Jurnal Teknik Informatika (Jutif)

Expert System for Diagnosing Autoimmune Diseases Using Dempster–Shafer and Fuzzy Logic: A Case Study of Prof. Dr. Margono Soekarjo Regional Hospital

Ragil Putri Rahmadani, Yohani Setiya Rafika Nur, Annisaa Utami — Sat, 18 Apr 2026 00:00:00 +0000

Autoimmune diseases, particularly lupus, pose a major challenge in healthcare because their symptoms are highly variable and often mimic other medical conditions. Delayed diagnosis can worsen patient outcomes, increase the risk of severe complications, and even lead to death, especially in healthcare facilities with limited autoimmune subspecialists, such as Prof. Dr. Margono Soekarjo Regional Hospital. This study aims to develop a web-based expert system to support early screening for lupus by combining the Fuzzy Tsukamoto method and the Dempster-Shafer theory. The Fuzzy Tsukamoto method is used to represent symptom uncertainty through fuzzification, while the Dempster-Shafer theory is used to combine evidence from individual symptoms to produce confidence levels for possible diagnoses. The research process included a literature review, expert interviews, construction of a symptom–disease knowledge base, design of fuzzy rules, implementation of mass function calculations, and development of a web-based diagnostic application. Testing was conducted using ten patient test cases with confirmed expert diagnoses. The test results showed an accuracy of 100%, with all system diagnoses matching the experts’ diagnoses. The strength of this research lies in the integration of two inference methods to improve the accuracy of evidence calculation, and in the use of symptom uniqueness and occurrence parameters that were validated directly by experts. This system has the potential to serve as an effective early screening tool for healthcare providers and patients, particularly in resource-limited settings. From an informatics perspective, this study contributes to the development of intelligent decision support systems by demonstrating the effectiveness of a hybrid reasoning approach in handling uncertainty in medical diagnosis. The integration of Fuzzy Tsukamoto and Dempster–Shafer methods enhances diagnostic consistency and reliability, making the proposed system relevant for research in expert systems and medical informatics.

Optimizing Multimodal Health Chatbots through the Integration of Medical Text and Images

Raditya Danar Dana, Mulyawan, Agus Bahtiar, Odi Nurdiawan — Wed, 15 Apr 2026 00:00:00 +0000

This study is motivated by the growing need for image-classification systems that remain accurate despite variations in image quality commonly found in real-world environments. Differences in image resolution often lead to decreased performance of Convolutional Neural Network (CNN) models, particularly in scenarios involving limited acquisition devices. This research aims to analyze the effect of image-resolution variations on CNN robustness by applying an adaptive augmentation strategy. An experimental approach was employed by manipulating independent variables namely image-resolution levels and augmentation techniques and observing their impact on accuracy, validation stability, and model generalization. The results show that medium-resolution images (128×128 px) combined with adaptive augmentation produce the best performance, yielding the highest validation accuracy and reduced overfitting compared to other configurations. The urgency of this study lies in its practical contribution to developing efficient image-classification models suitable for resource-constrained environments. Scientifically, the findings provide a structured mapping of the relationship between resolution, augmentation, and model stability, offering a foundation for designing more robust CNN architectures adaptable to real-world data variability.

Automated Facial Wrinkle Segmentation for Dermatological Assessment Using VGG-Based U-Net with Hybrid Augmentation

Wahyu Fajar Setiawan, Nanik Suciati — Sat, 18 Apr 2026 00:00:00 +0000

Manual and automated facial wrinkle segmentation remains challenging due to the fine-grained nature of wrinkles, uneven distribution across facial regions, severe class imbalance (~2% wrinkle pixels), and sensitivity to lighting variations—limiting the reliability of existing dermatological assessment tools. This study aims to evaluate VGG transfer learning with hybrid augmentation strategies for U-Net-based automated facial wrinkle segmentation. Using the FFHQ-Wrinkle dataset comprising 1,000 manually annotated high-resolution images (1024×1024 pixels), this study systematically evaluates three U-Net variants (Baseline, VGG16-based, VGG19-based) across four augmentation strategies: no augmentation, hierarchical image enhancement (CLAHE, gamma correction, bilateral filtering), geometric transformation (rotation, translation, shear, zoom, flip), and hybrid combination. A multi-component loss function integrating Focal Loss, Dice Loss, IoU Loss, and Boundary Loss addresses class imbalance while optimizing both region overlap and edge localization. The proposed VGG19-based U-Net with hybrid augmentation achieves state-of-the-art performance: Dice coefficient of 0.6585, IoU of 0.4970, precision of 0.6186, recall of 0.7344, and Boundary F1 of 0.9185. Key findings demonstrate that VGG19 transfer learning provides +21.54% Dice improvement over Baseline U-Net with 12.7-fold reduction in overfitting, while hybrid augmentation yields +4.87% Dice improvement with +2.24% synergistic gain beyond individual strategies. This research advances automated dermatological tools for precise skin health assessment, reducing subjectivity in clinical evaluations and providing actionable guidelines for practitioners developing automated wrinkle analysis systems.

Performance Analysis of Traditional Machine Learning Classifiers on LSTM-Extracted Features for Indonesian Sign Language System Recognition

Patricia Ho, Handri Santoso — Wed, 15 Apr 2026 00:00:00 +0000

Recognizing affix gestures in the Indonesian Sign Language System (SIBI) remains challenging due to subtle visual differences in hand shape and movement, often resulting in lower classification accuracy compared to other categories. This study aims to evaluate whether lightweight traditional and hybrid classifiers can provide competitive performance to deep learning models for SIBI recognition. Using a dataset of 21,351 gesture videos covering four categories (Affix, Alphabet, Number, and Word), features were extracted from MediaPipe keypoints and processed as frozen LSTM embeddings. Six classifiers (Random Forest, K-Nearest Neighbors, Naïve Bayes, Multilayer Perceptron, Support Vector Machine, and Hidden Markov Model) were evaluated with 5-fold stratified cross-validation using accuracy, precision, recall, and F1-score, with statistical significance tested through Friedman and Nemenyi analyses. Results show that MLP and RF achieved high performance in Alphabet, Number, and Word categories (above 96 percent accuracy), while Affix remained the most difficult, with MLP reaching 81.17 percent, outperforming the 68.17 percent from a prior BiLSTM model. This study provides a benchmark for hybrid model implementation in sign language recognition, showing that while traditional classifiers on deep features are effective and computationally lighter for general gestures, deep architectures remain superior for capturing the fine-grained temporal nuances critical for complex categories like affixes.

Improving the Performance of K-Means Algorithm using the Particle Swarm Optimization for Clustering Forest Fire

Putri Utami — Sat, 18 Apr 2026 00:00:00 +0000

Forest and land fires are a recurring ecological disaster in Indonesia, particularly in West Kalimantan, where peatlands and tropical climates contribute to high vulnerability. Effective identification of fire-prone areas is critical for mitigation efforts, yet conventional clustering methods such as K-Means suffer from limitations, especially in determining optimal cluster numbers and centroid initialization. This study proposes an enhanced clustering approach by integrating the Particle Swarm Optimization (PSO) algorithm with K-Means to improve the accuracy of hotspot clustering in Mempawah Regency. The research utilizes hotspot and weather datasets from January 2023 to March 2024, incorporating variables such as temperature, humidity, rainfall, and wind speed. Data preprocessing includes normalization using Min-Max Scaling. PSO is applied to determine the optimal number of clusters (K) within a range of 2 to 10 by evaluating Davies-Bouldin Index (DBI), Silhouette Coefficient (SC), and inertia. Experimental results show that the optimal configuration—20 iterations and 20 particles—yields a DBI of 0.897 and an SC of 0.464, indicating standard cluster quality. Visual validation using PCA demonstrates clear cluster separation, supporting the evaluation results. Compared to visual methods like Elbow, which suggest K=3–4, the PSO-KMeans approach identifies K=10 as optimal, providing better clustering performance. This research highlights the effectiveness of swarm intelligence in enhancing spatial data modeling and supports strategic decision-making for local wildfire mitigation efforts.

Bidirectional Cross-Attention and Uncertainty-Aware Ensemble for High-Precision Brain Tumor Classification on MRI Images

Gunawan Cholis Saputra, Muhammad Yuwanandra Risdyaksa — Wed, 15 Apr 2026 00:00:00 +0000

Accurate brain tumor diagnosis via Magnetic Resonance Imaging (MRI) is vital for effective neuro-oncological treatment. Although CNNs are widely regarded as the benchmark for local texture extraction, they frequently exhibit limitations in modeling long-distance global dependencies effectively. In contrast, Vision Transformers (ViTs), particularly the Swin variant, demonstrate superior capability in capturing global semantic context yet often fail to preserve the fine local granularity needed to delineate tumor boundaries significantly. To bridge this gap, we propose Bi-CA-UAE, a hybrid framework integrating Swin Transformer and EfficientNet-V2 through a novel Bidirectional Cross-Attention mechanism. Unlike static ensembles, our method enables dynamic information exchange between global and local feature maps before classification. Furthermore, we introduce an Uncertainty-Aware Gating Network to adaptively weigh each branch based on prediction confidence, reducing false positives in ambiguous cases. Validated on a multi-class MRI dataset of 7,023 images, the model achieved 99.85% accuracy and an Expected Calibration Error (ECE) of 0.02, matching the strongest baseline (Swin Transformer) while demonstrating superior training stability and calibration. Unlike naive concatenation ensembles that suffered from overfitting and performance degradation in later training stages, Bi-CA-UAE maintained robust peak performance. Additionally, the model attained perfect recall (1.00) for Glioma and a micro-average AUC of 1.00. t-SNE visualizations and reliability diagrams confirm the model's ability to learn highly discriminative and well-calibrated features, positioning it as a trustworthy clinical decision support system.

Performance Evaluation of Gradient Boosting Techniques for Predicting Customer Purchase Decisions

Wed, 15 Apr 2026 00:00:00 +0000

Customer purchase prediction remains a critical challenge in e-commerce and retail analytics, with significant implications for marketing strategies and business revenue. This research provides a detailed comparative evaluation of advanced gradient boosting techniques XGBoost, LightGBM, and CatBoost to predict customer purchasing behavior using review trends and demographic factors. The study employed a dataset of 100 customer records with attributes such as age, gender, review quality, and education level. Through systematic feature engineering, including age group categorization and categorical feature combinations, as well as addressing class imbalance using the Synthetic Minority Oversampling Technique (SMOTE), all three models were trained and evaluated using default hyperparameters with optimal settings. The experimental results show that CatBoost achieved the best performance, with 78.26% accuracy, 0.8011 precision, 0.7826 recall, and a 0.7775 F1-score, outperforming LightGBM (73.91% accuracy) and XGBoost (60.87% accuracy). The evaluation includes confusion matrix analysis, precision–recall metrics, and visual comparisons across all performance dimensions. These findings provide valuable insights for practitioners selecting appropriate machine learning algorithms for customer purchase prediction tasks, particularly in scenarios involving limited datasets and categorical features. This research contributes to the growing body of literature on the use of gradient boosting techniques for predicting consumer behavior and offers important practical implications for e-commerce applications. These findings offer important contributions to machine learning applications in customer behavior prediction.

