- or host-based detection solutions provide limited protection because they lack integrated visibility and contextual validation across detection layers. However, no existing research has specifically evaluated the integration of Suricata, Wazuh, and VirusTotal for endpoint-focused malvertising detection, creating a critical gap in multi-layer defense strategies. This study proposes a hybrid multilayer architecture combining Suricata as a Network Intrusion Detection System, Wazuh as a Host-based Intrusion Detection and Prevention System, and VirusTotal as an external Cyber Threat Intelligence source to provide correlated threat detection and automated mitigation. The system was evaluated in a controlled virtual laboratory consisting of attacker, victim, and SIEM environments replicating real malvertising scenarios. The results show that the proposed architecture successfully detected malicious payloads and completed an end-to-end detection-to-mitigation cycle in approximately 5-7 seconds while maintaining zero false positives under non-malicious conditions. This research contributes a practical and reproducible architecture for endpoint-based malvertising detection, demonstrating effective multi-layer correlation and rapid autonomous response. The limitation of this study lies in its reliance on signature-based detection and external API communication, which may reduce effectiveness against zero-day threats or offline deployments.

IoT Sensor Data Analysis for Early Fire Detection Using Dynamic Threshold

2025-10-27T08:21:22+00:00

Early fire detection using Internet of Things (IoT) technology plays a vital role in minimizing potential material losses and casualties. Conventional systems generally still rely on static thresholds that are less adaptive to environmental dynamics, leading to high false alarm rates and delayed detection. This study proposes a dynamic threshold approach based on a hybrid method of Fuzzy Logic–Random Forest–Adaptive Z-Score and compares it with the static threshold method. Testing was conducted using publicly available secondary datasets, and the algorithms were implemented and tested in Jupyter Notebook. Evaluation was performed using accuracy, false alarm rate (FAR), detection time, F1-score, precision, and recall metrics. The test results show that the dynamic threshold method provides better performance with an increase in accuracy from 59.5% to 74.8%, a decrease in FAR from 31.1% to 14.3%, and a reduction in detection time from 21 seconds to 0 seconds. In addition, the F1-score increased from 0.459 to 0.638, precision from 0.473 to 0.716, and recall from 0.446 to 0.575. These results show that the dynamic threshold approach is more adaptive and reliable in IoT-based fire detection systems than conventional static threshold methods.

CataractAsist: Convolutional Neural Network-Based Early Detection System for Cataracts

2025-12-03T06:13:24+00:00

Cataract disease is one of the leading causes of blindness worldwide, especially in developing countries with limited access to healthcare facilities. To address this challenge, this study aims to develop an automated cataract detection system using the Convolutional Neural Network (CNN) method. This system is designed to classify eye images into three classes, namely Normal, Mature, and Immature, by utilizing the "Senile Cataract" dataset from the Kaggle platform. The research methods include image pre-processing, feature extraction using the VGG16 model through transfer learning, model training with augmentation techniques, and performance evaluation using accuracy, precision, recall, f1-score, and confusion matrix metrics. The test results show that the model is capable of achieving 95% accuracy, with the highest f1-score in the Normal class at 0.96. Confusion matrix analysis shows excellent prediction rates for all classes, although there are slight classification errors between the Immature and Mature classes. In conclusion, this CNN-based cataract detection system is proven to be effective and accurate, and has great potential to be applied in web-based healthcare services as an automatic early diagnosis tool for eye diseases.

A Blockchain-Assisted Neural Network Model for Flood Detection and Data Integrity Assurance

2025-11-04T11:50:10+00:00

Flooding is one of the most frequent natural disasters and has substantial impacts on social, economic, and environmental conditions. Therefore, early detection plays a critical role in minimizing potential damage and supporting effective disaster response. This study proposes a Flood Detection System Using an Artificial Neural Network (ANN) with Blockchain-Based Data Integrity, which integrates predictive analytics and secure data management in a unified framework. The ANN model processes multisource environmental data such as satellite imagery, rainfall intensity, water level fluctuations, and soil moisture obtained from Google Earth Engine (GEE). Training is conducted using a sigmoid activation function and backpropagation algorithm to identify spatial and temporal patterns associated with flood-prone areas. The resulting classification outputs are stored in a blockchain ledger to ensure immutability, transparency, and protection against unauthorized data modification. Experimental evaluations demonstrate that the proposed hybrid approach achieves an accuracy of 95.82%, supported by precision, recall, and F1-score values that indicate consistent model performance across varying environmental conditions. The integration of blockchain provides verifiable and tamper-proof documentation of ANN predictions and related metadata. Overall, this research contributes a reliable, secure, and technically robust method for early flood detection, offering valuable support for data-driven decision-making in disaster mitigation and environmental risk management.

Enhancing Feature-Efficient Network Intrusion Detection Using Gradient Boosting and Chi-Square Selection on NSL-KDD

2025-12-04T15:32:52+00:00

This study examines the growing complexity of cyber threats that increasingly challenge the effectiveness of traditional Network Intrusion Detection Systems (NIDS). Modern attacks, particularly zero-day intrusions, require detection approaches capable of handling high-dimensional network traffic data. However, existing studies rarely examine the trade-off between feature efficiency and generalization performance in boosting-based NIDS under controlled feature-reduction strategies. Moreover, the role of statistical feature selection in mitigating overfitting in classical boosting models remains underexplored. This study evaluates the performance of NIDS by combining boosting ensemble algorithms, namely AdaBoost, Gradient Boosting, and XGBoost, with filter-based feature selection methods, including Information Gain, Chi-Square, and ReliefF. The NSL-KDD dataset is used as the primary benchmark, with Min–Max normalization applied during preprocessing to ensure numerical feature consistency. Model development is conducted using Orange Data Mining, and performance is assessed through 10-fold cross-validation. Experimental results show that Gradient Boosting achieves the highest baseline accuracy among the evaluated models. Further performance improvements are obtained through feature selection, with the Chi-Square method yielding the best result at 81.2% accuracy using 19 selected features. Information Gain also enhances performance, achieving 80.8% accuracy with 13 features, while ReliefF provides comparatively lower gains. These findings demonstrate that effective feature reduction improves generalization performance, reduces computational complexity, and mitigates overfitting. Overall, the proposed combination of Gradient Boosting and statistical feature selection provides a feature-efficient, generalizable intrusion detection strategy for modern network environments.

ELECTRE-Based Decision Support Model for LPG Base Location Optimization

2026-01-01T18:18:58+00:00

The kerosene to LPG (Liquefied Petroleum Gas) 3 kg conversion program since 2007 has successfully improved household energy efficiency, but equitable access to bases in remote areas is still an obstacle. In Tabanan Regency, Bali, eight villages do not have access to 3 kg LPG bases, making it difficult for the community to obtain LPG at the Highest Retail Price (HET) and timely supply. This research develops a decision-making model using the ELECTRE method to recommend optimal base locations based on a case study of four villages: Pupuan Sawah, Dalang, Mundeh, and Belatungan. The model integrates 15 criteria including population density, infrastructure accessibility, existing base distance, and the presence of public facilities with a multi-stakeholder approach. The model is expected to be a tool for LPG agents and policy makers in determining the optimal base location and supporting equitable distribution of subsidized energy.

Security Evaluation of Indonesian LLMs for Digital Business Using STAR Prompt Injection

2025-12-05T03:47:40+00:00

The adoption of Large Language Models (LLMs) in digital business systems in Indonesia is rapidly increasing; however, systematic security evaluation against Indonesian language prompt injection remains limited. This study introduces the Indonesian Prompt Injection Dataset, consisting of 50 attack scenarios constructed using the STAR framework, which combines structured instruction variations with sociotechnical context to expose potential model vulnerabilities. The dataset was used to evaluate three commercial LLM platforms ChatGPT using a GPT-4 class lightweight variant (OpenAI), Gemini 2.5 Flash (Google), and Claude Sonnet 4.5 (Anthropic) through controlled experiments targeting instruction manipulation in Indonesian. The results reveal distinct robustness profiles across models. Gemini 2.5 Flash exhibits moderate observed resilience, with 76% of scenarios classified as medium risk and 12% as high risk. ChatGPT demonstrates higher observed robustness under the tested scenarios, with 88% of cases classified as low risk and no high-risk outcomes. Claude Sonnet 4.5 shows intermediate observed resilience, with 72% low-risk and 28% medium-risk scenarios. High-risk cases primarily involve direct role override, urgency- or emotion-based prompts, and anti-censorship instructions, while structural ambiguities and multi-intent manipulations tend to result in medium risk, and mildly persuasive prompts fall under low risk. These findings suggest that while contemporary LLM defense mechanisms are effective against explicit attacks, contextual and emotionally framed manipulations continue to pose residual security challenges. This study contributes the first Indonesian-language prompt injection dataset and demonstrates the STAR framework as a practical and standardized approach for evaluating LLM security in digital business applications.

SVM-Based Pediatric Disease Classification Model from the Balinese Lontar Usada Rare Manuscript

2025-11-11T17:17:46+00:00

Lontar Usada Rare is a traditional Balinese manuscript containing pediatric medical knowledge based on local wisdom, yet its narrative format limits accessibility and utilization in modern contexts, while its physical fragility threatens long-term preservation. This study aims to develop a pediatric disease classification model using a Support Vector Machine (SVM) combined with Term Frequency–Inverse Document Frequency (TF-IDF) weighting to support the digitalization of Balinese traditional medicine. A total of 422 data samples were collected through expert interviews and manuscript analysis, covering symptoms, disease types, herbal ingredients, and treatment procedures. The research stages included text preprocessing (cleansing, tokenizing, stopword removal, stemming), manual labeling into 35 disease classes, and model evaluation using five train–test split ratios (80:20 to 60:40) with variations of the complexity parameter C (0.5, 1, 10, 100, 1000). The best performance was achieved using C=10 with an 80:20 ratio, resulting in 87.06% accuracy, 91.55% precision, 87.06% recall, and an F1-score of 87.96%. Confusion matrix analysis showed strong classification performance for most classes, although minority classes with overlapping symptoms exhibited misclassification. Overall, the TF-IDF and linear SVM combination effectively classifies pediatric disease symptoms from Lontar Usada Rare and contributes to the preservation and digital transformation of Balinese traditional medical knowledge for potential modern healthcare applications.

Feature-Level Fusion of DenseNet121 and EfficientNetV2 with XGBoost for Multi-Class Retinal Classification

2025-12-05T10:06:31+00:00

Accurate and efficient classification of retinal fundus images plays a critical role in supporting the early diagnosis of ocular diseases. However, models relying on a single deep learning backbone often struggle to capture the multi-scale and heterogeneous characteristics of retinal lesions, leading to unstable performance across visually similar disease classes. To address this limitation, this study proposes a novelty feature-level fusion framework that integrates complementary representations from DenseNet121 and EfficientNetV2-s, followed by classification using XGBoost. The fusion pipeline extracts 1024-dimensional features from DenseNet121 and 1280-dimensional features from EfficientNetV2-s, which are concatenated into a unified 2304-dimensional feature vector. Experiments were conducted on a dataset of 10,247 retinal fundus images spanning six categories: Central Serous Chorioretinopathy, Diabetic Retinopathy, Macular Scar, Retinitis Pigmentosa, Retinal Detachment, and Healthy. The proposed fusion model achieved an accuracy of 91.60%, outperforming DenseNet121 XGBoost (91.31%) and EfficientNetV2-s XGBoost (89.70%). Moreover, the fusion strategy demonstrated improved class-level stability, particularly for visually similar retinal disorders where single-backbone models exhibited higher misclassification rates. This study contributes a lightweight yet effective multi-backbone feature-level fusion approach that enhances discriminative representation and classification stability without increasing model complexity. In addition, the use of XGBoost introduces a tree-based decision mechanism that is inherently more interpretable than conventional fully connected layers, offering potential advantages for clinical analysis. Overall, the results highlight the effectiveness of multi-backbone feature fusion as a reliable strategy for automated retinal disease classification.