Job Recommendation for Fresh Graduates to Reduce Competency Gaps Using Content-Based Filtering and Retrieval-Augmented Generation

Iftitah Yanuar Rahmawati, Felda Mufarihati, Christian Sri Kusuma Aditya — Sat, 18 Apr 2026 00:00:00 +0000

Job recommendation systems are frequently used to help job seekers find suitable positions. Nevertheless, many existing systems focus primarily on accuracy and provide limited justification. This lack of openness can erode user confidence, particularly among recent grads who need a clear explanation of how their individual experiences fit the recommendations. Furthermore, these systems frequently lack sophisticated methods to explain the reasoning behind the recommendations, such as Retrieval-Augmented Generation (RAG), which makes them seem impersonal and difficult to trust. The purpose of this research is to develop an explainable job recommendation system that generates employment suggestions based on language comprehension by integrating RAG and Content-Based Filtering (CBF). User profiles and open positions are displayed using TF-IDF and Sentence-BERT, and then the experience level-based cosine similarity is calculated. To measure competency coverage, matching and absent skills are identified in an explicit skill-gap analysis. The Large Language Model and FAISS-based RAG modules leverage the explanations that are produced by finding matched and missing abilities as context. The CBF approach was used to evaluate recommendation relevance, while BLEU and ROUGE on ten test documents were used by HR specialists for validation. The system's mean ROUGE-1 F1 score was 0.4659, and its mean ROUGE-L score was 0.2918, based on 10 evaluation cases. Results show that the proposed recommendation system provides accurate and adequate guidelines based on HR references. This paper enriches Informatics by consolidating semantic similarity modeling, explicit competency-gap reasoning, and grounded text generation together to form a cohesive explainable recommendation framework targeted to cold-start job seekers.

Clustering And Classification Of Toddler Stunting Risk Using K-Means And Naive Bayes: A Case Study At Kembaran 1 Community Health Center

Lulu Amnah Fitriya Maharani, Purwadi, Debby Ummul Hidayah — Sat, 18 Apr 2026 00:00:00 +0000

Stunting continues to be a significant public health concern in Indonesia, with a frequency of 17.25% at Kembaran 1 Public Health Center, highlighting ongoing difficulties in early childhood nutrition and growth surveillance. This work seeks to assess and forecast stunting risk in toddlers by employing K-Means clustering and Naive Bayes classification to enhance early detection precision. The K-Means method was utilized on 1,168 toddler growth records to categorize stunting features, whereas the Davies–Bouldin Index (DBI) was employed to evaluate cluster quality. The ideal cluster was attained at k = 8, yielding a DBI value of 4.353, indicating compact and distinctly differentiated clusters. The Naive Bayes classifier subsequently predicted stunting potential with an accuracy of 93.56%, accurately categorizing 218 out of 233 test examples, yielding precision, recall, and F1-score values for the “short” class of 97.41%, 94.95%, and 96.18%, respectively. The findings indicate that the hybrid model successfully combines unsupervised and supervised learning, improving stunting prediction accuracy and cluster interpretability. The research provides a data-centric framework for localized stunting surveillance, aiding community health centers in formulating targeted early treatments and mitigating long-term developmental hazards.

Optimization Of Hybrid K-Means–Naïve Bayes Using Optuna for Classification of Global Plastic Waste Management Levels

Aulya Fani Madani, Poningsih, Zulia Almaida, Widodo Saputra — Sat, 18 Apr 2026 00:00:00 +0000

The rapid growth of plastic waste has become a serious global environmental challenge, while existing waste management analysis methods often struggle to handle large and heterogeneous environmental datasets. This study aims to improve the classification of global plastic waste management performance by integrating K-Means clustering and Naïve Bayes with Optuna-based hyperparameter optimization. Using a dataset of global plastic waste indicators from multiple countries during 2020–2024, K-Means is first applied to generate waste management level clusters, which are then classified using Naïve Bayes. The hybrid model is further optimized by tuning the var_smoothing parameter using Optuna. Experimental results show that the hybrid approach improves classification performance compared to the baseline Naïve Bayes model, while the optimized model increases accuracy from 89% to 95% along with improvements in precision, recall, F1-score, and ROC-AUC. These results indicate that combining clustering-based labeling with automated hyperparameter optimization can enhance the reliability of machine learning models for large-scale environmental data analysis. Therefore, the proposed approach can support more accurate evaluation of global plastic waste management and assist data-driven environmental policy development.

Early Detection Of Melanoma Skin Cancer Using Gray Level Co-Occurrence Matrix And Ensemble Support Vector Machine

Wed, 15 Apr 2026 00:00:00 +0000

Skin cancer is a major global health problem with incidence rates increasing every year. Melanoma, the most aggressive form of skin cancer, requires accurate early detection to reduce mortality risk. Conventional diagnostic methods such as visual examination and biopsy still face limitations in precision and consistency, highlighting the need for more objective and efficient technological approaches. This study proposes a classification method for melanoma using an ensemble of Support Vector Machine (SVM) and Random Forest (RF), supported by feature extraction through the Gray Level Co-occurrence Matrix (GLCM) and dimensionality reduction using Linear Discriminant Analysis (LDA). The research stages include image preprocessing using grayscale conversion to reduce data complexity, followed by GLCM-based texture feature extraction, and LDA transformation to enhance class separability. The classification model is developed using an ensemble voting mechanism that combines predictions from SVM and RF to produce a more stable and robust decision. Experimental results with a 60:40 train–test ratio show that the proposed method achieves an accuracy of 88.75%, outperforming each individual model tested. These findings indicate that the integration of GLCM–LDA features with the SVM-RF ensemble effectively improves melanoma detection performance. Overall, this study provides a significant contribution to the development of early detection systems in health informatics, offering potential improvements in patient safety and survival rates for individuals affected by skin cancer.

Aggregation Model to Determine Criteria Weights for Integrated Primary Health Care Information System (IPCIS) Implementation

Sat, 18 Apr 2026 00:00:00 +0000

Implementation of the Integrated Primary Health Care Information System (IPCIS) in integrated community health posts (posyandu) is influenced by various factors, including technical aspects, human resources, policies, and data governance. Given the diverse field conditions, the impact of each factor can vary, so it is important to understand the relative importance of each criterion. This study aims to determine the weight of the criteria that influence the implementation of IPCIS in posyandu. Ten people answered the questions correctly (out of 22 respondents), including cadres, sub-district staff, and health workers from Tirtorahayu Village. Respondent preferences were collected using three approaches: rank-based aggregation (Borda, Condorcet, Copeland), score-based aggregation (average), and voting-based aggregation (plurality and majority) to obtain the criteria weights (w) and a comparative analysis between the approaches. The findings demonstrate that the IPCIS criteria for security and protection of personal data were consistently given the highest weights. In the ranking-based aggregation approaches (w_Borda=0.11, w_Condorcet=0.20, w_Copeland=0.19). In score-based aggregation approaches (w=0.11). In voting-based aggregation approaches (w=0.15). It is indicating a strong group consensus regarding the importance of these aspects in IPCIS implementation. The combination of ranking-based and score-based aggregation resulted in stable IPCIS implementation criterion weights that reflected group consensus, with voting-based aggregation acting as validation. The practical implication is that the obtained weighted criteria can be used as a basis for determining program priorities and resource allocation when implementing IPCIS.

A Smart System for Non-Invasive Early Detection of Diabetes through Deep Learning-Based Nail Image Analysis and Expert Systems

Muhammad Zulfikri, Wirajaya Kusuma, Naufal A. Furqan — Wed, 15 Apr 2026 00:00:00 +0000

Public health in Indonesia faces significant challenges in the early detection of diseases, particularly in areas with limited medical services. Diabetes Mellitus can lead to serious complications, but its detection is often hindered by limited access to invasive and expensive diagnostic methods. This study aims to develop a non-invasive early detection system through nail image analysis using a deep learning method based on EfficientNet-B7 and a rule-based expert system. The system classifies nail images into five categories: Healthy, Beaus lines, Onycholysis, Onychomycosis, and Paronychia. The evaluation results show an accuracy of 97.11% on the test set, demonstrating excellent performance in detecting nail conditions associated with diabetes. The application of the expert system using Forward Chaining and Certainty Factor provides in-depth medical explanations for the model's predictions, making this system a potential solution for diabetes screening that is fast, affordable, and accessible across various healthcare facilities.

Forensic Evaluation of the Effectiveness of Private Browsing Modes in Google Chrome and Mozilla Firefox Using the National Institute of Standards and Technology Framework Integrated with Artificial Intelligence

Muhammad Syukri, Asep Ririh Riswaya, Dheni Apriantsani Budiman — Sat, 18 Apr 2026 00:00:00 +0000

As cyber threats and the misuse of personal data continue to increase, private browsing modes in web browsers such as Google Chrome and Mozilla Firefox are often perceived as solutions to enhance user privacy. However, these modes still leave traces of sensitive data in volatile memory (RAM), even though artifacts stored on disk-based storage are removed. This study evaluates the effectiveness of private browsing modes using the National Institute of Standards and Technology (NIST) framework integrated with Artificial Intelligence (AI) for forensic analysis. Simulation scenarios were conducted to assess the ability of private browsing modes to prevent data retention. The results indicate that although private browsing modes successfully eliminate disk-based traces, sensitive data such as account credentials can still be extracted from RAM. The integration of AI accelerates the detection of these artifacts. This research contributes to the field of digital forensics by providing a systematic framework for evaluating browser privacy mechanisms and offering insights for the development of real-time browser security tools.

Prophet with Google Trends for Forecasting Train Passengers in Java

Kiki Ferawati, Winita Sulandari, Nur Arina Bazilah Kamisan — Wed, 15 Apr 2026 00:00:00 +0000

As a popular transportation method for long-distance travel, trains were also a preferred choice during the homecoming period before Eid Al-Fitr, one of the major religious holidays in Indonesia. During this period, known locally as ‘mudik,’ millions of people travel from the urban cities back to their hometowns to celebrate with their families, creating a significant surge in transportation demand. However, since the holiday follows the Islamic calendar, which changes slightly every year, forecasting train passengers becomes tricky, thus requiring a different approach to achieve accurate predictions. This study utilizes the Prophet method to forecast train passengers in Java (excluding the Jabodetabek area) using the data from 2006 to 2024. We also incorporated the COVID-19 period as a fixed external regressor, along with external regressors from Google Trends data using the keywords ‘kereta api’, ‘mudik’, and ‘lebaran’, which are commonly searched by the public in relation to train travel and the Eid homecoming period. The results on the test set, 2024 data, showed that the word ‘mudik’ was the most effective in improving forecast accuracy, with a MAPE of 9.12 and RMSE of 797.76, a decrease of 11.57% and 9.34% compared to the updated baseline. This indicates that public search behavior around the term ‘mudik’ closely aligns with actual travel demand patterns. The findings of this study suggest that Prophet with external regressors are capable of forecasting train passengers and Google Trends can be a valuable addition for capturing data patterns related to specific phenomenon.

Exploring Ensemble Architectures on Lung X-Ray Multi-Class Image for Classification Using Convolutional Neural Network and Random Forest

Devin Garmenta Nuriansyah, Putu Desiana Wulaning Ayu, Dandy Pramana Hostiadi — Wed, 15 Apr 2026 00:00:00 +0000

The lungs are vital organs that play an important role in the respiratory and circulatory systems. Early detection of lung diseases through medical images, especially Chest X-Ray (CXR), is still a challenge due to the limited amount of data and complexity in image interpretation. This research aims to develop an effective image classification approach for lung disease detection by comparing two main methods: direct training using Convolutional Neural Network (CNN) and a hybrid method involving feature extraction from CNN model, feature selection using Chi-Square method, and classification using Random Forest algorithm.

To overcome data imbalance and increase variation, data augmentation techniques such as rotation, vertical and horizontal flipping, and zooming are used. Four popular CNN architectures are used in training, namely VGG16, ResNet-50, InceptionV3, and MobileNet. After training, features are extracted and stored in .csv format. Next, feature selection using the Chi-Square method and classification with Random Forest are performed.

The experimental results show that direct CNN training achieves high accuracy, with MobileNet reaching the highest performance at 98.83%. However, this approach requires significant computational resources and longer training time. In contrast, the hybrid method offers competitive accuracy with lower computational demands. The findings highlight the potential of combining deep learning and traditional machine learning to create efficient, accurate, and resource-friendly medical image classification systems. This research has significant implications for supporting early diagnosis of lung diseases, reducing diagnostic workload for medical professionals, and enabling the development of deployable AI-assisted healthcare solutions in resource-limited settings.

Deep Convolutional Generative Adversarial Network-Enhanced Data Augmentation for Imbalance Facial Acne Severity Classification Using a Fine-Tuned EfficientNet-B1

Khoirun Nisya, Sugiyarto Surono, Aris Thobirin — Thu, 16 Apr 2026 00:00:00 +0000

Imbalanced datasets often hinder the generalization capability of Convolutional Neural Networks (CNNs) in medical image classification, leading to overfitting and reduced performance on minority classes. This study aims to develop an acne severity classification model using EfficientNet-B1 combined with geometric and photometric augmentation, as well as and Deep Convolutional Generative Adversarial Network (DCGAN)-based augmentation to address class imbalance. The dataset consists of 1,380 facial images categorized into four acne severity levels: Normal, Level 0, Level 1, and Level 2. Preprocessing includes RGB conversion, bilinear resizing, and center cropping. The data are split into training (80%), validation (10%), and testing (10%) sets. Geometric and photometric augmentation applies horizontal flipping, 45° rotation, color jittering, and random resized cropping, while DCGAN generates synthetic samples to balance minority classes. The EfficientNet-B1 model is fine-tuned using compound scaling, MBConv blocks, Swish activation, Batch Normalization, Cross-Entropy loss, and AdamW optimizer, with 5-fold cross-validation for robustness. Experimental results demonstrate that DCGAN-based augmentation achieves superior performance, with a test accuracy of 94% and an average F1-score of 0.93, outperforming geometric and photometric data augmentation (90% accuracy and 0.88 F1-score). DCGAN augmentation also significantly reduces misclassification between visually similar acne severity levels, particularly Level 0 and Level 1. These findings indicate that integrating DCGAN with EfficientNet-B1 effectively enhances generalization on imbalanced medical image datasets, providing a robust and replicable framework for acne severity classification and related medical imaging applications.

Geodetically-Enhanced Hybrid GRU with Adaptive Dropout and Dynamic L2 Regularization for Earthquake Parameter Prediction in Indonesia

Najmuddin Mubarak MR, Susandri Susandri, Ahmad Zamsuri — Wed, 15 Apr 2026 00:00:00 +0000

Earthquake prediction remains challenging due to the nonlinear behavior and uncertainty of seismic activity. This study introduces a geodetically-enhanced hybrid GRU model integrating adaptive dropout and dynamic L2 regularization to improve robustness and accuracy in earthquake magnitude prediction. In addition to seismic sequence data, slip-rate values derived from scalar moment distribution were incorporated as a domain-informed feature to represent tectonic strain accumulation across Indonesia. The dataset consisted of BMKG records from 2010–2025 and was processed through outlier removal, normalization, temporal reshaping, and feature integration. The proposed model was evaluated against multiple deep learning baselines including CNN-1D, LSTM, standard GRU, Transformer-based models, and Neural ODE architectures. Performance assessment used RMSE, MAE, and R² metrics. The resulting hybrid GRU achieved improved predictive accuracy with an RMSE of 0.5176, MAE of 0.3973, and an R² score of 0.5997, outperforming both CNN-1D and standard GRU baselines. The integration of slip-rate features contributed to reduced prediction variance across tectonically active zones. These findings demonstrate that combining geodetic information with adaptive regularization strategies improves generalization and model stability for seismic forecasting. The approach offers potential applicability for rapid early-warning scenarios requiring low latency and reliable prediction accuracy.

Optimizing Heart Disease Classification Using C4.5, Random Forest, and XGBoost with ANOVA, Chi-Square, and AdaBoost

Andika Pratama, Setiawan Assegaff, Jasmir Jasmir, Nurhadi Nurhadi — Wed, 15 Apr 2026 00:00:00 +0000

Heart disease remains one of the leading causes of mortality worldwide, underscoring the need for accurate and scalable prediction models within clinical informatics. This study proposes a leakage-safe machine learning pipeline combining stratified splitting, SMOTE-based imbalance handling, and in-fold feature selection using ANOVA, Chi-Square, and AdaBoost-assisted ranking to enhance classification performance on a large heart-disease dataset consisting of 10,000 samples and 21 attributes. Three widely used algorithms, C4.5, Random Forest, and XGBoost, were evaluated to determine the optimal model-feature selection configuration for structured medical data. The results demonstrate that feature relevance contributes more significantly to predictive performance than increasing model complexity, with Random Forest achieving the highest accuracy, precision, recall, and F1-Score at 98.43% when combined with Chi-Square or ANOVA feature selection. C4.5 showed the greatest relative improvement, rising from 76.52% to 97.57% using AdaBoost-assisted selection, while XGBoost improved from 66.32% to 94.88% after statistical filtering. The dominant features identified such as CRP, BMI, blood pressure, fasting glucose, LDL, triglycerides, and homocysteine align with well-established cardiovascular biomarkers, supporting clinical validity. This research provides an important contribution to computer science by demonstrating an efficient and scalable hybrid FS-boosting framework capable of reducing unnecessary model complexity, improving generalization, and supporting low-latency deployment in clinical decision-support systems. The findings highlight the potential of structured-data machine learning to strengthen digital health diagnostics in resource-limited environments.

Comparative Analysis of IndoBERT and mBERT for Online Gambling Comment Detection in Indonesian Social Media

Mon, 20 Apr 2026 00:00:00 +0000

The rapid growth of illegal online gambling promotions in Indonesian social media comments requires automated detection systems capable of handling informal and noisy text. This study aims to evaluate the effectiveness of Transformer-based language models for detecting online gambling-related comments in Indonesian Twitter and YouTube data. Two pre-trained models, IndoBERT and mBERT, were fine-tuned and compared using a labeled dataset consisting of gambling and non-gambling comments. Model performance was evaluated using accuracy, precision, recall, and F1-score. Experimental results show that IndoBERT achieved 98% accuracy and F1-score, outperforming mBERT, which achieved 96% on the same dataset. Additionally, performance was compared against a recurrent neural network baseline to validate the effectiveness of Transformer-based architectures. The findings demonstrate that language-specific pre-training provides measurable advantages for detecting domain-specific content in Indonesian social media. This study contributes empirical evidence supporting the application of Transformer models for automated moderation of harmful online content in Indonesian digital platforms.

A Comparative Study of Generalized Linear Mixed Model and Mixed Effects Random Forest for Analyzing Data with Outliers

Reza Arianti, Khairil Anwar Notodiputro, Yenni Angraini — Wed, 15 Apr 2026 00:00:00 +0000

This study compares MERF and GLMM-NB in analyzing hierarchical data and focusing on the role of residual outliers and the application of winsorization. A two-stage analytical pipeline was implemented: (1) winsorization to reduce extreme residual values, and (2) model training using MERF and GLMM-NB. The dataset comes from the 2021 National Socio-Economic Survey (Susenas) in West Java Province, measuring tobacco consumption intensity. Two statistical approaches are compared, MERF and GLMM with a Negative Binomial distribution (GLMM-NB). Models were trained under two conditions: without winsorization (WIN0) and with two-sided 5% winsorization (WIN5). Winsorization was applied to the training data, and the test data were adjusted using thresholds from the training set. Model performance was assessed using Root Mean Squared Error (RMSE) and the train-test ratio. Under WIN0, GLMM recorded an RMSE of 49.65 for training and 42.27 for testing, while MERF achieved 35.96 and 39.94, respectively. After WIN5, GLMM showed a larger error reduction, with RMSE values of 34.90 (train) and 30.20 (test), while MERF dropped to 26.63 (train) and 28.64 (test). These results indicate that MERF provides higher predictive accuracy, whereas GLMM benefits more from winsorization. Household expenditure, employment status, age, and gender consistently emerged as key variables linked to tobacco consumption intensity. This study is the first to compare MERF and GLMM-NB with winsorization using Indonesia’s hierarchical data. The analytical framework helps inform public health policies aligned with SDG 3: Good Health and Well-being, particularly in reducing tobacco-related health risks.

Adaptive Gradient Boosting for Fuel Consumption Prediction in Mining Haul Trucks under Concept Drift Monitoring

Kusnawi, Mochamad Agung Wibowo , Ridwan Sanjaya — Sat, 18 Apr 2026 00:00:00 +0000

Fuel consumption prediction models deployed in mining operations often degrade in performance due to changes in the distribution of high-frequency telemetry data, a phenomenon commonly associated with concept drift. Static machine learning models trained on historical data may therefore lose reliability over time in dynamic operational environments. This study aims to develop an adaptive regression approach for predicting fuel consumption in mining haul trucks by integrating a Gradient Boosting Regressor with batch-wise performance monitoring and periodic retraining. Real-world telematics data were processed through systematic preprocessing and feature engineering to derive behavioral and operational indicators relevant to fuel usage. Model performance was evaluated using Mean Absolute Error (MAE), Mean Squared Error (MSE), and the coefficient of determination (R²), while drift monitoring employed a threshold-based MAE analysis over streaming batches. Experimental results show that the initial model achieved an MAE of 27.27 L/h and an R² of 0.759, and the adaptive retraining strategy provided marginal yet consistent performance stabilization without detecting significant drift within the observed period. Beyond the mining application, this framework contributes to the development of lightweight adaptive regression systems for real-time data stream processing, supporting computationally efficient predictive maintenance in industrial IoT environments.