Multi-Disease Retinal Classification Using EfficientNet-B3 and Targeted Albumentations: A Benchmark on Kaggle Retinal Fundus Images Dataset

2025-12-01T17:59:41+00:00

Retinal diseases remain one of the leading causes of blindness worldwide. This study develops a deep learning pipeline for multiclass retinal disease classification using EfficientNet-B3 combined with Albumentations to improve generalization. We target four classes: cataract, diabetic retinopathy, glaucoma, and normal. We use the Kaggle Retinal Disease dataset (4,217 fundus images) divided into 70% training, 10% validation, and 20% testing. Images are resized to 224×224 and augmented with horizontal flip, random brightness contrast, CLAHE, shiftscale rotate, crop, gamma correction, and elastic transformation. The EfficientNet-B3 backbone is refined after head training with warm-up and learning rate regularization (batch normalization, dropout). After 50 epochs, the best validation performance reaches 0.9526, and on the hold-out test set, the model achieves 95.38% overall accuracy. The F1 scores per class were 1.0000 (diabetic retinopathy), 0.9685 (cataract), 0.9255 (normal), and 0.9184 (glaucoma). Confusion analysis showed that most errors involved glaucoma being misclassified as normal, likely due to optic disc similarities. These results demonstrate that EfficientNet-B3 with targeted augmentation provides accurate and reliable multi-disease screening of fundus images, with the potential to support faster and more consistent triage in clinical workflows. Future research should expand clinical validation and explore attention mechanisms or multimodal input to reduce glaucoma-normal ambiguity.

Efficient CNN-Based Classification of SARS-CoV-2 Spike Gene Sequences Using Alignment-Free Encoding

2025-12-14T05:15:29+00:00

The COVID-19 pandemic caused by SARS-CoV-2 continues to challenge the global health system through the emergence of various variants with genetic characteristics that affect vaccine transmission and effectiveness. Conventional identification methods such as Whole-Genome Sequencing (WGS) have high accuracy but are constrained by significant cost and time. Most classification studies today still rely on complex hybrid architectures such as CNN-LSTM or image-based representations that increase computational load. This study aims to develop an efficient and lightweight pure Convolutional Neural Network model based on alignment-free encoding to classify five Variant of Concern (VOC) variants of SARS-CoV-2 (Alpha, Beta, Delta, Gamma, and Omicron) with an exclusive focus on the Spike gene sequence. The dataset consists of 5,000 Spike gene sequences that are represented using integer encoding and standardized with zero-padding. CNN proposed Lightweight architecture consists of four 1D convolution layers with a total of approximately 1.6 million parameters. The test results show that the model achieves excellent performance with an overall accuracy of 98.93%. The precision, recall, and F1-score values averaged 0.99, while the analysis of the ROC curve showed AUC values above 0.99 for all variants. This approach has proven to be efficient and effective, offering a fast, scalable, and resource-efficient solution to support real-time genomic surveillance systems in future pandemic mitigation.

Heart Disease Classification Using Optimised XGBoost and Random Forest with SHAP Explanations

2025-11-11T17:04:48+00:00

Heart disease remains one of the leading causes of global morbidity, creating a need for accurate and interpretable computational tools to support early diagnosis. However, many existing studies on the Cleveland Heart Disease dataset rely on limited validation protocols, apply only a single hyperparameter optimisation strategy, or provide narrow explainability analyses, which can lead to optimistic performance estimates and inconsistent clinical insight. This study addresses these gaps by proposing a classification-based prediction framework that evaluates Random Forest and XGBoost for binary heart-disease classification under three hyperparameter optimisation strategies random search, Bayesian optimisation, and particle swarm optimisation (PSO) within a nested, anti-leakage cross-validation design, while SHAP is employed to analyse model interpretability across the best-performing configurations. The experimental results show that the ensemble classifiers achieve strong and consistent performance, with ROC–AUC values ranging from 0.8908 to 0.9089 across all scenarios; Random Forest optimised with PSO obtained the highest ROC–AUC (0.9089 ± 0.0146) and F1-score (0.8188 ± 0.0206), whereas XGBoost with Bayesian optimisation reached comparable performance without statistically significant differences. SHAP analyses identified oldpeak, ca, thal, cp, thalach, and exang as the most influential features, in line with established clinical indicators of myocardial ischemia and perfusion abnormalities. These findings indicate that combining tree-based ensemble classifiers with systematic hyperparameter optimisation and SHAP-based interpretability can enhance the reliability and transparency of heart-disease classification on the Cleveland dataset, while highlighting the need for further validation on contemporary, multi-centre clinical data.

Performance Evaluation and Optimization of an IoT-Based Fish Smoking Monitoring System for Ensuring Product Quality

2025-12-22T02:56:07+00:00

Fish smoking is a widely used preservation method; however, the quality of smoked fish is highly dependent on the stability of temperature, humidity, and smoking duration. Manual control of these parameters has limitations and may reduce product quality. Existing studies on fish smoking monitoring systems primarily focus on temperature control without providing quantitative evaluation of how multi-parameter process stability affects product quality and shelf life. This study aims to design and implement an Internet of Things (IoT)-based monitoring system for fish smoking equipment to ensure the quality of smoked fish. The research method used is Research and Development (R&D), which includes needs analysis, system design, development, testing, and evaluation stages. The system integrates temperature and humidity sensors, a microcontroller, and an IoT platform for real-time monitoring. The test results show that the system is capable of monitoring the smoking chamber temperature within a range of 60–80 °C with an average error of ±1.5 °C compared to a standard measuring instrument, and maintaining an optimal temperature of 70 °C during the smoking process. Quality testing of the smoked fish indicates uniform doneness, a golden-brown color, firm texture, and an average moisture content reduction of 35%. Shelf-life testing shows that the smoked fish can last up to 7–10 days at room temperature and up to 21 days under cold storage without significant changes in aroma and texture. Unlike previous works, this study provides quantitative evidence that improved stability of multiple smoking parameters through IoT-based monitoring significantly enhances product quality consistency and extends the shelf life of smoked fish.

Implementation of YOLOv11 for Food Detection to Support Nutritional Information in Stunting Prevention

2025-11-21T08:51:16+00:00

Stunting remains a persistent public health challenge in Indonesia, mainly due to chronic malnutrition and limited parental literacy regarding balanced diets. To address this issue, this study developed an integrated nutrition education system using YOLOv11 and Generative AI, structured based on the ADDIE framework. This system aims to bridge the literacy gap by automating food identification and transforming technical nutritional data into easy-to-understand insights for stunting prevention. The study used a dataset of 2,413 images, which was expanded to 4,687 through augmentation. Technical evaluation showed strong performance with a Mean Average Precision (mAP@0.5) of 97%, ensuring reliable detection of important protein sources such as eggs. In addition to accuracy, the system applies a heuristic nutritional assessment algorithm visualized through a ‘Traffic Light’ system to reduce the cognitive load on users. Qualitative evaluation with posyandu cadres showed a significant increase in nutritional understanding, with 90% of users able to explain appropriate dietary interventions based on AI recommendations. These results conclude that the integration of computer vision with structured educational design effectively transforms mobile devices into real-time decision support systems for stunting prevention initiatives at the community level.

Facial Expression Recognition for Monitoring Learning Satisfaction in Smart Learning Environments Using MobileNetV2

2025-11-20T01:41:08+00:00

This study develops a lightweight, privacy-aware Facial Expression Recognition (FER) framework to monitor learning satisfaction in Smart Learning Environments (SLEs). Using MobileNetV2 with a two-stage training scheme on the FER2013 dataset and evaluated on 35,000 test samples, the system addresses two main questions: (1) how effectively a customized MobileNetV2 recognizes core student expressions under authentic classroom conditions, and (2) how temporal aggregation and confidence calibration improve the stability of a Learning Satisfaction Index (LSI). The model achieves 0.39 accuracy and 0.34 macro-F1, with strong performance for happy, neutral, and surprise, while challenges remain for fear–surprise and neutral–sad. Temporal smoothing reduces prediction noise and enhances the reliability of LSI signals for instructional decision-making. The findings highlight practical implications for education, particularly in supporting real-time formative assessment and improving teachers’ awareness of student engagement through privacy-preserving, on-device affect monitoring.

Hybrid Machine Learning Predictive Model for Resource Allocation Optimization and Project Risk Management

2026-01-08T16:37:48+00:00

IT project management faces critical challenges related to inaccurate resource allocation estimation and project risk assessment, which complicates decision-making and threatens project performance. Although machine learning techniques have been widely adopted in this domain, existing studies predominantly rely on single models or simple ensemble strategies, limiting their ability to capture heterogeneous interactions among organizational, technical, and risk-related factors. This study proposes a hybrid machine learning–based decision support framework that integrates feature-level representation learning and probabilistic decision fusion. Gradient Boosting is reconceptualized as a feature selection and nonlinear interaction modeling mechanism, while Artificial Neural Networks generate latent feature embeddings representing complex project characteristics. These representations are fused through a Naive Bayes classifier to produce calibrated probabilistic predictions, supported by a weighted fusion strategy with F1-score–based threshold optimization to improve stability under imbalanced risk conditions. Experimental evaluation is conducted using 5,997 synthetic IT project records from PT Anugerah Nusa Teknologi. Model performance is evaluated using accuracy, precision, recall, F1-score, and ROC-AUC metrics. Compared to standalone Gradient Boosting, Artificial Neural Network, and Naive Bayes models, the proposed hybrid framework consistently demonstrates superior predictive performance, achieving an accuracy of 0.85, an F1-score of 0.8485, and a ROC-AUC of 0.9050. Theoretically, this study contributes to project management research by demonstrating that IT project outcomes are more effectively modeled through multi-perspective learning rather than isolated predictors. Practically, the proposed framework provides actionable decision support to assist project managers in optimizing resource allocation and prioritizing risk mitigation under uncertainty.

Classification of Instagram and TikTok Addiction Levels among University Students Using the Naive Bayes Classifier

2025-11-14T10:54:17+00:00

The widespread use of gadgets and internet connectivity has become an essential aspect of daily life, especially through intensive interaction with social media platforms. Excessive usage can lead to addictive behaviors that disrupt students’ academic productivity and concentration. Although research on social media addiction continues to grow, few studies specifically examine platform-level addiction (Instagram vs. TikTok) using multi-class classification approaches. Therefore, this study aims to assess the level of social media addiction among university students, focusing on users of Instagram and TikTok at Telkom University Purwokerto. The analysis employs the Naive Bayes Classifier algorithm using data collected from 100 respondents. Model performance is evaluated through a multi-class confusion matrix to compute accuracy, precision, recall, and F1-score. Separate datasets for Instagram and TikTok are used to enable platform-specific behavioral assessment. The results show that the Naive Bayes Classifier achieves strong performance, with 93% accuracy for the Instagram dataset and 90% for the TikTok dataset. Precision scores reach 95% and 91%, recall values 93% and 90%, and F1-scores 93% and 90%, respectively. These findings confirm that Naive Bayes is effective for classifying students’ levels of social media addiction. Overall, this research contributes a reliable machine-learning–based approach for evaluating digital behavior and provides insights for early detection, enabling universities to design targeted interventions for students at risk of problematic usage. The methodology may also be extended to analyze engagement patterns on emerging social media platforms in future studies.