Early Detection of Depression Levels Among Gen-Z Using TikTok Data and Extra Trees Ensemble Classifier

Achmad Solichin, Helmi Zulqan, Painem, Anindya Putri Pradiptha — Wed, 15 Apr 2026 00:00:00 +0000

Mental health disorders, particularly depression, have become an increasingly critical issue, especially among young people aged 15–29 years. Social stigma and limited awareness often hinder early detection and intervention. In the digital era, social media platforms such as TikTok provide opportunities to observe users’ behavioral patterns that may reflect their psychological conditions. This study proposes an early depression detection model based on TikTok social media data using an ensemble machine learning approach, namely the Extra Trees classifier. Data were collected from 263 undergraduate students through an online survey combined with automated crawling of respondents’ TikTok accounts. Depression levels were labeled using the Patient Health Questionnaire-9 (PHQ-9) and categorized into four classes: none, mild, moderate, and severe. After data selection, feature extraction, and class balancing using SMOTE, the final dataset consisted of 600 instances with 24 features, including demographic attributes, TikTok activity metrics, and social network analysis features. Experimental results indicate that the Extra Trees classifier achieved the highest performance, with an accuracy, precision, recall, and F1-score of 91%, outperforming Decision Tree, Random Forest, XGBoost, LightGBM, and CatBoost. The model demonstrated stable performance across all depression levels and efficient prediction time suitable for near real-time web-based applications. These findings confirm that integrating behavioral and network-based social media features with validated psychological assessments can support effective early depression screening. This research contributes to mental health informatics and social media analytics within the field of computer science by demonstrating the effectiveness of ensemble learning for depression detection using TikTok-based digital behavioral data.

User Acceptance Analysis of SINAGA Digital Attendance System Using Integrated UTAUT and SCT Models with PLS-SEM for Civil Servants in Purbalingga Regency

Imam Sofarudin Latif, Rujianto Eko Saputro, Azhari Shouni Barkah — Sat, 18 Apr 2026 00:00:00 +0000

This study combines two main theories, namely the Unified Theory of Acceptance and Use of Technology (UTAUT) and Social Cognitive Theory (SCT), to analyze the level of user acceptance of the SINAGA digital attendance system among civil servants in Purbalingga Regency. This study aims to identify factors that influence technology adoption through an integrated UTAUT approach with SCT moderation, particularly self-efficacy. The method used was a survey of 102 respondents, with analysis using Partial Least Squares-Structural Equation Modeling (PLS-SEM) involving testing of outer and inner models through the Slovin approach. The results show that factors such as Performance Expectancy (PE), Effort Expectancy (EE), Social Influence (SI), and Facilitating Conditions (FC) significantly influence Behavioral Intention (BI). Self-Efficacy (SE) and Outcome Expectancy (OE) also act as moderating factors that strengthen the relationship between PE and BI, as well as EE and BI. With an R² value of 78%, this model has a high explanatory power regarding users' behavioral intentions in adopting the system. This study contributes to the development of technology acceptance theory in the public sector, particularly for e-government systems, and suggests improving users' digital competence and optimizing infrastructure to support further technology acceptance with the integration of artificial intelligence (AI) technology in the system for more efficient dynamic monitoring. The main contribution of this research is the development of digital systems within the Indonesian government, in line with the sustainability of technology adoption in the public sector.

Comparative Analysis of GPT-2 Augmentation, ALBERT, and Similarity Measures for Cyberbullying Detection

Zidane Hidayat, Hasan Dwi Cahyono, Fajar Muslim — Wed, 15 Apr 2026 00:00:00 +0000

The effectiveness of cyberbullying detection is influenced by the availability of sufficient, diverse, and contextually rich training data, which is often limited in low-resource languages such as Indonesian. To address dataset limitations, researchers have extensively explored data augmentation (DA) as a promising approach to improving model performance. DA generates new data instances by applying transformations to existing data, thereby increasing both dataset size and variability. Prior studies have demonstrated that applying Easy Data Augmentation (EDA) with Support Vector Machine (SVM) classification improved cyberbullying detection performance, even when it faced challenges in capturing semantic and contextual nuances. In this paper, we investigated Indonesian DA methods using the Transformer-based GPT-2 model. The augmented sentences were evaluated and filtered based on context, semantics, diversity, and novelty, with similarity measures such as Euclidean Distance (ED), Cosine Similarity (CS), Jaccard Similarity (JS), and BLEU Score (BLS) ensuring the quality of the augmentation. Furthermore, we compared text classification performance using both SVM and the Transformer-based ALBERT model. Experimental results revealed that incorporating similarity measures and GPT-2 as a DA method failed to improve cyberbullying detection performance, potentially due to the semantic drift introduced by GPT-2 and the inadequacy of similarity measures in capturing nuanced contextual information. However, we found that ALBERT outperformed SVM as a classification model, achieving average F1-scores of 91.77% and 91.72%, respectively. This study contributes to the informatics field by exploring the potential of Transformer-based augmentation and similarity evaluation in enhancing low-resource text classification, while acknowledging the limitations in data quality and model adaptation.

Balinese Statue Image Classification Using Transfer Learning: A Comparative Study of MobileNetV3 and EfficientNetV2

Thu, 16 Apr 2026 00:00:00 +0000

Balinese sculpture is an important form of cultural heritage that exhibits high visual diversity in terms of shape, structure, and carving style, which makes manual identification and documentation challenging. Previous studies on automated statue classification have generally focused on limited sculpture categories and therefore do not fully represent the visual diversity of Balinese sculptures. This study aims to develop an automatic image classification model capable of recognizing multiple Balinese statue categories using transfer learning and fine-tuning strategies. The proposed approach compares two convolutional neural network architectures, MobileNetV3 and EfficientNetV2, across eight statue classes: Dewa, Dewi, Mitologi, Penabuh, Pengapit, Punakawan, Raksasa, and Wanara. A dataset of 8,400 images was constructed from three-dimensional video documentation to capture multiple viewing angles of each statue. The images were processed through frame extraction, resizing, normalization, data augmentation, and dataset splitting. Model training was conducted in two stages, consisting of transfer learning followed by fine-tuning using reduced learning rates. Experimental results indicate that both models achieve high classification performance on the test dataset. MobileNetV3 obtained the highest test accuracy of 99.64% with a loss value of 0.0119, while EfficientNetV2 achieved an accuracy of 98.56% with a loss of 0.0613. These findings demonstrate that lightweight architectures can deliver competitive performance when supported by appropriate training strategies. This study provides a comparative evaluation of efficient deep learning models for cultural heritage image classification and supports the development of more reliable and systematic digital documentation of Balinese sculptures.

Enhancement of YOLOv9 Model for Traffic Vehicle Detection using Augmentation Techniques

Imam Ahmad Ashari, Wahyul Amien Syafei, Adi Wibowo — Wed, 15 Apr 2026 00:00:00 +0000

Traffic vehicle detection is a crucial component in developing intelligent transportation systems, with object detection models like YOLO (You Only Look Once) often preferred for their speed and accuracy. However, challenges remain in detecting vehicles under diverse lighting conditions and small object scales, even with advanced models such as YOLOv9. To address these limitations, image augmentation techniques are employed to enhance model robustness by providing broader data variation. This study investigates the impact of multiple image augmentation methods on the YOLOv9t model for traffic vehicle detection. The techniques evaluated include Blur, Brightness Adjustment, Contrast Adjustment, Color Jitter, Cropping, Flipping, Noise Injection, Rotation, Scaling, and Zoom-In. Results reveal that Scaling and Brightness Adjustment significantly improve detection accuracy, achieving mAP50-95 values of 0.450 and 0.449, respectively. Conversely, methods such as Contrast Adjustment, Rotation, and Cropping produced unsatisfactory outcomes, with Contrast Adjustment performing the worst at only 0.167. Without augmentation, the baseline mAP50-95 was 0.378, emphasizing the vital role of augmentation in improving detection performance, especially under challenging conditions. These findings highlight the importance of selecting appropriate augmentation techniques to optimize YOLOv9t performance, with further improvements possible through combining multiple methods. Compared to approaches that solely focus on enhancing model architecture, the proposed augmentation-based strategy proves more effective in addressing real-world challenges, strengthening resilience against lighting variations and small object detection. This contribution supports the development of more accurate and reliable multilabel vehicle detection systems, advancing safer and more efficient intelligent transportation solutions.

Enhancing Diagnostic Accuracy of Polycystic Ovary Syndrome Classification in Ultrasound Images Using a Hybrid Deep Learning Model of VGG16 and AlexNet

Wed, 15 Apr 2026 00:00:00 +0000

Diagnosis of Polycystic Ovary Syndrome (PCOS) using ultrasound (USG) imaging still faces a major challenge in the form of inter-observer variability, which can lead to inconsistent diagnostic outcomes and increase the risk of misclassification. This limitation highlights the urgent need for an automated artificial intelligence (AI)–based system capable of performing ultrasound image classification with greater objectivity, accuracy, and consistency. This study aims to develop an automated PCOS classification model based on a hybrid Convolutional Neural Network (CNN) architecture that integrates VGG16 and AlexNet through a feature concatenation mechanism, following preprocessing and data augmentation steps to enhance model generalization. The model’s performance was evaluated using accuracy, precision, recall, F1-score, and specificity as key metrics. Experimental results demonstrate that the VGG16–AlexNet hybrid model achieved the best performance, with an accuracy of 98.26%, precision of 97.90%, recall of 97.90%, F1-score of 97.90%, and specificity of 98.52%. These results outperform other hybrid configurations such as VGG16–MobileNetV2, VGG16–ResNet50, and VGG16–InceptionV3, each of which achieved accuracies above 96%. These findings confirm that combining the feature depth of VGG16 with the computational efficiency of AlexNet enables more comprehensive extraction of spatial and textural patterns in ultrasound images. Consequently, the proposed hybrid model offers a promising AI-driven diagnostic support system that not only enhances the accuracy of PCOS detection but also assists clinicians in making faster, more objective, and consistent medical decisions.

Optimization of MobileNet SSD Using Pruning, Quantization, and Transfer Learning for Real-Time Vehicle Detection in IoT-Based Security Systems

Afit Miranto, Purwono Prasetyawan, Iqbal May Aryanto — Wed, 15 Apr 2026 00:00:00 +0000

Security is a critical requirement in modern public and private environments, especially in systems that rely on resource-constrained IoT devices. This research aims to optimize the MobileNet SSD (Single Shot MultiBox Detector) model to achieve fast and reliable real-time vehicle and human detection on low-power hardware. The proposed optimization pipeline integrates three techniques: pruning to reduce network redundancy, quantization to accelerate inference and decrease memory usage, and transfer learning using six relevant object classes (person, car, motorcycle, bicycle, bus, and truck). Experiments were conducted on a Raspberry Pi 5 equipped with a camera and local dashboard interface. The optimized MobileNet SSD v2 model achieved a mean Average Precision (mAP) of 0.724 and mAP@0.5 of 0.951, while improving inference speed from 21 FPS to over 24 FPS. These results indicate a balanced trade-off between accuracy, speed, and resource efficiency, enabling stable real-time performance on constrained IoT platforms. The findings contribute to the body of knowledge in embedded and edge AI by demonstrating how integrated model-level optimization can significantly enhance deep learning inference on low-power systems, offering scientific and practical implications for smart surveillance and intelligent traffic monitoring.