Academic Performance Prediction from Student–VLE Bipartite Interaction Graphs Using Centrality Features A Comparative Study with Classical Classifiers

2026-01-12T05:06:18+00:00

The rapid growth of digital learning platforms has increased the availability of student academic records and fine-grained interaction logs, creating opportunities for Educational Data Mining (EDM) to support early academic monitoring. However, many predictive models still rely mainly on individual tabular attributes and underutilize relational signals embedded in learning interactions. This study proposes a graph-mining feature approach for predicting student academic performance using a bipartite Student–VLE interaction graph. Centrality measures—degree, weighted degree, HITS hub, PageRank, and eigenvector centrality—are extracted to form a centrality feature set and combined with standard student information features. Using the public OULAD dataset, we compare three supervised classifiers: Random Forest, Support Vector Machine, and XGBoost. Experiments show that adding the centrality feature set consistently and substantially improves performance across all models compared to baseline tabular features. On the test set, XGBoost achieves the strongest results with accuracy 0.842, ROC-AUC 0.922, PR-AUC 0.902, and MCC 0.684, while Random Forest is close behind (accuracy 0.834, ROC-AUC 0.916, PR-AUC 0.894, MCC 0.672). The SVM model also benefits (accuracy 0.800, ROC-AUC 0.869, PR-AUC 0.811, MCC 0.599), confirming the robustness of the graph-derived signal. Scientifically, this study provides empirical evidence that a multi-centrality representation offers more systematic and transferable predictive value than relying on a single graph metric, across multiple classical model families under the same evaluation protocol. These findings indicate that graph-mining centrality features capture complementary structural information about learning engagement that is not represented by tabular attributes alone, and they offer a practical, interpretable enhancement to classic EDM pipelines for academic performance prediction.

User Satisfaction in Moo Opinion App: Machine Learning for Cooperative Segmentation

2025-11-26T17:28:48+00:00

This study addresses the critical need to understand digital application user satisfaction within the agricultural cooperative sector, specifically for the Moo Opinion application at the Village Unit Dairy Cooperative (KUD). The study's primary novelty lies in the implementation of an integrated, sequential Machine Learning framework—combining Random Forest (RF), Principal Component Analysis (PCA), and K-Means Clustering—to provide a granular analysis of user behavior in a specialized dairy ecosystem. The methodology first utilized RF for key feature selection, followed by PCA for dimensionality reduction, and K-Means for precise user segmentation. Primary data was collected from 40 respondents (20 farmers, 20 customers). Key findings reveal that Service Quality (0.42) and Milk Quality (0.36) are the most significant drivers of satisfaction, considerably outweighing economic factors like Milk Price (0.08). PCA identified two core satisfaction dimensions: Quality-Service Synergy (explaining 56.7% variance) and Structural-Economic Factors (explaining 25.7% variance), confirming the dominance of non-economic aspects. K-Means Clustering successfully identified three segments: Highly Satisfied (45%), Moderately Satisfied (38%), and Low Satisfaction (17%), with high cluster validity (Silhouette Coefficient 0.71). A recognized limitation of this study is the small sample size (N=40), which may affect the generalizability of the findings to larger cooperative populations. However, the results offer significant practical implications, highlighting the need for KUD to prioritize digital service quality and product value over pricing strategies to enhance loyalty and prevent churn.

Analysis of Factors Causing Toddler’s Malnutrition in Medan City Using the Random Forest Method

2025-10-05T04:28:24+00:00

Malnutrition and severe malnutrition in toddlers remain critical public health concerns that impair physical growth, cognitive development, and long-term productivity. Deficiencies in essential nutrients increase the risks of stunting, weakened immunity, and developmental delays. Although interventions such as supplementation and routine anthropometric monitoring are implemented, comprehensive identification of multidimensional causal factors is still limited, reducing the effectiveness of targeted policies. This study aims to predict toddler nutritional status using a quantitative data mining approach. A dataset consisting of 328 samples and 17 features was collected from health facilities in Medan City, including Puskesmas, the Health Office, and Posyandu. A Random Forest Classifier was developed with missing-value handling, feature engineering, and feature importance analysis to identify dominant predictors of nutritional outcomes. The model achieved an overall accuracy of 92.42 percent and showed strong performance in identifying the “Normal” class, although predictive sensitivity for minority classes such as “Gizi Kurang” and “Gizi Buruk” remained comparatively lower. Feature importance analysis indicated that complete immunization and health insurance ownership were the most influential determinants of nutritional status. This research provides a machine learning–based tool for early nutritional risk prediction and offers data-driven insights to support more precise malnutrition interventions. Future enhancement may include expanding feature diversity and applying advanced interpretability techniques to strengthen model reliability. The findings reinforce the importance of evidence-based nutrition policy strategies that prioritize early prevention and improved child health outcomes.

Towards Adaptive Learning: A Bayesian Knowledge Tracing Approach to Student Skill Prediction Bayesian Knowledge Tracing for Modeling Daily Living Skills in Children with ASD

2025-11-29T07:21:59+00:00

Autism Spectrum Disorder (ASD) presents challenges in mastering Activities of Daily Living (ADLs), which are essential for independence. This study applies Bayesian Knowledge Tracing (BKT) to model the mastery of five ADL skills—eating, dressing, toothbrushing, combing, and bathing—using data from 27 learners (1,350 responses). BKT parameters, including initial mastery, learning transition, guessing, and slipping, were used to estimate individual learning trajectories. Results showed that eating was the easiest skill (predicted mastery = 0.78), while bathing and combing were the most difficult (<0.55). The model achieved an overall accuracy of 0.62, with strong detection of actual mastery (TP = 722) but a high false-positive rate (FP = 429), indicating sensitivity to the guessing parameter. Learning curves and heatmaps revealed substantial inter-student variability. A comparative evaluation with the Performance Factors Analysis (PFA) model showed that BKT achieved higher overall predictive accuracy (BKT = 0.6356; PFA = 0.5917), while PFA demonstrated a higher AUC (0.6747) but exhibited strong positive-class bias in classification. These findings demonstrate the usefulness of BKT in modeling ADL development and highlight its potential for adaptive learning systems that support personalized interventions for ASD learners.

Comparative Analysis of XGBoost, KNN, and SVM Algorithms for Heart Disease Prediction Using SMOTE-Tomek Balancing

2025-10-24T05:47:36+00:00

Heart disease remains one of the leading causes of death worldwide, making early detection crucial for improving patient outcomes. This study aims to evaluate and compare the performance of several machine learning algorithms in detecting heart disease using the 2015 BRFSS dataset, which includes responses from 253,680 individuals. The three algorithms examined are Extreme Gradient Boosting (XGBoost), K-Nearest Neighbors (KNN), and Support Vector Machine (SVM). The data preprocessing steps involved feature encoding, class imbalance handling using the Synthetic Minority Over-sampling Technique combined with Tomek Links (SMOTE-Tomek), and hyperparameter tuning through RandomizedSearchCV. The models were assessed on a hold-out validation set using several metrics, including accuracy, Receiver Operating Characteristic-Area Under the Curve (ROC-AUC), F1-score, precision, and recall. The results demonstrated that XGBoost achieved the highest performance, with an accuracy of 94%, a ROC-AUC score of 0.98, and an F1-score of 0.94. In comparison, KNN achieved an accuracy of 87% (ROC-AUC 0.95), while SVM attained an accuracy of 79% (ROC-AUC 0.86). These findings suggest that XGBoost is a robust model for large-scale heart disease classification and holds potential for implementation in clinical decision support systems.

Line-of-Sight Dominance Over Vegetation: Simulation-Based LoRa Performance in Tropical Forest Terrain

2025-12-04T15:15:36+00:00

Low-Power Wide-Area Network (LPWAN) technologies, especially LoRa, are receiving considerable interest for applications involving environmental monitoring in difficult terrain conditions. However, existing research predominantly examines vegetation attenuation or terrain elevation effects separately, leaving a critical research gap in understanding their combined and interactive impacts on LoRa connectivity in tropical forest environments. Furthermore, most studies rely on simplified propagation models that inadequately represent the complex radio environment of tropical forests, and few investigations systematically compare the relative importance of vegetation density, elevation, and line-of-sight conditions. This work addresses these gaps through an in-depth simulation-based investigation of LoRa network behavior in the University of Brawijaya (UB) Forest, which serves as a typical tropical forest setting in Indonesia. We performed detailed simulations using Python and LoRaSim, employing fine-resolution elevation datasets and precise vegetation classification to examine how dense vegetation, medium vegetation, and elevation parameters influence LoRa communication performance. Our findings indicate that, in contrast to traditional propagation models, nodes located in dense vegetation zones reached a 90.0% success rate, as opposed to 65.0% in zones without vegetation. Additional investigation shows that line-of-sight presence (28.6% versus 0.0% success rate) and relative elevation relative to the gateway (11.1% versus 27.3% success rate for nodes positioned above and below the gateway, respectively) represent more crucial factors for connectivity compared to vegetation attenuation by itself. These outcomes offer important guidance for enhancing LoRa-based environmental monitoring systems in tropical forest settings through strategic node positioning that considers elevation characteristics and line-of-sight availability.

Tourism Destination Recommendation Using Blockchain Technology and MCDM Approach

2025-11-16T17:32:34+00:00

The rapid advancement of digital tourism services has revolutionized how travelers search and select destinations, yet privacy and trust issues remain major challenges in centralized recommendation systems. User data such as preferences, location history, and feedback are often stored on centralized servers, making them vulnerable to data breaches and manipulation. This research proposes a Blockchain-Driven Multi-Criteria Decision Making (MCDM) Approach to develop a privacy-preserving and trustworthy tourist recommendation system. The proposed framework integrates blockchain technology to ensure secure, transparent, and immutable data management, while MCDM techniques such as the Analytic Hierarchy Process (AHP) and TOPSIS are employed to evaluate and rank tourist destinations based on multiple criteria, including popularity, cost, safety, accessibility, and sustainability. The blockchain layer enforces decentralized data verification through smart contracts and cryptographic consensus, ensuring that user privacy is protected without sacrificing system transparency. The experimental results indicate improved recommendation accuracy, reduced privacy risks, and enhanced user trust compared to conventional systems. The proposed model achieved 12.5% higher recommendation accuracy and 30% lower privacy risk compared to centralized models. This study demonstrates that combining blockchain and MCDM can effectively support transparent and fair decision-making in digital tourism, offering a scalable and secure foundation for next-generation recommendation systems.

Emotion-Based Multi-Class Sentiment Analysis Of FirstMedia Customers Reviews Using SVM With Kernel Comparison

2025-12-06T05:06:20+00:00

The advancement of digital technology has made users increasingly reliant on online services, with user reviews serving as an essential resource for evaluating the quality of service provided by companies such as FirstMedia. However, these valuable data have not undergone comprehensive analysis to assess users’ emotional responses. This study aims to classify FirstMedia customers’ emotions into four categories (joy, sadness, anger, and neutral) and to evaluate the Support Vector Machine (SVM) method using four different kernel functions. Most existing studies primarily focus on polarity-based sentiment analysis and do not explicitly examine multi-emotion classification or kernel comparison in machine learning models. A total of 4,001 reviews were collected through web scraping from the Google Play Store and the X app and processed through several preprocessing steps. Emotion classification was conducted using the NRC Indonesian Emotion Lexicon, while word significance was determined using TF-IDF weighting. After preprocessing, 3,069 labeled reviews were retained and distributed as 1,065 neutral, 748 anger, 692 joy, and 564 sadness reviews, which were used for emotion classification. Model performance was evaluated using a hold-out validation scheme with an 80:20 train-test split and assessed through a confusion matrix. To address class imbalance, undersampling was applied, resulting in a balanced dataset for model training. The evaluation results show that the Linear kernel achieved the highest performance, with an accuracy of 82.63%, precision of 82.86%, recall of 82.63%, and an F1-score of 82.60%, outperforming the Gaussian, Polynomial, and Sigmoid kernels. This study demonstrates that multi-emotion sentiment analysis provides a more comprehensive understanding of user perceptions beyond conventional sentiment polarity, thereby supporting more informed evaluations of digital service quality.

Music-Structure Segmentation in Balinese Gamelan (Tabuh Lelambatan) with SSM, Checkerboard Novelty, and HMM

2025-11-05T16:06:23+00:00

This study aims to automatically segment the musical structure of Balinese gamelan by combining the Self-Similarity Matrix (SSM) method, the Checkerboard Novelty kernel, and Hidden Markov Models (HMM). Balinese gamelan has a complex musical structure that is cyclical and based on a colotomik system, requiring an adaptive analytical approach to repetitive patterns and transitions between musical sections. The research data consists of 30 Tabuh Lelambatan gamelan audio recordings obtained from public digital sources and validated through expert annotation to produce ground truth. The segmentation process was carried out through feature extraction using Constant-Q Transform (CQT), SSM formation to detect acoustic similarity patterns, application of the checkerboard kernel to mark transitions between segments, and temporal sequence modeling using HMM to refine boundary detection. System performance evaluation was carried out by comparing the segmentation results with ground truth using precision, recall, and F1-score metrics. The test results showed an average macro precision value of 0.998, a recall of 0.705, and an F1-score of 0.818, indicating that this method is capable of detecting the main boundaries of musical structures with high accuracy and consistent stability. However, the model still tends to miss gradual micro transitions. This research contributes to the field of Music Information Retrieval (MIR) and supports efforts to preserve traditional Balinese music through data-based analysis and the development of music computing technology.

Optimizing URL-Based Phishing Detection Using XGBoost and Relief Feature Selection

2025-12-05T03:45:37+00:00

Phishing is a significant cybersecurity threat in which attackers exploit manipulated URLs to deceive users and obtain confidential information. As phishing attacks continue to grow in complexity, automated machine learning based detection methods have become essential to strengthen digital security. This study proposes a URL based phishing detection model using boosting algorithms while analyzing the role of feature selection in improving classification performance and computational efficiency. The experiments were conducted on a dataset consisting of 10000 instances with 50 features and balanced class labels. After data preparation, 48 features were retained as input variables, and min max normalization was applied to ensure uniform feature scaling. Three boosting algorithms namely Gradient Boosting, XGBoost, and AdaBoost were evaluated using accuracy, precision, recall, and F1 score. Among these methods, XGBoost achieved the highest accuracy of 98.8 percent, demonstrating its effectiveness in learning complex URL patterns. Subsequently, three feature selection techniques namely Information Gain, Chi Square, and ReliefF were applied and evaluated using 10 fold cross validation. The results indicate that ReliefF provides the most effective feature reduction by selecting 37 features while maintaining the same classification accuracy. Unlike previous studies that mainly focus on classifier comparison, this study demonstrates that integrating XGBoost with ReliefF enables significant feature dimensionality reduction without compromising predictive accuracy. This finding highlights an efficient trade off between detection performance and computational complexity. Overall, the proposed framework offers a robust, efficient, and scalable solution for fast and adaptive phishing detection in modern cybersecurity environments.

Optimization of Web-Based Printing Order Management System Using Redis Database for Efficient Data Handling

2025-11-07T03:03:17+00:00

The rapid advancement of information technology has encouraged small and medium-sized enterprises to shift from manual operational procedures to structured digital systems. However, many small printing businesses continue to face delays, data inconsistencies, and limited real-time monitoring due to conventional order management practices. These challenges highlight the need for a more responsive and efficient ordering system capable of improving transaction accuracy and service delivery speed. This study addresses the issue by developing a web-based ordering system using an iterative Agile Scrum approach, followed by a comprehensive performance evaluation through simulated concurrent user testing. The results show a substantial improvement in system responsiveness, with user data retrieval time decreasing from 11,228 ms to 2,148 ms (an 80.9% improvement) and order processing time reduced from 16,954 ms to 4,697 ms (a 72.3% improvement), resulting in an overall average efficiency gain of 76.6%. The integration of Redis caching significantly enhances system performance, stability, and load distribution, addressing the current gap in Redis implementation for small-scale printing environments. This study demonstrates that adopting a hybrid data-handling architecture can provide a scalable, reliable, and high-performance solution for digital ordering processes, enabling small enterprises to improve operational efficiency and customer satisfaction.

An OWL-Based Ontology Model of Food Production and Distribution in Indonesian

2025-12-05T03:48:05+00:00

Food security in Indonesia is influenced by the dynamics of production, distribution, and availability between regions. However, many existing information systems still rely on conventional data structures without semantic integration, which limits interoperability and hinders interregional analysis. To address this gap, this study developed an ontology model based on the Web Ontology Language (OWL) that formally represents the relationships between food production, commodity characteristics, distribution flows, food insecurity conditions, and geographical context. The ontology was built using Protégé through stages of literature review, official data collection from BPS, FAO, and the Ministry of Agriculture, conceptual model design, implementation, and evaluation. Conceptual validation was conducted through Focus Group Discussions (FGD) with food supply chain experts to ensure the suitability of the ontology structure and the actual conditions of the national food system. The technical evaluation involved consistency testing using the Pellet reasoner and Competency Question (CQ) testing through SPARQL queries to assess the ontology's ability to respond to essential information needs. The resulting ontology consists of five core classes (FoodProduction, FoodItem, FoodDistribution, FoodSecurityStatus, and GeographicRegion) which collectively represent the semantic structure of Indonesia's food supply chain. The evaluation results show that the ontology is structurally consistent and capable of producing outputs that are in line with CQ, including the retrieval of production-distribution information and the initial identification of commodity surpluses and deficits based on instance data. These findings indicate that the developed ontology provides a coherent semantic foundation for modeling food systems and has strong potential to support the development of knowledge-based food security management applications.

Comparison of XGBoost and Naive Bayes Models in Type 2 Diabetes Prediction with RFE Feature Selection

2025-11-11T17:11:37+00:00

Type 2 diabetes mellitus is a chronic disease with an increasing prevalence rate that can cause serious complications if not detected early. The application of machine learning algorithms can aid prediction, but selecting the right model and features greatly determines the accuracy of the results. This study aims to compare the performance of the Extreme Gradient Boosting (XGBoost) and Naive Bayes algorithms in predicting type 2 diabetes with and without Recursive Feature Elimination (RFE) feature selection. The data used were from the UCI Machine Learning Repository, comprising 768 samples and eight clinical features. The research process included data preprocessing, dividing the data into 614 training data and 154 testing data, applying RFE to select the most influential features, model training, and evaluation using accuracy, precision, recall, F1-score, and AUC. The results show that Naive Bayes without RFE achieves 70.77% accuracy, 0.57377 precision, 0.648148 recall, F1-score 0.608696, and 0.772778 AUC, while Naive Bayes with RFE increases the accuracy to 74.02% and the AUC to 0.793333. Meanwhile, XGBoost with RFE provided the best results with an accuracy of 74.67%, precision of 0.653061, recall of 0.592593, F1-score of 0.621359, and the highest AUC of 0.804259. Besides, applying RFE also improves the computational efficiency. These findings indicate that applying RFE significantly improves classification and computation time performance. The practical implication is that this model could aid early detection of diabetes in clinical settings. Further research can be conducted by optimizing parameters and using more diverse datasets.

IoT-Based Stress Monitoring Using CNN for HRV-GSR Analysis

2025-12-05T08:50:38+00:00

Stress has become a major global health concern affecting both physical and mental well-being. Conventional stress assessment methods rely on subjective self-reports that cannot capture real-time physiological changes. Existing systems are often limited to controlled laboratory environments or depend on traditional machine learning approaches requiring extensive manual feature engineering.

This study aims to develop an Internet of Things–based stress monitoring system using deep learning to enable objective, continuous, and practical real-world stress detection. The system incorporates wearable sensors using an ESP32-DevKit V1 microcontroller, a MAX30102 photoplethysmography sensor, and a Grove-GSR module for real-time acquisition of Heart Rate Variability and Galvanic Skin Response signals. A dual-branch Convolutional Neural Network architecture processes preprocessed HRV and GSR time-series data to automatically learn discriminative features without manual feature engineering. Data were collected from 30 participants, resulting in 8,100 labeled samples across four stress levels. The proposed CNN model achieved 91.3% classification accuracy, outperforming baseline machine learning models such as Support Vector Machine (78.4%), Random Forest (81.7%), and XGBoost (84.3%). Real-time system evaluation demonstrated an average latency of 1.47 seconds and battery endurance exceeding 13 hours, confirming the feasibility of continuous wearable stress monitoring. The integration of IoT infrastructure with deep learning provides an effective framework for physiological stress assessment, offering potential applications in preventive healthcare, workplace health management, and personalized mental-wellness monitoring.

Causal Analysis of Stunting Determinants Using the Peter-Clark and Greedy Equivalence Search Algorithms

2025-11-02T14:30:12+00:00

Child stunting remains a major public health challenge, reflecting the long-term effects of inadequate nutrition, limited maternal education, and restricted access to health services. However, most existing studies rely on correlational analysis, leaving the underlying causal mechanisms insufficiently understood. This gap limits the development of effective interventions, as policymakers require evidence on how determinants interact causally. To address this issue, this study applies two causal discovery algorithms Greedy Equivalence Search (GES) and Peter Clark (PC) to identify causal relationships among eight key determinants of stunting using secondary data from the West Bangka District Health Office (2024). The variables include anthropometric indicators, maternal characteristics, and environmental conditions. Causal assumptions such as causal sufficiency, acyclicity, and faithfulness were imposed to ensure identifiability of the resulting graphs. Model performance was evaluated using Directed Density (DD) and Causal Density (CD) metrics. GES generated a parsimonious causal structure highlighting maternal education, posyandu visits, and exclusive breastfeeding as dominant causal candidates affecting height-for-age (TB/U) and weight-for-age (BB/U). In contrast, the PC algorithm produced a more complete and dense structure, achieving DD = 1.0 and CD = 0.12, compared with GES (DD = 0.80; CD = 0.10). These results indicate that PC is more exploratory in mapping complex causal interactions, while GES offers a simpler and more conservative model. Collectively, the findings demonstrate that combining score-based and constraint-based discovery approaches yields complementary insights into the mechanisms driving stunting.

Adoption of Artificial Intelligence in Vocational High Schools: A Systematic Review of Teachers’ Perspectives

2025-12-14T05:06:49+00:00

The purpose of this study is to examine Vocational High School teachers' perceptions of the adoption of artificial intelligence (AI) in vocational education. Using a systematic literature review (SLR) approach with the PRISMA 2020 protocol, 1,346 articles were identified from six main databases: ACM, IEEE Xplore, ScienceDirect, Google Scholar, and Semantic Scholar. Of the 1,346 articles identified, 263 were excluded at an early stage because they did not meet basic criteria, such as authorship, publication language, and document type. A total of 1,083 articles were screened, yielding 208 reports for in-depth screening. After the screening and feasibility assessment, 29 studies were included in the final analysis. The analysis showed that 72% of studies reported positive results, 24% reported moderate results, and 3% reported exploratory results. The dominant factors influencing teachers' perceptions included infrastructure readiness, digital competence, institutional support, the relevance of the vocational curriculum, and ethical and privacy issues. These findings emphasize the need for a holistic strategy for implementing AI in vocational schools that encompasses teacher training, education policy, and ethical considerations.