A Locally Grounded Retrieval-Augmented LLM-Based Chatbot for Bilingual Stunting Prevention Consultation among Health Cadres in Indonesia

Wed, 15 Apr 2026 00:00:00 +0000

Stunting remains a major public health challenge in Indonesia, affecting 21.6% of children under five nationally and 18.34% in Nusa Tenggara Barat (NTB), which strains the capacity of health cadres to deliver timely and accurate nutrition education. This study aims to develop a consultation chatbot by integrating Large Language Models (LLMs) with Retrieval-Augmented Generation (RAG) to provide context-aware stunting prevention guidance. A total of 45 journal articles and 7 books were curated to construct 7,642 question–answer pairs using a RAG-based pipeline. Text preprocessing involved segmentation, embedding, and Byte Pair Encoding tokenization, followed by fine-tuning a LLaMA 3 model on an NVIDIA L4 GPU. Model performance was evaluated using ROUGE and BERTScore metrics, complemented by a small pilot usability assessment. The RAG-integrated model achieved a ROUGE-1 score of 81.03% and a BERTScore F1 of 93.48%, consistently outperforming baseline models. These findings demonstrate the potential of RAG-enhanced LLMs to support scalable and accessible health informatics solutions for empowering health cadres in resource-limited and rural settings.

Optimizing YOLO11 for Dense Crowd Counting under Severe Occlusion via Head-Detection Fine-Tuning

Joko Sutrisno, Sri Winarno , Affandy — Sat, 18 Apr 2026 00:00:00 +0000

Accurate and real-time people counting is essential for crowd management and public safety, yet achieving precision in high-density environments remains a challenge due to severe visual occlusion. While the recently released YOLO11 architecture introduces advanced features such as C3k2 and C2PSA modules, its performance as a pre-trained model for people counting tasks has not been fully explored. This study evaluates the efficacy of a head-detection-based fine-tuning strategy using the YOLO11 model, compared against the default pre-trained baseline. The fine-tuning performance is analyzed across three distinct scenarios: S1 (full fine-tuning at 960 pixels), S2 (partial backbone freezing at 960 pixels), and S3 (partial freezing at 640 pixels). The fine-tuning process was conducted using the CC_Mach_1 dataset from Roboflow Universe, which consists of high-density images annotated for head detection. The results demonstrate that the baseline pre-trained YOLO11, which relies on full-body features, exhibits extremely limited performance with an mAP@0.5 of 0.017 and a Mean Absolute Error (MAE) of 100.3. In contrast, the fine-tuned scenarios achieved substantial improvements, led by S1 which reached the highest accuracy with an mAP@0.5 of 0.682 and reduced the MAE by 62% to 37.8. While S2 remained highly competitive with an MAE of 39.6, the performance in S3 declined to 46.9, confirming that lower input resolutions limit the model's ability to identify small-scale head features. These findings provide empirical evidence that domain-specific fine-tuning for head detection substantially improves the robustness of YOLO11 against occlusion. Beyond technical accuracy, this detection-based approach offers a more computationally efficient alternative to traditional density-map-based methods, making it highly suitable for deployment in real-time surveillance systems for large-scale public monitoring.

Optimizing E-commerce Personalization through Hybrid Decision Tree–Nearest Neighbor Recommendation Integration

Wed, 15 Apr 2026 00:00:00 +0000

Single-method recommendation systems face critical limitations: content-based filtering suffers from overspecialization while collaborative filtering struggles with data sparsity and cold-start problems. This research introduces an innovative hybrid recommendation framework that synthesizes Content-Based Filtering (CBF) utilizing Decision Trees with Collaborative Filtering (CF) employing Nearest Neighbor algorithms. Our approach addresses the inherent limitations of singular recommendation methodologies by integrating product attribute analysis with collective user behavior patterns. We conducted comprehensive evaluations using a shopping behavior dataset comprising 3,900 consumer records with diverse demographic and product interaction data. Our findings reveal that an asymmetric hybrid configuration—weighted at 70% for CBF and 30% for CF—achieves optimal performance with a Root Mean Square Error (RMSE) of 0.7422. The system incorporates an interactive user interface that facilitates a natural shopping experience: browsing available items, receiving personalized recommendations, and providing explicit feedback on suggested products. Through feature importance analysis, we identified key product attributes that significantly influence recommendation quality, including size variations and specific color preferences. The hybrid approach demonstrates 42% greater category diversity and 37% more recommendation diversity compared to pure content-based filtering, while maintaining superior accuracy metrics. Our research contributes to understanding optimal hybrid architectures and provides practical insights for implementing effective personalization strategies in real-world e-commerce environments.

Benchmarking Brain-Training Apps Using DEGREE and Fuzzy Logic: Lumosity vs Elevate

Heri Sismoro, Mei Parwanto Kurniawan — Sat, 18 Apr 2026 00:00:00 +0000

This study provides an actionable benchmark of two popular brain-training apps—Lumosity and Elevate—by applying the 14-factor DEGREE framework as a structured UX evaluation tool and using fuzzy scoring to improve interpretability. We recruited 190 Computer Science undergraduates; each participant evaluated both apps, yielding 380 app evaluations using a counterbalanced two-sheet questionnaire. Fourteen factors covering usability, engagement, and perceived learning were rated on a five-point Likert scale. Reliability was strong for both apps (Cronbach’s α = 0.822 for Lumosity; 0.847 for Elevate). Descriptive results showed mid-to-high perceptions overall, with mean scores of 3.51 (Lumosity) and 3.44 (Elevate). Fuzzy aggregation transformed subjective ratings into a 0–1 index, producing overall scores of 0.503 (Lumosity) and 0.490 (Elevate), indicating a small global advantage for Lumosity. At the factor level, Lumosity was slightly higher on most DEGREE dimensions, whereas Elevate showed relative advantages on Learnability and Confidence, suggesting potential benefits for early onboarding and self-efficacy. Overall, the proposed DEGREE–Fuzzy pipeline yields a transparent, reproducible benchmark that translates multi-factor perceptions into decision-ready recommendations for selecting apps aligned with instructional goals.

Sentiment Analysis Using Bidirectional Encoder Representations from Transformers for Indonesian Stock Price Prediction with Long Short-Term Memory and Gated Recurrent Unit Models

Dwi Utari Iswavigra, Very Dwi Setiawan, Mutia Ulfa, Brieva Ommr — Wed, 15 Apr 2026 00:00:00 +0000

The advancement of artificial intelligence based market analytics has driven the need for stock price prediction models capable of representing market behavior both technically and psychologically. This study aims to improve stock price forecasting in the Indonesian capital market by integrating sentiment analysis with deep learning time-series models. It evaluates whether public sentiment can contribute to enhancing prediction accuracy when combined with historical stock data. Textual sentiments were extracted using IndoBERT and converted into positive, negative, and neutral scores, which were then merged with historical stock prices. These data were modeled using LSTM, GRU, and a hybrid LSTM–GRU architecture. Model evaluation was conducted using MSE, MAE, RMSE, and MAPE metrics across six Indonesian stocks ANTM, BBCA, BBRI, SCMA, TLKM, and UNVR. The hybrid LSTM–GRU model produced the lowest prediction errors for BBCA and BBRI, with MSE scores of 0.151 and 1022.062, respectively. GRU delivered the best performance for highly volatile stocks, such as SCMA MAPE 1.65% and UNVR MAPE 0.51%, while LSTM demonstrated the most stable performance for TLKM with an MSE of 606.93 and RMSE of 24.63. Across all cases, sentiment scores improved model responsiveness, particularly during price spikes ANTM mid-2025 and price declines BBRI early year. The integration of sentiment significantly enhances prediction relevance by combining psychological market indicators with technical price trends. This framework provides more reliable decision-making support for investors, strengthens algorithmic trading strategies in Indonesia, and contributes to intelligent financial analytics that reflect local market behavior.

Cloud Computing-Based U-Net Integration for Post-Landslide Satellite Image Segmentation

Wed, 15 Apr 2026 00:00:00 +0000

Landslides are geological disasters that cause severe impacts on human life, infrastructure, and ecosystems, highlighting the need for post-disaster mapping methods that are fast, accurate, and scalable. This study aims to develop a post-landslide satellite image segmentation framework based on U-Net integrated with cloud computing to support large-scale and operational disaster mapping. While U-Net has been widely applied for landslide analysis, most existing studies focus on local-scale assessments or susceptibility mapping and lack integration with cloud-based pipelines and multi-source data for post-disaster operations. The novelty of this research lies not in modifying the U-Net architecture, but in integrating multi-source geospatial data, system workflow, and scalable cloud deployment. The proposed framework utilises a global multi-source dataset consisting of RGB imagery, Normalized Difference Vegetation Index (NDVI), slope, and elevation to enhance robustness and generalisation across diverse geomorphological conditions. Experimental results show stable model convergence with a final loss of 0.0357, an F1-score exceeding 0.75, and an AUC-PR of 0.8391. Evaluation on the testing dataset achieves a precision of 0.7692, recall of 0.7519, F1-score of 0.7604, and Intersection over Union of 0.6135. Qualitative analysis demonstrates strong spatial agreement between predicted segmentation and ground truth, with minor deviations mainly along complex slope boundaries. From an Informatics perspective, this study contributes by operationalizing deep learning through cloud computing to enable scalable computation, parallel processing, and system-level deployment, while providing object-level estimates of landslide events and affected areas to support disaster response and risk mitigation.

Deep Learning-Based Recognition of Indonesian Sign Language (BISINDO) Alphabetic Gestures Using Skeletal Feature Extraction and LSTM

Teuku M Arief Afwan, Rahmat Gernowo, Helmie Arif Wibawa — Wed, 15 Apr 2026 00:00:00 +0000

Communication is a fundamental aspect of human life, and for the deaf community, sign language serves as the primary medium of interaction. In Indonesia, the Indonesian Sign Language (BISINDO) is widely used, however, research on automatic BISINDO recognition remains limited due to the scarcity of representative datasets. This study presents the development of a BISINDO recognition system based on deep learning by integrating the Long Short-Term Memory (LSTM) architecture with the MediaPipe Holistic framework. To address data limitations, a custom dataset comprising 866 BISINDO alphabetic gesture videos was collected, involving recordings from both expert and non-expert signers to capture stylistic variations. Extracted skeletal landmark features were processed through a three-layer LSTM network followed by dense layers for sequential modeling and classification. Experimental results show that the proposed model achieved a validation accuracy of approximately 93%, outperforming static image–based methods and demonstrating the effectiveness of skeletal features in representing dynamic gestures. The model also exhibited real-time applicability with promising performance, although challenges such as misclassification of visually similar gestures and dataset imbalance remain. This study contributes to the underexplored field of BISINDO recognition by providing a baseline system and dataset, and further advances the domains of computer vision and human–computer interaction within informatics through an inclusive, data-driven framework for Indonesian Sign Language recognition and future AI-assisted accessibility technologies.

Implementation of Agile Method and Apriori Algorithm for Recommendation System in Outdoor Equipment Rental Services

Raditya Prama Elfreda, Muhammad Lulu Latif Usman — Sat, 18 Apr 2026 00:00:00 +0000

Drop Outdoor Purwokerto faces inefficiencies in its outdoor equipment rental process, where customers are required to visit the store directly to check item availability, often resulting in miscommunication and suboptimal transaction management. To address this issue, this study aims to design and develop a web-based outdoor equipment rental information system that enables real-time availability checking and efficient online booking. The system is developed using the Agile methodology to accommodate dynamic user requirements and iterative system improvements. In addition, the Apriori algorithm is implemented to analyze historical rental transaction data and generate item recommendations based on association rule mining. The analysis results indicate that several outdoor equipment items exhibit strong association patterns, with the highest lift value exceeding 1, signifying meaningful relationships beyond random co-occurrence. These patterns are utilized as the basis for the recommendation feature within the system. Functional testing using Black Box Testing shows that all system features operate as expected, achieving a 100% success rate across tested scenarios, including transaction processing, cart management, and recommendation display. The findings demonstrate that integrating the Agile development approach with Apriori-based data mining can effectively support data-driven decision-making in outdoor equipment rental services. This study contributes to the development of recommendation systems for small and medium-sized rental businesses by highlighting the practical application of association rule mining on rental transaction data, which exhibits characteristics distinct from conventional retail datasets.

Optimizing Automatic Irrigation Duration for Grapevines in Greenhouses Using Multiple Linear Regression Analysis

Kharisma Monika Dian Pertiwi, Trenady Alfarabi — Wed, 15 Apr 2026 00:00:00 +0000

Greenhouses offer a controllable microclimate for high‑value horticulture, yet manual irrigation and single‑sensor threshold rules remain inefficient and error‑prone for grapevine cultivation in tropical conditions. This study designs and implements an Internet‑of‑Things (IoT) automatic irrigation system that employs an interpretable multiple linear regression (MLR) model as the decision core, using air temperature and soil moisture—acquired via DHT11 and capacitive soil‑moisture sensors—to estimate irrigation duration in real time. The model is trained on greenhouse measurements and deployed for low‑latency edge inference to actuate valves with duration‑to‑volume conversion, enabling precise and adaptive water delivery. Experimental evaluation shows strong predictive performance (MSE = 0.15, MAPE = 1.44%, R² = 0.98), indicating high accuracy and reliable generalization for operational control. The primary contributions are: (i) a lightweight, explainable regression formulation tailored to tropical grapevines that outperforms single‑parameter baselines; (ii) an end‑to‑end, edge‑deployable IoT pipeline that reduces computational and energy costs while maintaining real‑time autonomy; and (iii) an engineering blueprint that is scalable and maintainable for smallholder contexts. The impact for Informatics/Computer Science lies in demonstrating a practical ML‑on‑the‑edge reference design—combining interpretable modeling, sensor fusion, and actuation—that advances sustainable computing for precision agriculture, improves resource efficiency, and supports robust, replicable deployment of smart‑irrigation systems in data and power‑constrained environments.

Explainable Ensemble Learning for Depression Risk Classification Using Multidomain Behavioral Features

Erfian Junianto, Siti Nurkhodijah — Wed, 15 Apr 2026 00:00:00 +0000

Depression is a growing global health concern, particularly among adolescents and university students. Despite the availability of standardized assessments, delays in early detection remain a major barrier to effective treatment. Digital behavioral data holds considerable potential for mental health assessment, but its utilization remains limited due to the absence of integrated and interpretable computational models. This study presents an interpretable machine learning framework for classifying depression risk using multi-domain behavioral features extracted from simulated digital life datasets. Three public datasets were integrated and mapped to five psychological clusters based on DSM-5 criteria: self-regulation, negative affect, cognitive strain, comparison and avoidance, and sleep disturbance. Two ensemble classifiers, Random Forest and XGBoost, were applied and evaluated using 10-fold stratified cross-validation. Depression risk was categorized into three levels: Low, Medium, and High. The Random Forest model achieved the highest accuracy (81%) and macro-averaged F1-score (0.81), showing strong performance especially in identifying transitional Medium-risk users. To enhance transparency, both global and local model interpretations were performed using SHapley Additive exPlanations (SHAP). Results revealed that digital stressors such as excessive screen time and disrupted sleep patterns were prominent in high-risk classifications, while mood stability and mindfulness were protective factors in low-risk groups. The proposed framework offers a scalable and explainable for early depression screening by integrating psychological theory with artificial intelligence methods. The findings contribute to the field of behavioral informatics by demonstrating the practical value of interpretable models in enhancing the reliability, transparency, and applicability of digital mental health systems and personalized behavioral monitoring.

Mixed-Data K-Means Clustering with Hyperparameter-Tuned Random Forest for OSS-Based MSME Investment Profiling and Policy Targeting

Wed, 15 Apr 2026 00:00:00 +0000

Administrative data of Micro, Small, and Medium Enterprises collected through the Online Single Submission system are highly heterogeneous, combining numerical and categorical attributes that hinder conventional investment segmentation and early-stage policy mapping. This study aims to develop a predictive clustering framework for enterprise investment profiling using mixed-type administrative data. The proposed methodology applies robust preprocessing, including RobustScaler for numerical variables and one-hot encoding with singular value decomposition for categorical features. Mixed-type similarity is computed using Gower distance, followed by a hybrid Gower–K-Means clustering approach, where the optimal number of clusters (k = 3) is determined using Silhouette, Calinski–Harabasz, and Davies–Bouldin indices. A comparative evaluation of clustering algorithms is conducted, with K-Prototypes performing best in the initial assessment and K-Means achieving superior performance after optimization. Cluster membership is subsequently predicted using a Random Forest classifier with hyperparameters optimized through randomized search. Experiments on 20,857 enterprise records identify three distinct clusters representing low-capital micro enterprises, transitional firms, and asset-intensive corporate entities. The optimized K-Means model achieves a Silhouette score of 0.97 and a Davies–Bouldin Index of 0.54. Compared with the untuned baseline, the tuned Random Forest model improves recall from 0.25 to 0.75 (200% increase) and increases the F1-score from 0.40 to 0.86 (114% improvement), while achieving 99.89% accuracy. These gains correspond to an estimated 20–30% improvement in MSME investment mapping effectiveness compared with traditional profiling approaches, providing a scalable AI-based blueprint for targeted regional economic governance.

Rule-Based Expert System for Personalized GERD Food Recommendations Using Forward Chaining and Certainty Factor in Indonesia

Alifani Maulia, Purwadi, Bagus Adhi Kusuma — Wed, 15 Apr 2026 00:00:00 +0000

This study develops a personalized and transparent rule-based expert system to support dietary decision-making for Indonesian patients with gastroesophageal reflux disease (GERD), addressing a critical gap in existing expert-system applications that focus mainly on diagnosis rather than daily diet management. The system integrates knowledge derived from the Indonesian GERD Consensus (2022) and its 2024 addendum with local nutritional evidence to construct if–then rules that classify foods into safe, limit, and avoid categories. A forward-chaining inference engine processes user-specific inputs—including symptoms, trigger sensitivities, eating behaviors, and dietary restrictions—while the Certainty Factor (CF) model quantifies confidence levels to accommodate individual tolerance variability. The system was implemented using Python and deployed through a Gradio-based wizard interface, enabling stepwise data collection and producing Top-N food recommendations with explainable “reason traces.” Functional evaluations across mild, moderate, and severe profiles demonstrated consistent alignment with national dietary guidelines, steering users toward low-fat, non-spicy, soft-textured, and clear-broth menu options, while eliminating high-risk trigger foods. Preliminary expert validation indicated high agreement with guideline principles, emphasizing the system’s interpretability and practical relevance. This research contributes to the field of health informatics by operationalizing forward chaining and CF for personalized dietary support, offering an auditable and computationally efficient alternative to black-box recommendation systems. Future developments include expanding the food dataset, refining CF calibration, and conducting structured clinical validation to enhance performance and applicability in real-world mHealth environments.

Decision Support Systems in Electronic Procurement for Public Sector Procurement: A Systematic Literature Review on Machine Learning Integration

Teguh Cahyono, Alva Hendi Muhammad, Sri Ngudi Wahyuni, Hanif Al Fatta — Sat, 18 Apr 2026 00:00:00 +0000

This study analyzes the evolution of Decision Support Systems (DSS) and Multi-Criteria Decision Making (MCDM) in public sector procurement between 2020 and 2025. Using bibliometric analysis of Scopus and Web of Science articles, the research focuses on themes such as e-procurement, supplier selection, public procurement, and the integration of intelligent technology. Network visualization, overlays, and density mapping were applied to explore keyword relationships, temporal trends, and research intensity. Findings reveal that in 2020, studies concentrated on transparency and digitalization in public e-procurement, with classical MCDA methods, fuzzy TOPSIS, and semantic DSS dominating the approaches. By 2022–2023, the emphasis shifted toward intelligent technologies, including artificial intelligence, neuro-fuzzy systems, and data mining algorithms. These innovations expanded DSS functions from evaluation to predictive analytics and optimization. Core themes such as supplier selection, optimization, and public procurement remained central, while emerging topics like sustainability and clinical decision support systems pointed to new research directions. A significant gap was identified in the university context. Although public sector e-procurement has been widely studied, no research has specifically addressed DSS–MCDM applications in higher education procurement systems. Consequently, future agendas should prioritize adaptive DSS tailored to universities, blockchain integration for transparency, and AI applications in clinical and humanitarian systems.

Hybrid LSTM-CNN-GRU Deep Learning for Integrating IoT and Social Media Sentiment Analysis in Indonesian Higher Education Reputation Management

Kresno Murti Prabowo, Ikbal Nidauddin, Endro Andiono — Wed, 15 Apr 2026 00:00:00 +0000

Higher education institutions in Indonesia face critical challenges in managing digital reputation. Despite 85% of prospective students using social media for university research, only 23% of institutions have integrated monitoring systems, resulting in 67% experiencing undetected reputation crises with substantial financial losses. This research proposes a novel framework integrating IoT campus data with social media sentiment analysis using hybrid deep learning architecture. The system employs LSTM-CNN networks with multi-head attention mechanisms for sentiment classification and GRU networks for reputation trend prediction, enhanced with data fusion strategy. Data collected from 428 IoT sensors and 3.2 million social media posts across five Indonesian universities over six months underwent advanced preprocessing including Indonesian-specific slang normalization and Sastrawi stemming. The hybrid LSTM-CNN architecture with attention achieved 90.3% sentiment classification accuracy (Macro-F1: 0.903), significantly outperforming baseline methods including Naive Bayes (76.2%), traditional LSTM (84.5%), and IndoBERT (87.1%). IoT integration contributed 18.2% RMSE improvement in trend prediction (R²: 0.874). The early warning system predicted reputation crises with 85.7% precision and 82.4% recall, providing critical intervention windows averaging 14.3 days before incidents. The real-time dashboard achieved 98.5% availability with sub-3-second response time and excellent usability (SUS score: 82.4). This research contributes: (1) novel IoT-sentiment integration framework with demonstrated effectiveness, (2) context-aware deep learning architecture optimized for Indonesian language achieving state-of-the-art performance, (3) validated early warning system enabling proactive reputation management, and (4) practical implementation with significant improvements over existing methods, advancing educational data analytics and AI-based decision support systems.