Comparing XGBoost and LightGBM for Optimizing Health Content Categories

2025-11-10T18:01:50+00:00

Indonesia’s social media platforms contain large amounts of unverified health information. Research on Indonesian health-text mining still rarely focuses on disease-based classification, leaving a gap compared with studies that only address sentiment or general topic categorization. This study proposes a multi-class classification approach that uses IndoBERT embeddings combined with gradient-boosting classifiers (XGBoost and LightGBM) to categorize tweets into diabetes, hypertension, and heart disease. The dataset comprises 4,075 tweets collected from platform X (Twitter). Preprocessing involves text cleaning, anonymization, normalization, and the extraction of 768-dimensional IndoBERT embeddings. Experiments are conducted in Google Colab (Intel Xeon CPU, 13 GB RAM, optional NVIDIA T4 GPU) using stratified five-fold cross-validation.The best results are obtained by the IndoBERT × LightGBM pipeline, which achieves an accuracy of 0.8526 and a macro-averaged F1-score of 0.8527, outperforming the IndoBERT × XGBoost model (accuracy 0.8325 and macro F1-score 0.8326). Feature-importance analysis shows that contextual terms related to blood sugar, the heart, and blood pressure strongly influence the predictions. Overall, the proposed method provides an effective baseline for monitoring health-related text and supporting disease-oriented analytics in Indonesian-language social media.

A Systematic Literature Review of Machine Learning for Endurance Running Performance Prediction

2025-12-22T13:31:30+00:00

This study systematically reviews the application of machine learning methods for predicting running performance, with particular emphasis on short-middle distance events such as the 5 km. Although machine learning based performance prediction has been widely explored in endurance sports, a comprehensive review synthesizing models, predictors, and pipelines across running distances remains limited. The review followed the PRISMA 2020 framework. Articles published between 2020 and 2025 were retrieved from ScienceDirect, Google Scholar, and PubMed using predefined keyword combinations related to machine learning and running performance. Studies were included if they focused on running (excluding cycling, triathlon, or other sports), applied predictive modeling, and reported model evaluation metrics. A total of 26 studies met the inclusion criteria and were assessed using quality appraisal criteria inspired by TRIPOD and QUADAS-2. The analysis identified four main research themes: (1) application of machine learning models for running performance prediction, (2) physiological and anthropometric predictors, (3) non-physiological and contextual factors, and (4) personalized athlete training and monitoring. Ensemble learning models (Random Forest, XGBoost, LightGBM) consistently outperformed traditional linear regression by capturing non-linear interactions, while deep learning approaches (LSTM, GRU) demonstrated strong capability in modeling temporal training dynamics. A generalized machine learning pipeline for running performance prediction was also synthesized. This review contributes a structured framework that integrates modeling approaches, predictor categories, and evaluation strategies, and highlights research opportunities for explainable and personalized prediction systems, particularly for 5 km running performance.

Integrating Agile Development and Content-Based Filtering for Personalized Digital Cultural Heritage Applications: A Case Study of Sri Ranggah Rajasa Sang Amurwabhumi

2025-11-25T16:39:24+00:00

The preservation of Indonesia’s cultural heritage increasingly requires digital innovation that not only archives historical material but also engages users through adaptive interaction. However, existing digital cultural platforms seldom provide personalized learning experiences and often lack iterative user-centered development, creating a clear gap in adaptive digital cultural heritage applications. This study aims to design and develop a cultural application titled Sri Ranggah Rajasa Sang Amurwabhumi using a hybrid framework that integrates the Agile Development Method with a Content-Based Filtering (CBF) approach. Agile was applied through iterative cycles of design, development, implementation, integration, and testing, enabling continuous enhancement based on user feedback. Meanwhile, the CBF algorithm was used to generate personalized cultural content recommendations by analyzing semantic similarities among historical items. The novelty of this research lies in the unified hybridization of Agile and CBF to support adaptive, personalized digital cultural learning centered on a specific Indonesian cultural figure. Data were gathered from 30 respondents, including students and cultural practitioners, through usability testing and structured questionnaires. Results indicate high performance across key aspects: functionality (91%), usability (90%), recommendation accuracy (88%), and user satisfaction (93%). These findings demonstrate that combining Agile and CBF strengthens technical reliability while improving engagement through adaptive content delivery. Agile supports iterative refinement of user interfaces and system responsiveness, whereas CBF enables intelligent personalization in cultural learning environments. Nevertheless, this study is limited by its modest sample size and its focus on a single cultural topic, which may reduce generalizability. Future work will expand the dataset, incorporate multimodal cultural content, and validate the hybrid framework across broader Indonesian cultural domains..

A Hybrid YOLOv11 and LightFM Model for Emotion-Driven Anime Recommendation

2025-11-13T05:36:58+00:00

Existing anime recommendation systems focus on genre preferences and viewing history without considering users' emotional states, leading to context-blind recommendations that may exacerbate negative moods and reduce satisfaction. Most existing systems employ outdated architectures with limited accuracy and lack diversification mechanisms to prevent filter bubbles. This study develops an emotion-based anime recommendation system integrating YOLOv11 for facial emotion recognition with hybrid collaborative filtering using LightFM and Maximum Marginal Relevance diversification. The primary novelty lies in seamlessly combining YOLOv11's superior emotion recognition, LightFM's hybrid matrix factorization for cold-start mitigation, and MMR diversification for preventing filter bubbles while maintaining emotional congruence. The methodology employed the KDEF dataset (3,597 images, five emotion classes) for training YOLOv11 with data augmentation, and the MyAnimeList dataset (744,330 interactions) for recommendation modeling. Emotion-to-genre mappings informed by survey data from 51 participants were implemented with MMR diversification to balance relevance and variety. The YOLOv11 model achieved 93.70% validation accuracy, outperforming CNN-LSTM approaches by 37.55 percentage points. The hybrid recommendation model demonstrated test AUC of 0.8567 and Precision@10 of 0.1457, representing 417% improvement over pure collaborative filtering, while diversification increased genre representation by 20.9% with minimal precision loss. This system demonstrates real-time applicability for streaming platforms through camera-based emotion capture and immediate recommendation generation, enhancing user engagement and emotional well-being. The integration represents a significant advancement toward affective computing in entertainment media.

Workload-Aware Performance Evaluation of Sequential and Parallel DAG-Based Machine Learning Orchestration on Single-Node Systems

2026-01-14T04:36:35+00:00

The increasing adoption of machine learning in production systems has intensified the need for structured, automated, and reproducible pipelines, commonly modeled as directed acyclic graphs. However, unlike most existing workflow scheduling studies that focus on distributed or multi-node environments, this work addresses the lack of controlled, workload-aware analysis for orchestration strategies specifically in single-node systems. A controlled experimental methodology is applied by executing an identical machine learning pipeline under two orchestration modes: sequential execution using a local orchestrator and parallel execution using a workflow orchestration engine. Two scenarios are evaluated by explicitly controlling task execution duration to represent light and heavy computational workloads. The results show that under light workloads, parallel execution increases the average makespan significantly, yielding a speedup of only 0.71, which indicates performance degradation due to dominant orchestration overhead. In contrast, under heavy workloads, parallel execution reduces the average makespan from 1013 seconds to 532 seconds, achieving a speedup of 1.90. System-level monitoring reveals higher central processing unit utilization during parallel execution, while evaluation metrics, including root mean square error and coefficient of determination, remain stable across all experimental runs. This study contributes empirical evidence of a workload threshold beyond which orchestration overhead becomes negligible and pipeline parallelism becomes beneficial. These findings demonstrate that the performance benefits of parallelism are strictly workload-dependent, highlighting the importance of selecting orchestration strategies based on computational workload characteristics.

Fuzzy Time Series Chen Model for Dual-Commodity Agricultural Forecasting: Evidence from Indonesia’s Rice and Corn Production

2025-12-05T06:37:41+00:00

Indonesia's strategic food commodities, particularly rice and corn, exhibit strong seasonal fluctuations and irregular production shocks driven by climate anomalies and policy changes, generating nonlinear time-series patterns that conventional statistical models often fail to capture. This study evaluates the forecasting capability of the standard Chen Fuzzy Time Series (FTS) model for dual-commodity agricultural data under varying seasonal and anomaly conditions. Monthly production data from January 2021 to March 2025 from the Indonesian Central Bureau of Statistics (BPS) were processed through a complete FTS pipeline: universe-of-discourse construction, triangular membership function design, fuzzification, FLR and FLRG formation, and midpoint-based defuzzification. Forecast accuracy was assessed using MAE, MSE, RMSE, MAPE, and R², with residual distribution analysis, Shapiro-Wilk tests, and scatter plots conducted to validate model stability. The model achieved high precision with overall MAPE of 4.37% for rice and 8.12% for corn, both classified as Highly Accurate. Monthly accuracy revealed consistent stability during May-December, while transitional months (January-March) showed greater variability due to extreme anomalies such as the January 2024 production collapse. Residual analysis confirmed near-normal error distribution for rice (p = 0.062) and mild deviation for corn (p = 0.031), while scatter plots demonstrated strong linear relationships (Rice R² = 0.9876; Corn R² = 0.9654). The findings establish Chen's FTS as a transparent and operationally reliable baseline method for national food production forecasting, although its sensitivity to structural breaks highlights the need for future hybridization with climate and policy indicators.

Experimental Characterization of ESP-Mesh Performance for Resilient Medical IoT Monitoring Systems

2026-01-17T01:35:46+00:00

The reliability of medical Internet of Things (IoT) systems is critically dependent on network resilience, particularly in indoor hospital environments where conventional Wi-Fi infrastructures are vulnerable to single points of failure. Although ESP-Mesh has emerged as a promising self-healing communication protocol, its performance characteristics in medical IoT monitoring contexts remain insufficiently explored. This study aims to experimentally characterize the performance of ESP-Mesh networks for resilient medical IoT monitoring systems by analyzing multi-hop latency behavior, signal degradation, and communication stability under indoor medical-like conditions. A multi-parameter monitoring prototype integrating infusion volume, drip rate, and heart rate sensors was deployed as an experimental platform. Network performance was evaluated through controlled measurements of RSSI, end-to-end latency, and self-healing behavior, while MQTT was employed to assess cloud-based transmission efficiency. The results demonstrate that ESP-Mesh maintains stable self-healing communication with an average multi-hop latency of 0.714 s across distances up to 5 m, with latency increasing consistently as RSSI decreases. MQTT cloud transmission achieved a lower average latency of 0.247 s with zero packet loss, confirming its suitability for lightweight medical data delivery. Sensor evaluation revealed high accuracy for infusion volume monitoring (95.42%), while heart rate and drip rate measurements exhibited lower reliability due to signal interference and environmental sensitivity. These findings provide empirical insights into the performance limits and trade-offs of ESP-Mesh networks in medical IoT environments. The study confirms the feasibility of ESP-Mesh as a resilient communication backbone for medical monitoring, while highlighting the necessity of advanced signal processing to achieve clinical-grade sensing reliability.

Digital Transformation of Toddler Posyandu Services via an Android-Based Application

2025-09-23T12:04:59+00:00

Abstract: Posyandu Balita as a community-based health service holds an essential role in improving maternal and child health in Indonesia. Nevertheless, the dependency on manual documentation frequently causes delays in reporting immunization, incomplete records, and limited access for parents to monitor child growth. This study sought to design and assess an Android-based Posyandu Balita application by applying a Research and Development (R&D) model combined with the System Development Life Cycle (SDLC) approach. The development process covered several phases: needs analysis, system design, application construction, pilot implementation, and evaluation through the Technology Acceptance Model (TAM).

The pilot, which involved 10 health cadres and 10 parents, revealed that the application reduced data loss, facilitated more accurate immunization tracking, and encouraged stronger parental involvement. Functional testing indicated that the main features—digital medical records, reminder notifications, and growth chart visualization—worked consistently as intended. Based on TAM analysis, perceived usefulness (PU) and perceived ease of use (PEOU) significantly shaped users’ behavioral intention to utilize the system (PU = 62%, PEOU = 58%). Moreover, the level of parental compliance in child health monitoring increased, where 85% of parents actively accessed the digital platform compared to only 40% before the trial.

Overall, the results demonstrate that mobile health applications developed with user-centered approaches can improve the effectiveness and efficiency of community-based services. The Posyandu Balita application is a promising innovation to support Indonesia’s digital health transformation. Further research is required to examine large-scale implementation, integration with national health information systems, and strategies for long-term sustainability.

Keywords: Community Health, Toddler Posyandu, Android-based Application, Mobile Health, Technology Acceptance Model, Digital Innovation

Adaptive Learning System Based on Human-in-the-Loop for PDF Template Data Extraction

2025-11-26T16:54:34+00:00

PDF template data extraction remains a substantial challenge due to semi-structured document formats and variations. While large pre-trained models achieve high accuracy, they require extensive computational resources and labeled datasets, making them impractical for resource-constrained environments. Conversely, rule-based approaches are efficient but rigid. This research addresses this gap by developing an adaptive learning system that integrates rule-based approaches with Conditional Random Fields (CRF) in a hybrid framework, designed for data-scarce scenarios. The system implements parallel extraction strategies with confidence-based selection and Human-in-the-Loop (HITL) feedback for incremental learning. Pattern learning updates rule-based strategies, while CRF models are retrained incrementally. Evaluated on synthetically generated documents across diverse template types, the system achieves 98.61% accuracy with minimal training data and 7% user correction rate, demonstrating high learning efficiency (1.88 corrections per percentage point). The improvement is statistically significant (paired t-test, p < 0.001, Cohen’s d = 8.95). The system operates on CPU-only hardware with 50-100 MB footprint and 0.1-0.5 seconds processing time. This work fills a practical gap in document extraction, providing a middle-ground solution balancing high accuracy, minimal data requirements, low resource consumption, and real-time adaptability—suitable for small organizations and rapid deployment where large models are impractical. The evaluation uses synthetic data to ensure reproducibility and controlled assessment, though real-world validation would strengthen practical applicability.

Implementation of YOLOv12 and PaddleOCR for Indonesian Bank Statement Table Extraction

2025-10-05T04:29:25+00:00

The increasing reliance on digital financial documents has highlighted the need for automated methods to extract structured information from bank statements. Traditional optical character recognition (OCR) systems often fail to capture complex tabular structures, leading to incomplete or error-prone transaction records. To address this challenge, this research proposes a two-stage detection and recognition pipeline that combines YOLOv12 for table and structural element detection with PaddleOCR for text extraction, followed by automated Excel conversion. The objective of this study is to improve accuracy in localizing tables, detecting rows and columns, and generating structured financial data that can be directly utilized for downstream applications. The methods involve training a YOLOv12-n model in two stages: Stage 1 focuses on detecting entire table regions, while Stage 2 focuses on identifying row and column structures within the detected tables. A lightweight AdamW optimizer with conservative augmentation strategies was applied to preserve the geometric integrity of document layouts. Results show that Stage 1 achieved precision of 0.998, recall of 1.0, and mAP50-95 of 0.989, while Stage 2 achieved precision of 0.992, recall of 0.964, and mAP50-95 of 0.899, demonstrating strong localization and structural recognition. The conclusions confirm that the proposed two-stage pipeline is effective for financial document processing, with potential applications in digital banking, auditing, and automated record management. Future research may focus on expanding datasets and addressing domain-specific variability.

Sarcasm Detection in Indonesian YouTube Comments using Fine-Tuned IndoBERT with Class Imbalance Handling

2025-11-26T16:38:58+00:00

Sarcasm detection in Indonesian social media faces challenges in natural language processing due to implicit meanings and limited labeled datasets. YouTube, with 143 million users in Indonesia, represents a largely unexplored source of sarcastic expressions. This study aims to develop an automatic sarcasm detection system for Indonesian YouTube comments using fine-tuned IndoBERT and evaluate the performance of two IndoBERT variants. A dataset of 5,291 YouTube comments was collected and automatically labeled using GPT-4o with structured prompts based on linguistic indicators of sarcasm. Two IndoBERT variants (IndoNLU and IndoLEM) were fine-tuned with three class imbalance mitigation strategies: imbalanced, under-sampling, and class weighting. Zero-shot evaluation was conducted as a baseline to measure fine-tuning effectiveness. Models were evaluated using accuracy, precision, recall, and F1-score metrics. Pre-trained models without fine-tuning showed very limited sarcasm detection capability with F1-scores of 0.1613 for IndoNLU and 0.3519 for IndoLEM. Fine-tuning with under-sampling dramatically improved F1-scores to 0.6499 for IndoNLU and 0.6568 for IndoLEM, showing improvements up to 303%. IndoBERT-IndoNLU provided more balanced performance with 0.6424 accuracy, while IndoLEM showed higher sarcasm recall of 0.7639. Fine-tuning IndoBERT is effective for detecting sarcasm in Indonesian YouTube comments. This study contributes by providing a new labeled dataset, demonstrating the effectiveness of automatic labeling using large language models, and providing empirical evidence of the significant value of fine-tuning for Indonesian sarcasm detection.

Comparison of IndoBERT and SVM Performance in Sentiment Analysis of Digital Education Platforms

2025-11-02T17:04:11+00:00

Sentiment analysis on user-generated reviews is essential for understanding the quality and effectiveness of digital education platforms. This study compares the performance of Support Vector Machine (SVM) and IndoBERT in classifying sentiments from Ruangguru user reviews. The original dataset contains 111,838 reviews, from which a stratified sample of 10,000 entries was selected for experimentation to maintain class proportion. Text preprocessing applied standard/light normalization (case folding and light cleaning, handling URLs/users/hashtags and repetition) without stopword removal to preserve polarity cues. Auto labels are validated on 139 manually annotated samples (accuracy 0.763, Cohen’s κ 0.644), indicating reliable yet imperfect alignment. To ensure a fair, leakage-safe comparison, we use a fixed 20% standard test split for all models; within the remaining data, 10% is used for validation, and IndoBERT checkpoints are selected based on validation macro-F1 (early stopping). The SVM baseline combines word- and character-level TF-IDF with class-balanced LinearSVC and grid search, achieving accuracy 0.888 and macro-F1 0.543, strong on positives but limited for the neutral class. IndoBERT yields more balanced performance: the class-weighted variant attains the best macro-F1 0.601 (accuracy 0.857), while the baseline reaches the highest IndoBERT accuracy (0.867) with macro-F1 0.596. These results show that Transformer models provide a more balanced trade-off under severe imbalance, whereas SVM remains a competitive accuracy-oriented baseline. In practice, platforms should prioritize macro-F1, use optimized IndoBERT when minority opinions matter, and invest in expanded manual labeling and advanced imbalance handling to improve neutral detection further.

Improving Machine-Learning Malware Detection Through IQR-Based Feature Reduction

2025-11-29T07:50:58+00:00

Malware detection is a significant challenge in cybersecurity due to the complex and evolving nature of threats. This study evaluates the effectiveness of machine learning algorithms, specifically XGBoost and LightGBM, in detecting malware. The approach includes data cleaning, normalization, feature selection, and the use of the Interquartile Range (IQR) technique to select relevant features. The initial dataset contained 21,752 files, evenly split between malicious and benign files. After data cleaning, the number of samples decreased to 19,256 files, with numerous features that were reduced after applying IQR. Results show that XGBoost outperforms other algorithms, achieving 99.20% accuracy, an improvement over the 98.99% accuracy without IQR. The IQR technique enhances data quality by filtering out features with significant differences between malware and benign files, improving model performance. Additionally, reducing the feature set helps prevent overfitting and strengthens the model's generalization ability. The study concludes that machine learning, particularly with algorithms like XGBoost and LightGBM, can effectively improve malware detection. By using IQR in feature selection, model performance is enhanced, leading to reduced false positives and increased detection efficiency. The research highlights the importance of feature selection techniques like IQR in boosting the predictive power of machine learning models, making them more efficient in identifying malware. Future work will explore additional feature selection methods to further improve malware detection accuracy.

Blockchain and SVM Integration for Distributed DDoS Attack Detection

2025-11-02T14:06:16+00:00

Rapid developments in information technology have increased dependence on network services, but have also triggered an increase in cyber threats such as Distributed Denial of Service (DDoS). These attacks can paralyze systems by flooding servers with simultaneous fake traffic. Conventional rule-based detection methods are now less effective in dealing with dynamic attack patterns, requiring an adaptive approach based on machine learning. This research develops a Support Vector Machine (SVM) model enhanced with Blockchain technology to improve accuracy and data security in detecting DDoS attacks. The dataset used is CICDDoS2023 from the Canadian Institute for Cybersecurity, which contains various variants of modern DDoS attacks. The research stages include data pre-processing, training the SVM model using the RBF kernel, and integrating Blockchain with training data hash recording through a smart contract using Remix Ethereum to ensure data integrity. Performance evaluation was carried out using accuracy, precision, recall, and F1-score metrics based on the confusion matrix results. The integration of SVM and Blockchain showed an increase in security and detection accuracy compared to conventional SVM models. This approach not only improves the reliability of the DDoS attack detection system, but also creates a transparent and tamper-proof data validation mechanism. The research results are expected to contribute to the development of adaptive, decentralized network security systems with a high level of confidence in attack detection results.

Comparative Evaluation of YOLOv8 and YOLOv11 for Student Behavior Detection in Classroom CCTV Environments

2025-12-01T16:51:15+00:00

Monitoring student behavior during classroom learning is important for supporting learning quality and teacher performance. This study presents a pilot comparison between YOLOv8 and YOLOv11 for detecting student classroom behaviors from CCTV images. Six elementary behaviors are consistently defined and used throughout the work: lookup, raise-hand, read, stand, turn-head, and write. The available SCB dataset contains 4,934 labeled images, but this study deliberately uses a front-facing subset of 100 images that best represent clear posture and behavior. After augmentation, the dataset grows to 220 images, split into 180 training, 30 validation, and 10 testing images. Both models are trained for 25 epochs on a T4 GPU with comparable configurations. At the detector level, YOLOv11 achieves higher mean average precision (mAP) of 42.9% compared to 28.9% for YOLOv8. At the behavior level, overall classification accuracy on the test set is 43.3% for YOLOv8 and 37.5% for YOLOv11. These results indicate a trade-off: YOLOv11 provides stronger bounding-box detection performance, while YOLOv8 produces slightly more stable behavior-level predictions on this very small and imbalanced dataset. The study emphasizes that these findings are exploratory baselines rather than definitive benchmarks, because the dataset is small and no statistical significance testing is performed. Future work must use a larger portion of the SCB dataset, more balanced class distributions, repeated experiments, and statistical analysis to obtain more robust conclusion.

Evaluation of Machine Learning Algorithm for Automatic Assessment of School Students' English Essay

2025-11-02T16:30:12+00:00

The manual assessment of essays in English language learning often faces challenges related to objectivity and efficiency, especially on a large scale. With advancements in artificial intelligence technology, machine learning-based approaches have begun to be adopted to automate this process through Automated Essay Scoring (AES) systems. However, most existing AES models tend to rely solely on the final scores from the dataset without considering the structural quality of the writing, such as coherence between paragraphs. This study aims to evaluate the effectiveness of machine learning algorithms in assessing school students' essays by adding coherence features as predictor variables in a regression model. This approach uses linguistic feature representation techniques to explicitly build coherence indicators. The proposed model achieved a QWK improvement from 0.69 to 0.89 using SMOTE and coherence features. Meanwhile, human evaluation results showed that the pair of Rater 1 and Rater 2 achieved a QWK of 0.82, the pair of Rater 1 and Rater 3 scored 0.79, and the pair of Rater 2 and Rater 3 scored 0.81. These values indicate a high level of agreement among raters, suggesting that the assessment instrument used is stable. The main contribution of this study is introducing the coherence feature as an explicit predictor in the AES model, filling the gap not provided by standard datasets and proving that coherence improves model accuracy. This research provides practical benefits such as speeding up the evaluation process, reducing teachers' workload, and improving the objectivity and consistency of assessment in language education and evaluation.