Security Assessment of JWKS-Based Authentication: Mitigating JWT Attack Vectors Through Penetration Testing

Ferry Andhika Pratama, Agus Hermanto, Geri Kusnanto — Sat, 18 Apr 2026 00:00:00 +0000

JSON Web Tokens (JWT) have become the de facto standard for stateless authentication in modern web applications and microservices architectures. However, improper implementation exposes systems to critical vulnerabilities including algorithm confusion attacks, signature bypass, and key injection exploits. This paper presents a comprehensive resilience analysis of JSON Web Key Set (JWKS)-based authentication mechanisms against known JWT attack vectors through a systematic penetration testing approach. We implemented and evaluated a production-grade courier management system (City Courier) featuring dynamic JWKS key rotation, RFC 7517-compliant public key distribution, and encrypted private key storage. Our penetration testing methodology systematically evaluated the system against 10 critical JWT attack vectors including algorithm confusion (CVE-2022-29217), kid parameter injection, weak secret exploitation, and signature verification bypass. Results demonstrate that proper JWKS implementation with dynamic key rotation, strict algorithm validation, and comprehensive audit logging provides robust defense against all tested attack vectors. The system successfully mitigated algorithm confusion attacks through explicit algorithm whitelisting, prevented kid injection via UUID-based key identifiers, and maintained security during key rotation events. Performance analysis shows minimal overhead (less than 50ms) for JWKS endpoint queries with aggressive caching. This research contributes practical implementation patterns for secure JWT authentication, providing both empirical evidence for JWKS-based security controls and a validated blueprint to neutralize critical vulnerabilities in modern microservices architectures.

Enhanced Lung Cancer Detection Using ANN with Random Oversampling, RFE-Based Feature Selection, and GridSearchCV Hyperparameter Tuning

Mon, 20 Apr 2026 00:00:00 +0000

Amid the most predominant mortality factors on a global scale, Lung cancer constitutes one of the most significant oncological burdens, chiefly because most patients receive a diagnosis only at later stages. The limitations of conventional diagnostic approaches underscore the urgent need for artificial intelligence–based detection systems that can improve both diagnostic accuracy and efficiency. This study aims to develop a lung cancer prediction model using an Artificial Neural Network (ANN) optimized through an integrated strategy that includes data preprocessing, class balancing via Random Oversampling (ROS), feature selection using Recursive Feature Elimination (RFE), and hyperparameter tuning with Grid Search. The evaluation of model effectiveness employs accuracy, precision, recall, F1-score, along with a confusion matrix. Experimental results demonstrate an accuracy of 98%, with average precision, recall, and F1-score values of 0.95. Statistical validation using McNemar’s test confirms a significant performance improvement over the baseline model (χ² = 18.05, p < 0.001), accompanied by a large effect size (Cohen’s h = 0.82). Furthermore, the model exhibits balanced performance in identifying both lung cancer and non-cancer cases, reflecting the effectiveness of the data balancing and feature selection strategies. These findings suggest that the optimized ANN model has strong potential as a foundation for a medical decision support system for early lung cancer detection, contributing to more reliable diagnoses and more accurate clinical decision-making.

Long Short Term Memory and Gradient Boosting Model for One Day Ahead Forecasting of ANTAM Gold Bar Prices

Annisa Ashari, Zakarias Situmorang, Rika Rosnelly — Sat, 18 Apr 2026 00:00:00 +0000

This study develops and optimizes a hybrid LSTM-XGBoost forecasting model for daily ANTAM gold bar prices. The model utilizes historical time-series data of ANTAM gold prices, enriched with macroeconomic variables including the USD/IDR exchange rate and Brent oil prices, as well as derived features such as returns, lags, rolling statistics, and calendar effects. The LSTM component captures medium-term sequential patterns from the price series and macroeconomic variables, while the XGBoost component exploits a rich set of tabular features to model nonlinear relationships and volatility dynamics. Both models are trained and tuned separately, then combined through a weighted ensemble scheme in which the optimal weight is selected by minimizing Mean Absolute Percentage Error (MAPE) on the validation set. Experimental results on the test set show that the proposed hybrid model achieves Mean Squared Error (MSE) of 26,891,172.36, Root Mean Squared Error (RMSE) of 16,398.53, MAPE of 0.0058 (approximately 99.42% accuracy), and coefficient of determination \mathbit{R}^\mathbf{2} of 0.9971, outperforming a naïve baseline that assumes “tomorrow’s price equals today’s price”. The optimized LSTM-XGBoost hybrid model proves highly effective for short-term ANTAM gold price forecasting, providing reliable decision support for Indonesian gold market stakeholders.

Evaluating SMOTE Performance for Imbalanced Multi-Label Sentiment Classification in MLSE Usability Testing of Mobile App Reviews

Hasan Basri, Wahyu Noviani Purwanti, Ihsan Alparisi — Wed, 15 Apr 2026 00:00:00 +0000

Imbalanced data poses a significant challenge in multi-label classification tasks, especially when combining sentiment analysis with usability testing of mobile application reviews. This study investigates the effectiveness of the Synthetic Minority Over-sampling Technique (SMOTE) in improving classification performance on a multi-label dataset consisting of 10,000 Indonesian language user reviews from the Google Play store. The classification labels represent a combination of usability criteria and sentiment polarity, with strong imbalance observed across several classes. Three machine learning algorithms SVM, Decision Tree, and Random Forest were evaluated on datasets of increasing sizes (1,000 to 10,000 entries), each tested under both original and SMOTE-balanced conditions using stratified 10-fold cross-validation with accuracy and F1-score as the primary metrics. Experimental results show that SMOTE significantly improves the performance of Decision Tree mainly on smaller datasets but exhibits inconsistent gains as the dataset grows, provides modest and stable improvements for Random Forest, and negatively impacts SVM, whose performance remains consistently better without SMOTE. This study concludes that SMOTE is not a universally effective solution and must be applied selectively based on model characteristics. These findings contribute to the Machine Learning for Software Engineering (ML4SE) domain and the field of informatics by highlighting the importance of aligning resampling techniques with algorithmic behaviour when dealing with highly imbalanced multi-label text classification tasks.

Cybersecurity Risk Detection Based on Roblox User Review Analysis Using TF-IDF and Comparison of Naïve Bayes and Support Vector Machine

RG Guntur Alam, Huda Ibrahim — Sat, 18 Apr 2026 00:00:00 +0000

The rapid growth of online gaming platforms increases user engagement while also exposing users to technical and cybersecurity risks. User reviews represent a rich yet underutilized textual source that can serve as early indicators of such risks. Unlike prior studies focused on sentiment polarity, this study positions user reviews as early cybersecurity risk signals by mapping complaint patterns into operational security risk categories relevant to system developers. This study compares Naïve Bayes (NB) and Support Vector Machine (SVM) in detecting cybersecurity risks from imbalanced textual data derived from Roblox user reviews. A total of 3,000 reviews were collected from the Google Play Store via web scraping and preprocessed using case folding, normalization, tokenization, stopword removal, and stemming. Reviews were classified into four cybersecurity risk categories (account access issues, suspicious behavior, connection instability, and data loss) based on rule-based security keyword mapping. Text representation employed TF-IDF with unigram and bigram features, while class imbalance was handled through undersampling. Model evaluation used three train–test splits (80:20, 70:30, and 60:40) and was assessed using Accuracy, Macro F1-score, AUC-PR, training time, and statistical testing. Results show that SVM consistently outperforms Naïve Bayes, achieving higher accuracy (0.86–0.88) and substantially better Macro F1-scores (0.73–0.77), indicating more balanced detection of minority cybersecurity risks. These differences are statistically significant (p < 0.05). The novelty of this study lies in transforming user reviews into a structured cybersecurity risk detection framework and empirically demonstrating the robustness of SVM in identifying rare but critical risks from imbalanced data.

Improving Imbalanced Data Classification Using Stacked Ensemble Learning with Naïve Bayes Variants and Random Forest

Helen Sastypratiwi, Yulianti, Hafiz Muhardi — Wed, 15 Apr 2026 00:00:00 +0000

Classification in imbalanced and heterogeneous datasets poses significant challenges in informatics, particularly in agricultural domains where minority classes are often underrepresented and feature redundancy affects model performance. This research aims to improve classification performance by developing a stacked ensemble learning framework that integrates probabilistic and tree-based learners to address class imbalance and enhance model interpretability. The framework combines Gaussian Naïve Bayes (GNB), Multinomial Naïve Bayes (MNB), and Random Forest (RF) as base learners with Logistic Regression as the meta-learner. Feature selection was performed using Chi-Square and ReliefF to identify the most relevant predictors, while SMOTE was applied to balance the dataset. Two ensemble configurations were evaluated: Ensemble A (GNB + MNB) and Ensemble B (GNB + RF). Experimental results demonstrate that Ensemble B achieved 97% accuracy and a macro F1-score of 0.97, with a 5.7% accuracy improvement over the best individual classifier and an 18% improvement in minority-class recall. The integration of probabilistic and tree-based models within a stacked architecture provides an interpretable and effective solution for data-driven decision systems in informatics, particularly valuable for domains requiring both high accuracy and model explainability in handling imbalanced datasets.

Integrated Maturity Assessment of Information Security for Land and Building Tax Management System Using National Institute of Standards and Technology Cybersecurity Framework 2.0, International Organization for Standardization/International Electrotechnical Commission 27002:2022, and Cybersecurity Capability Maturity Model 2.1.

Dhenok Prastyaningtyas Paramesvari, Jatmiko Endro Suseno, Catur Edi Widodo — Thu, 16 Apr 2026 00:00:00 +0000

Regional tax information systems such as the Sistem Informasi Manajemen Objek Pajak (SISMIOP) are vulnerable to cybersecurity threats due to the sensitivity of taxpayer data and the persistence of ad-hoc security management practices. These conditions pose risks to data confidentiality, integrity, and service availability, potentially undermining public trust and the effectiveness of local government services. This study aims to assess the information security maturity of SISMIOP operated by the Badan Pengelolaan Pendapatan, Keuangan, dan Aset Daerah (BPPKAD) through an integrated application of the NIST Cybersecurity Framework (CSF) 2.0, ISO/IEC 27002:2022, and the Cybersecurity Capability Maturity Model (C2M2) 2.1. A qualitative case study approach was employed. An organizational profile was developed using interviews, observations, and document analysis, followed by mapping 38 relevant NIST CSF subcategories to ISO/IEC 27002 controls and C2M2 capability domains. Security maturity was evaluated using questionnaires and interviews based on the C2M2 Maturity Indicator Levels (MIL0-MIL3), and a gap analysis was conducted against the target maturity level of MIL2. The results show that most cybersecurity functions, Govern, Identify, Detect, Respond, and Recover, remain at MIL1, indicating that practices are performed but not yet formalized or consistently implemented. The Protect function partially achieved MIL2. The largest gaps were identified in governance and risk management domains. Based on these findings, 38 prioritized strategic recommendations were formulated to improve policy formalization, risk management, technical controls, monitoring, and incident handling. This study contributes a practical and replicable multi-framework maturity assessment model to strengthen information security governance in public-sector tax information systems.