Forecasting Hotel Demand with Time Series Prediction Model Using Random Forest Regression

2025-12-05T03:46:43+00:00

The tourism sector, as one of the main contributors to national foreign exchange, relies heavily on the growth of the hospitality industry. Improvements in this sector are expected to enhance service quality and strengthen the overall image of tourism. However, the hospitality industry is highly dynamic, with fluctuating room demand influenced by both internal and external factors, creating challenges for accurate demand forecasting. This study develops a hotel demand prediction model using internal variables (occupancy rate, reservations, cancellations, and lead time) and external variables (events and visitor numbers). The Random Forest Regression method was employed, with predictive performance evaluated through a proxy demand index. The dataset was obtained from Adiwana Unagi Suites, Ubud, Bali, covering historical time series data from November 2021 to July 2025 with a total of 18.674 transactions. Evaluation metrics included Mean Absolute Error, Mean Square Error, Root Mean Square Error, and R-squared, applied to each hotel room type. The results demonstrate strong predictive performance, with R-squared values of 99.83% for test data, 99.95% for training data, and 88.24% for three-month prediction data, accompanied by low error values across all metrics. The lower performance in the three-month forecast may be due to the proxy demand index not fully representing actual demand. Overall, the findings highlight the potential of machine learning approaches, particularly Random Forest Regression, to support decision-making in hotel management. The model can serve as a reference for room pricing, allocation, and operational strategies, enabling stakeholders to adapt effectively to fluctuating market demand.

IndoBERT-Based Pediatric Disease Classification and Symptom-Based Traditional Medicine Recommendation from Lontar Usada Rare

2025-11-10T08:11:32+00:00

This study aims to develop a Balinese traditional text-based pediatric disease classification model using a fine-tuned IndoBERT model on the Lontar Usada Rare dataset. The dataset used consists of 422 entries containing disease symptoms, disease types, medicinal ingredients, and treatment procedures obtained from transliteration of lontar manuscripts and interviews with traditional medicine experts. Pre-processing was done through case folding, cleansing, and normalization, followed by label encoding on 35 disease classes. The IndoBERT model was fine-tuned using the AdamW optimizer with a learning rate of 5e-5, batch size 8, and 15 epochs. Evaluation results showed the model was able to achieve 90.59% accuracy, 94.71% precision, 90.59% recall, and 90.99% F1-score, indicating excellent performance in understanding the linguistic context of traditional medical text. The developed recommendation system integrates model prediction with TF-IDF-based cosine similarity method to provide the most relevant treatment recommendations based on user symptom input. This research makes an important contribution to the digitization and preservation of Balinese traditional medical knowledge through the development of a structured and widely accessible digital knowledge base.

Evaluation of MobileNet-Based Deep Features for Yogyakarta Traditional Batik Motif Classification

2025-12-05T03:44:26+00:00

Batik is an Indonesian intangible cultural heritage that embodies profound philosophical, aesthetic, and cultural values. Yogyakarta batik motifs, such as Parang, Kawung, and Truntum, reflect Javanese wisdom and identity through distinctive geometric and floral patterns. In the digital era, artificial intelligence based image processing provides a promising approach to support the preservation and automatic recognition of traditional batik motifs. The objective of this study is to evaluate the effectiveness of MobileNet-based feature extraction combined with Support Vector Machine (SVM) classification for Yogyakarta batik motif recognition. The proposed method employs MobileNet as a convolutional feature extractor and SVM as a decision model to separate motif classes in the feature space. Experiments were conducted on 685 batik images consisting of three motif classes, with class imbalance handled using Synthetic Minority Over-sampling Technique (SMOTE). Model performance was evaluated using weighted accuracy, precision, recall, and F1-score under five-fold cross validation. The results show that MobileNetV3Large achieved the best performance with a weighted accuracy of 98.36%, followed by MobileNetV3Small and MobileNetV4Small. Statistical significance tests using the Friedman test and Wilcoxon signed-rank analysis confirm that the performance differences among the evaluated models are statistically significant. These findings indicate that MobileNetV3 architectures provide robust and discriminative feature representations for batik motif classification on limited yet structured datasets. This study contributes a validated MobileNet–SVM framework for batik recognition and supports ongoing efforts in the digital preservation of Indonesia’s cultural heritage. Future work will explore larger motif sets and cross-dataset evaluation to further improve generalization performance.

Blockchain Disaster-Relief DApps with SVM and Data Anchors for Fraud-Prevention

2025-11-02T14:40:30+00:00

VoucherAid and DataAnchor are prototype DApps for disaster-relief voucher processing that integrate on-chain rule enforcement, cryptographic data anchoring through fixed-size hash commitments, and an off-chain SVM-based analytics gateway. VoucherAid issues non-transferable vouchers, restricts redemption to certified merchants, and emits auditable events, while DataAnchor records time-stamped digests to support provenance verification without exposing sensitive content. A 200-record dataset was generated from on-chain logs and enriched with behavioral–temporal features derived from redemption activity. Experiments conducted in a single-node Ganache environment using a 70:30 split show that the SVM achieves 0.75 accuracy with perfect precision but limited recall for fraud (1.00 precision, 0.32 recall, 0.48 F1), indicating that the model cannot serve as a reliable stand-alone detector and is more appropriate as a conservative decision-support tool under human oversight. The prototype demonstrates that separating on-chain enforcement from off-chain analytics can enhance auditability and support model evolution without contract redeployment. However, the findings remain constrained by the small, partially synthetic dataset, the single-node evaluation environment, and programmatic labeling. Future work will expand datasets, incorporate richer temporal and graph-based features, adjust thresholds and class weights, and evaluate the system on multi-node networks to improve fraud recall while maintaining usability and inclusion.

A Comparative Study of MobileNetV2 and ResNet50 for Multi-Class AI- Generated and Real Image Classification

2025-12-08T07:45:09+00:00

This study aims to classify AI-generated and real images using Convolutional Neural Network (CNN) architecture by comparing the performance of MobileNetV2 and ResNet50. Previous studies on AI-generated image detection have primarily focused on binary classification without explicitly analyzing object-level context in multi-class scenarios, leaving a gap in understanding model performance across diverse visual categories. The dataset consists of 23,941 images divided into two main classes of real and fake and five subclasses of human, animal, art, view, and vehicle. The training process employs data augmentation and a K-Fold Cross Validation strategy on the training and validation set to maintain balanced class proportions, while a separate unseen test set is used exclusively for final performance evaluation. Model evaluation is performed based on accuracy, precision, recall, and F1-score metrics on test data. The results showed that MobileNetV2 achieved the best accuracy of 89% at the 10th epoch, but experienced a decline in performance at the 30th and 50th epochs, indicating overfitting. In contrast, ResNet50 showed the most stable performance with the highest accuracy of 93% at the 30th epoch and consistently high precision, recall, and F1-score values. Thus, ResNet50 was found to be the most effective architecture for classification of AI-generated and real images on multi-class datasets, while MobileNetV2 remains relevant for implementation on devices with computational limitations.

Machine Learning Analysis of Jakarta Bay Water Quality: Comparing Models

2025-11-07T03:40:29+00:00

Jakarta Bay experiences persistent anthropogenic pressures that produce spatially heterogeneous water-quality conditions. This study develops a regulation-aligned, explainable classification framework using a 2024 in-situ dataset collected at 53 stations across two sampling periods (March and August). After preprocessing—including unit harmonization, outlier screening, missing-value imputation, and treatment of below-detection-limit measurements—the dataset yielded 104 complete samples classified into Good (n=46), Lightly Polluted (n=28), and Moderately Polluted (n=34) categories based on KEPMEN LH No. 51/2004. Three ensemble algorithms (LightGBM, CatBoost, and Random Forest) were evaluated using stratified cross-validation to maintain class balance and prevent spatial leakage. CatBoost achieved the best overall performance (Accuracy = 0.8338; F1 = 0.8257), followed by Random Forest, while LightGBM showed the highest variability across folds. Class-level metrics indicate that CatBoost produced the most balanced predictions, particularly for the borderline Lightly Polluted class. SHAP analysis identified turbidity/TSS, nutrients, dissolved oxygen, salinity, and spatial gradients as dominant predictors, enabling transparent interpretation of model decisions. The resulting framework provides a reproducible and operationally deployable approach for rapid screening, hotspot detection, and decision support in Jakarta Bay’s water-quality management.

Enhanced Stacked GRU Model for Monthly Rice Production Forecasting in Bali Province

2025-12-23T07:53:50+00:00

Rice production has a seasonal pattern that depends on the planting cycle and environmental conditions, requiring forecasting methods that can accurately model temporal dynamics. This study aims to predict monthly rice production in Bali Province using the Stacked Gated Recurrent Unit (GRU) architecture. Monthly rice production data from 2018 to 2024 from the Central Statistics Agency (BPS) was used as the main source. The preprocessing stage includes data cleaning, Min-Max normalization, and feature engineering in the form of creating sin_month and cos_month features to capture seasonal patterns, as well as a 3-month rolling mean to extract short-term trends. The proposed stacked design with dual-layer GRU combined with seasonal features improves temporal pattern extraction compared to single-layer GRU baselines. The model was tested using three configurations, and Scheme 3 provided the best performance with an MAE value of 1610.21, an RMSE of 2055.90, and a MAPE of 14.29%, which is considered good accuracy. The model was able to follow seasonal production trends, including an increase at the beginning of the year and a decrease during the planting period. Long-term predictions for the next 12 months and quarterly forecasts per district/city also showed patterns consistent with historical data. The results of the study indicate that Stacked GRU is effective in forecasting seasonal rice production and can be used as a basis for decision support in food security planning in Bali.

Comparative Analysis of Four Machine Learning Algorithms for Smoke Detection Using SMOTE-Rebalanced Sensor Data

2025-11-10T09:48:09+00:00

Smoke detection plays a critical role in preventing fire-related hazards, particularly in intelligent monitoring and early warning systems. Conventional smoke sensors often exhibit limited responsiveness in dynamic environmental conditions, prompting the adoption of IoT-based sensor data combined with machine learning techniques. This study presents a comparative evaluation of four supervised classification algorithms, K-Nearest Neighbors (KNN), Decision Tree, Random Forest, and Gradient Boosting, using the Smoke Detection Dataset from Kaggle. The methodology integrates SMOTE to address class imbalance and Z-score normalization for feature standardization. Hyperparameter tuning was performed using GridSearchCV with 5-fold cross-validation, and model performance was assessed based on accuracy and execution time. Experimental results show that KNN achieved the highest accuracy (98.33%) with the lowest execution time (0.0327 s), whereas Decision Tree recorded the lowest accuracy (84.17%) but remained computationally fast (0.0406 s). Random Forest and Gradient Boosting demonstrated strong predictive capability (97.22% and 96.94%, respectively), but at higher computational costs (1.4338 s and 8.3819 s, respectively). Almost all models achieved perfect scores (1.00) for precision, recall, and F1-score following SMOTE-based balancing, except KNN which obtained slightly lower values (0.99). The findings indicate a trade-off between predictive performance and computational efficiency, suggesting that lightweight models such as KNN are better suited for real-time IoT-based smoke detection. In contrast, ensemble models may be more appropriate for backend analysis. This research contributes an integrated evaluation framework that combines data rebalancing, multi-model benchmarking, and time-based performance analysis, providing practical insights for the development of responsive and scalable early smoke detection systems.

An Integrated K-Means and Composite Risk Scoring Framework for Urban Dengue Vulnerability Mapping

2025-12-24T19:14:58+00:00

The rising incidence of dengue hemorrhagic fever (DHF) in Indonesian urban areas highlights the urgent need for analytical frameworks capable of capturing spatial heterogeneity in vulnerability while supporting targeted public health interventions. However, most existing dengue vulnerability studies rely on clustering or indicator-based scoring in isolation, limiting interpretability and reducing their operational relevance for policy-driven decision making. This study explicitly addresses this gap by proposing an integrated spatial clustering and epidemiologically weighted composite risk scoring framework for urban dengue vulnerability mapping. Using Semarang Municipality as a case study, K Means based spatial clustering was combined with composite risk scoring to analyze dengue vulnerability across administrative subdistricts. Seven key indicators consisting of population density, area size, total population, morbidity, mortality, incidence rate, and health facility availability were processed through systematic imputation, normalization, and attribute selection to ensure data consistency and analytical robustness. The optimal number of clusters was determined using the Elbow Method and Silhouette Score, after which K-Means clustering was applied to generate spatially coherent vulnerability groupings. A composite risk scoring mechanism was subsequently employed to classify regions into five operational risk categories: Low-Risk, Moderate-Risk, High-Risk, Very High-Risk, and Emergency-Priority. The results reveal clear structural differentiation in dengue vulnerability patterns, where Emergency-Priority and Very High-Risk clusters are not only characterized by elevated epidemiological indicators but also by constrained health service availability, amplifying outbreak susceptibility. Specifically, 13 subdistricts (7.5%) were identified as Emergency-Priority and 22 subdistricts (12.4%) as Very High-Risk, together accounting for approximately 20% of the study area. Beyond numerical classification, the integration of spatial clustering and composite risk scoring enhances interpretability by linking cluster structure with epidemiological severity and service capacity, thereby improving policy relevance compared to conventional clustering-only approaches. Validation through heatmap visualization, risk category distribution, and cluster ranking confirms the stability and interpretive clarity of the proposed framework. By moving beyond descriptive clustering toward an integrated analytical model, this study contributes a scalable and adaptive decision-support framework for dengue risk mapping. The findings provide actionable insights for policymakers, enabling evidence-based prioritization, optimized resource allocation, and the development of responsive intervention strategies to mitigate dengue burden in complex urban environments.

Comparative Study of Baseline and CBAM-Enhanced ResNet50 and MobileNetV2 for Indonesian Rupiah Banknote Classification

2025-11-26T17:09:46+00:00

This study investigates the performance of Convolutional Neural Network (CNN) architectures enhanced with Convolutional Block Attention Module (CBAM) for Indonesian banknote classification. Although attention mechanisms have shown strong potential in improving fine-grained visual recognition, their effectiveness for the classification of banknotes with fine textures and similar color patterns remains underexplored, forming a key research gap addressed in this work. Four architectures, ResNet50, ResNet50+CBAM, MobileNetV2, and MobileNetV2+CBAM, were evaluated using K-Fold cross-validation on a dataset of 1,281 images representing seven banknote denominations. Experimental results show that ResNet50 achieves strong baseline performance with a weighted Train accuracy of 99.14% and a Val accuracy of 96.72%, while the integration of CBAM further improves feature discrimination, with ResNet50+CBAM obtaining the highest average accuracy across all folds with a weighted Train accuracy of 100% and a Val accuracy of 99.45%. MobileNetV2 showed lower performance due to its lightweight capacity with a Train accuracy of 91.88% and a decrease in Val accuracy of 85.71%. However, the addition of CBAM provided measurable improvements and greater stability with a Train accuracy of 99.61% and Val accuracy of 92.82%. Overall, CBAM improved CNN’s ability to focus on spatial information and salient channels, resulting in more reliable classification. ResNet50+CBAM emerged as the best-performing model, offering the best balance between accuracy and consistency. These findings support the development of reliable computer vision systems for financial technology applications, including automatic banknote recognition, counterfeit detection, and secure transaction verification.

Enhanced Performance Evaluation of VGG16 and ResNet50 for Deepfake Detection Using Local Ternary Pattern

2025-11-14T08:56:29+00:00

Deepfake video generation has become increasingly sophisticated, posing challenges for detection methods that rely solely on convolutional neural networks (CNNs without explicit texture enhancement). Many existing approaches have limited robustness in capturing subtle texture inconsistencies caused by manipulation, compression, and noise. This study investigates the integration of Local Ternary Pattern (LTP)–based texture enhancement with transfer learning models for deepfake video detection. Specifically, VGG16 and ResNet50 architectures are evaluated using the Celeb-DF (v2) dataset. LTP is employed to extract fine-grained texture features due to its higher robustness to illumination variations and noise compared to conventional descriptors such as Local Binary Pattern (LBP). Video frames are processed individually and used to train CNN classifiers, followed by evaluation at both frame and video levels. Experimental results show that ResNet50 outperforms VGG16, achieving a test accuracy of 93% with a validation loss of 0.2228, while VGG16 reaches an accuracy of 88% with a validation loss of 0.2636. Further testing on 20 withheld videos demonstrates that ResNet50 correctly classifies all samples, whereas VGG16 misclassifies two real videos, indicating lower robustness to real-video misclassification. These results demonstrate that LTP-based texture enhancement effectively supports CNN-based deepfake detection and that deeper architectures benefit more from enriched texture representations. This study provides empirical insights into improving robustness and reliability in deepfake video classification.

Compliance-by-Design for Bonded Warehouses: A Web-Based Inventory System with CEISA-Verified Gate Control and D&M Success Evaluation

2026-01-14T04:35:38+00:00

In bonded warehouses, manual inventory procedures pose serious risks to both operational effectiveness and regulatory compliance. This study aims to develop and then implement inventory management of a web-based information system intended to reduce these risks using an action research methodology. The DeLone & McLean IS Success Model was used to evaluate the system's efficacy. Information quality, system quality, and service quality are strong positive drivers of both use and user satisfaction, according to an analysis based on data from 15 key users. The system's Net Benefits, which were measured by a 99% decrease in reporting time and a substantial reduction of data entry errors, are thus highly predicted by these factors. The high scores in Use (4.8) and Net Benefits (4.8) confirm that the system was not only technically proficient but also delivered tangible, strategic value to the organization. Structural model analysis confirmed that Information Quality, System Quality, and Service Quality significantly influence Use and User Satisfaction (> 0.40, p < 0.01), with an R² of 0.645 indicating substantial explanatory power. This study concludes that the system is a highly effective solution for enhancing compliance, accuracy, and operational performance in a bonded warehouse environment. This action research study successfully addressed the complex operational and regulatory challenges inherent in bonded warehouse management. The primary limitation is the single-site case study design. Future work could involve a comparative study across multiple bonded warehouses to identify more universal best practices and contextual factors that influence system success.

Explainable Machine Learning-Based Decision Tree Model for Early Detection of Hypertension Risk

2025-11-22T04:11:46+00:00

Hypertension is one of the leading causes of cardiovascular disease and is often referred to as a “silent killer” because it typically remains asymptomatic until serious complications, such as stroke or kidney failure, occur. Early detection of hypertension risk is therefore essential to enable timely intervention and prevention. This study aims to develop an explainable machine learning–based Decision Tree model for early detection of hypertension risk using clinical and lifestyle data. The balanced dataset includes variables such as age, body mass index (BMI), blood pressure, family history, smoking habits, stress levels, and sleep duration. The dataset used in this study was obtained from the “Hypertension Risk Prediction Dataset” available on the Kaggle platform, consisting of 1,985 patient records and 11 main features covering variables such as age, body mass index (BMI), systolic and diastolic blood pressure, family history, smoking habits, stress level, physical activity, and sleep duration. The dataset is balanced between the hypertension and normal categories, enhancing the reliability of the classification results. The model was constructed using a Decision Tree Classifier implemented in Scikit-learn and validated through cross-validation to minimize overfitting. Model performance was assessed using accuracy, precision, recall, F1-score, and AUC-ROC metrics. The results indicate that the model achieved an accuracy of 96% and an AUC of 0.9645, demonstrating excellent classification performance. The motivation behind this research lies in the growing need for interpretable artificial intelligence models in healthcare, where transparency and explainability are critical for clinical trust and ethical decision-making. Unlike black-box models, the Decision Tree approach allows clinicians to trace each prediction path, understand contributing variables, and apply insights in real-world medical settings. The primary advantage of this model lies in its transparency, as each prediction can be interpreted through explicit decision rules. Overall, this explainable and high-performing model shows strong potential as a clinical decision support tool for early hypertension screening and prevention programs.

Decision Model for Best Contraceptive Technique Recommendation Based on Patient's Ideal Profile

2025-10-05T04:27:07+00:00

Choosing the right contraceptive method is essential to support the success of family planning programs. Many patients still choose methods without considering their medical conditions, which can lead to failure or side effects. This study designed a decision-making model based on Profile Matching to recommend contraceptive methods according to the patient’s ideal profile. The dataset was obtained from Faskes Level 1 Udayana Denpasar. Validation was conducted through discussions with midwives as experts, referring to the KLOP KB Wheel as the standard issued by the WHO. The evaluation results show a high level of agreement between the model’s recommendations and expert judgments, indicating that the model provides more objective and easily understood recommendations compared to manual approaches.

Implementing a Payment Gateway in the Mount Slamet Hiking Ticketing System

2025-12-04T15:12:19+00:00

Background—Many hiking basecamps in Indonesia still process bookings manually, causing queues, quota uncertainty, and errors in payment verification that hinder operations. Objective— Design and implement a web-based information system (e-ticketing) for Mount Slamet hiking, integrated with the Midtrans payment gateway; validate transactions in near-real-time and issue ticket IDs for gate inspection. Methods—Development followed Agile/Scrum. Requirements were gathered through observation and interviews; the design employed use-case, activity, ERD, and payment-flow models. Implementation used React (UI), Express and Prisma ORM (API), MySQL, and Midtrans Snap, with signature-verified, idempotent webhooks. Trials covered end-to-end black-box testing (booking; transitions among pending, paid, expired, and canceled; ticket-ID issuance; and check-in), cross-browser compatibility (Chrome, Edge, Firefox, Safari on desktop and mobile), and the System Usability Scale (SUS; n = 13). We also monitored propagation time from settlement to order update and behavior in the admin panel (route, quota, and date-closure management). Results—All functional scenarios passed; behavior was consistent across major browsers; mean SUS = 75.0 (> 68) indicates acceptable usability. Webhooks ensured automatic, duplicate-free status updates, with propagation on the order of seconds, so the reservation–payment–e-ticket chain operated end-to-end and was traceable via ticket-ID logs. Conclusion—The proposed e-ticketing system is technically feasible for basecamp operations and provides an architectural blueprint, core data schema, and a replicable Midtrans integration pattern. Future work will refine the public interface, add refund/void features, and conduct production-grade performance and security testing.

Comparative Study on Machine Learning Algorithms for Code Smell Detection

2025-10-22T06:13:18+00:00

Detecting code smells is crucial for maintaining software quality, but rule-based methods are often not very adaptive. On the other side, existing machine learning studies often lack large-scale comparisons on modern datasets. The goal of this research is to comprehensively compare the performance of various machine learning algorithms for multi-label code smells classification in terms of effectiveness and efficiency. The dataset used in this research is SmellyCode++, containing more than 100,000 samples. Seven models: Logistic Regression, Linear SVM, Naive Bayes, Random Forest, Extra Trees, XGBoost, and LightGBM combined with Binary Relevance were trained on data balanced using random undersampling and multi-label synthetic minority over-sampling. The performance of each model was evaluated using the F1-Macro, Hamming Loss, and Jaccard Score metrics. A non-parametric statistical analysis was also conducted to validate the findings. The experiment found that ensemble-based models statically significantly outperformed the linear and probabilistic models. The performance among the top ensemble models was found to be statistically equivalent. With this statistical equivalence in accuracy, computational efficiency measured with training time became the critical tiebreaker. BR_RandomForest, BR_XGBoost, and BR_ExtraTrees proved highly efficient, while BR_LightGBM was significantly slower. This study concludes that BR_RandomForest offers the best overall trade-off in providing top tier accuracy combined with excellent computational efficiency, making it a robust choice for practical applications.