An Evaluation of Self-Attentive Sequential Recommendation (SASRec) Algorithm Using Hyperparameter Tuning

Agung Toto Wibowo, Hasmawati, Hani Nurrahmi, Imtitsal Ulya Salsabila — Sat, 18 Apr 2026 00:00:00 +0000

Sequential recommendation is a branch of Recommender Systems that aims to predict the next item a user will interact with based on their historical sequence of interactions. The main challenge in SR is to capture both short-term and long-term dependencies among items within a sequence. Self-Attentive Sequential Recommendation (SASRec) is a self-attention-based deep learning model designed to recognize sequential interaction patterns. Despite its effectiveness, the performance of SASRec is highly dependent on hyperparameter configurations, yet comprehensive evaluations remain limited. This research aims to evaluate the influence of SASRec's configuration through hyperparameter tuning on sequential recommendation performance. The hyperparameters used are hidden_size, inner_size, number of attention heads (num_heads), and number of layers (num_layers). The evaluation was conducted on two public datasets with different sparsity characteristics: MovieLens-1M (Sparsity ≈ 95.80%) and Amazon Musical Instruments (Sparsity ≈ 99.99%). In this study, Recall@k and MRR@k were used as performance metrics. The test results showed that hidden_size and inner_size had a significant positive impact on performance, especially on the dense dataset. The optimal hidden_size was obtained at hidden_size = 64 on the Amazon Musical Instrument dataset, and at hidden_size = 256 on the Movielens 1M dataset. The optimal inner_size was obtained at inner_size = 256 on both datasets. Meanwhile, the num_heads and num_layers hyperparameters did not provide a significant performance improvement. Furthermore, in the comparison between SASRec, GRU4Rec, and BERT4Rec, SASRec outperforms GRU4Rec and BERT4Rec in handling highly sparse datasets such as Amazon Musical Instruments obtained average recall@20 = 0.0678, and average MRR@20 = 0.0223.

Augmentation Strategy and Hyperparameter Optimization Using Optuna for Potato Leaf Disease Classification in Uncontrolled Environment

Harri Kurniawan Rofiqi, Edi Noersasongko, Sri Winarno, M. Arief Soeleman — Wed, 15 Apr 2026 00:00:00 +0000

Image-based classification of potato leaf diseases presents a significant challenge, particularly when data are collected in uncontrolled field environments. While Convolutional Neural Networks (CNNs) and Computer Vision have been widely used for plant disease identification, most previous studies relied on laboratory datasets with uniform lighting and backgrounds, limiting their real-world applicability. This study proposes an integrated framework that combines data augmentation, class balancing using the Synthetic Minority Over-sampling Technique (SMOTE), and automated hyperparameter optimization through Optuna to enhance the robustness and accuracy of CNN-based models. A total of 3,076 high-resolution potato leaf images representing seven disease classes were evaluated across five CNN architectures and three training scenarios. The MobileNetV3-Large model achieved the best baseline performance with an accuracy of 0.863 and F1-score of 0.868, while Optuna-based optimization further improved performance to 0.895 accuracy, 0.913 precision, 0.906 recall, and 0.904 F1-score, demonstrating the effectiveness of adaptive optimization in improving model generalization. The integration of augmentation, SMOTE, and Optuna resulted in an intelligent and efficient system resilient to environmental variability, showing strong potential for automatic early detection of potato leaf diseases in real agricultural settings. This research contributes to the advancement of Informatics and Artificial Intelligence by promoting adaptive computer vision approaches for smart agriculture and real-world image-based diagnostic systems.

Geographic Information System for Land Suitability Mapping of Partner Farmers at Okiagaru Indonesia Agricoop Using Rule-Based System and Prototype Methodology

Wed, 15 Apr 2026 00:00:00 +0000

Geographic Information System (GIS) for land suitability assessment integrates spatial and attribute data to evaluate and map areas for food crops and horticulture. The system applies parameters such as temperature, rainfall, water pH, clay CEC, organic carbon, and other soil characteristics, analyzed through a rule-based approach. Its main goal is to optimize agricultural land use by aligning crop selection with physical and environmental conditions. GIS-based analysis enables accurate digital mapping and categorizes land into highly suitable, moderately suitable, marginally suitable, and unsuitable classes, providing valuable insights for farmers, governments, and stakeholders in sustainable land management. System development employed the Prototype methodology, emphasizing iterative stages of requirement gathering, rapid design, prototype construction, user evaluation, and refinement. The Land Suitability GIS (SigKL) was tested at five partner-farmer sites in Cianjur and Sukabumi. Black Box Testing confirmed that all 20 functional features achieved a 100% success rate. The system supports the identification of potential new agricultural areas and offers recommendations for improving less productive land. The novelty of this research lies in integrating FAO (1976) classification with interactive digital mapping and locally tailored knowledge rules, enabling real-time accessibility. Unlike prior studies limited to static analysis, SigKL introduces an adaptive, rule-based GIS prototype with interactive visualization, directly supporting decision-making for sustainable agriculture. This innovation enhances transparency and accessibility, contributing to the Sustainable Development Goals (SDGs) related to food security and sustainable farming.

Enhancement Of The C4.5 Decision Tree Algorithm With Anova For Predicting Academic Achievement Of Students At Smpn.16 Kota Jambi

Rice Osviarni, Setiawan Assegaff, Jasmir, Nurhadi — Wed, 15 Apr 2026 00:00:00 +0000

This study aims to improve the accuracy of predicting student academic achievement by integrating the Analysis of Variance (ANOVA) method with the C4.5 Decision Tree algorithm. In the context of information systems, this research holds significant importance for the development of more reliable Decision Support Systems (DSS) or early warning systems in school environments. The research was conducted at SMPN 16 Jambi City using secondary data from three academic years (2022/2023-2024/2025) covering academic variables, attendance, and parental income. The main issue addressed was the limitations of the C4.5 algorithm in handling irrelevant features and unbalanced data, which, at the system implementation level, can lead to inaccurate recommendations or alerts.This research method employed a data mining approach with stages including data cleaning, numeric conversion, missing value imputation, formation of derived variables, and categorization of the target variable "Achievement." The initial C4.5 model produced 72.81% accuracy on the training data and 69.71% accuracy on cross-validation. After feature selection using ANOVA, one insignificant variable was removed, resulting in a hybrid C4.5+ANOVA model with nine key features. Test results showed an increase in accuracy to 80.44% on the training data and 73.66% on the cross-validation data, representing an improvement of 7.63 and 3.95 percentage points, respectively.This improvement in model performance directly translates to an enhancement in the quality of the information system's output, yielding more reliable reports and predictions for teachers and school management.

Machine Learning Decision Support System for Heart Disease Prediction with Optuna and Threshold Optimization

William Ramdhan, Jeperson Hutahaean, Deny Jollyta, Abdul Karim — Sat, 18 Apr 2026 00:00:00 +0000

Cardiovascular disease remains a major global health challenge, necessitating accurate and reliable decision support systems for early detection. This study proposes a machine learning–based decision support system that integrates ensemble learning, automated hyperparameter optimization using Optuna, and decision threshold tuning. The system was evaluated using several baseline machine learning models, including Logistic Regression, SVM, KNN, Decision Tree, and Random Forest, with the Random Forest model selected for optimization. Hyperparameter tuning with Optuna and decision threshold optimization led to a significant improvement in accuracy (95.0%) and ROC–AUC (0.977), with the optimized model outperforming all baseline models. This approach demonstrates improved sensitivity, reduced false negatives, and enhanced predictive performance, offering a clinically reliable tool for early heart disease detection. The results emphasize the importance of model optimization and decision threshold calibration in clinical decision support systems.

Analysis of Public Sentiment Indonesia’s Personal Data Protection Law: A Comparison of SVM and IndoBERT on X Platform

Wed, 15 Apr 2026 00:00:00 +0000

The high number of data misuses, thefts, and leaks led to the enactment of the PDP Law, which regulates the rights and obligations of data owners and electronic system providers. The purpose of this study is to examine the public’s response to the implementation of the law through the X platform, using tweet harvest as a scraping tool, and to evaluate model performance through a comparative approach between SVM and BERT. The feature extraction used in this study is TF-IDF for SVM and BERT with IndoBERT. The accuracy results indicate that BERT is better with an accuracy of 86% compared to SVM with a training and test data ratio of 85:15. This advantage is because BERT can understand linguistic context that SVM cannot. On the other hand, SVM has advantages in computational efficiency and faster processing, making it a suitable choice in situations with limited computational resources.

The sentiment analysis result revealed that data protection, digital footprint and the institution's role were the most frequently discussed topics. Furthermore, periodic or real-time evaluations can be conducted on the public's response to the PDP Law to ensure it remains aligned and relevant to technological developments and societal needs.

Comparative Analysis of Explainable AI Methods LIME, SHAP, and ELI5 on Random Forest Based Indonesian E-Commerce Sentiment Classification

Haditya Pandu Winanta, Muhammad Yusril Hana, Firman Noor Hasan — Sun, 19 Apr 2026 00:00:00 +0000

The rapid growth of e-commerce platforms in Indonesia has generated a massive volume of product reviews, making sentiment classification essential for understanding customer perceptions and supporting data-driven decision making. This study aims to develop a sentiment classification model for Indonesia e-commerce product reviews while enhancing model transparency through Explainable Artificial Intelligence (XAI). The proposed approach employs a Random Forest classifier eith Term Frequency-Inverse Document Frequency (TF-IDF) for feature extraction. The dataset consists of 23,194 product reviews from the fashion and electronics categories, classified into positive, negative, and neutral sentiment. Model performance is evaluated using accuracy, precision, recall, and F1-Score metrics. Experimental results show taht the Random Forest model achieves an accuracy of 93.74%, with the best performance observed in the postive sentiment class. To improve interpretability, three XAI methods-LIME, SHAP, and ELI5-are applied. The analysis indicates that LIME is effective for local explanations, SHAP provides consistent global and local feature importence, and ELI5 offers concise and computationally efficient global explanations. This study contributes to the field of computer science by demostrating how comparative XAI analysis can bridge the gap between high-performing black-box models and interpretable sentiment classification in high-dimensional extual data, thereby supporting transparent and accountavle AI system in e-commerce applications.

Deep Learning Based MobileNet Optimization For High Accuracy Classification Of Toddler Stunting

Anan Wibowo, Rahmat Widia Sembiring, Solikhun — Wed, 15 Apr 2026 00:00:00 +0000

This study aims to develop and optimize a MobileNet-based deep learning model for toddler stunting classification using whole-body images. A progressive optimization strategy was applied through three scenarios: (1) a baseline MobileNet feature-extraction model, (2) an optimized fine-tuned model, and (3) a final model enhanced with an adaptive ReduceLROnPlateau scheduler. Using a private dataset of 571 images, the proposed model achieved significant improvements—from 97.47% accuracy in the baseline model to a perfect 100% accuracy, precision, recall, and F1-score in the final scenario. These results highlight the novelty of this study, namely the use of whole-body images combined with progressive MobileNet optimization, which substantially outperforms prior studies relying solely on facial image analysis. The proposed approach demonstrates strong potential as a highly accurate and efficient computational tool for clinical stunting screening.