Jurnal Teknik Informatika (Jutif)

Image-Based Classification of Rice Field Conversion: A Comparison Between MLP and SVM Using Multispectral Features

Anisya, Sumijan, Anna Syahrani — Thu, 16 Oct 2025 00:00:00 +0000

The conversion of farmland into non-agricultural purposes has emerged as a pressing concern in many urban regions, including Koto Tangah District, Padang City. In this area, agricultural land experienced a 4% shift in land use between 2022 and 2024. If this trend continues, it could lead to a notable decline in rice production and ultimately threaten food security. This research focuses on examining spatial transformations of rice fields from 2022 to 2024 by utilizing Sentinel-2 satellite imagery along with advanced classification techniques. Vegetation and moisture features were extracted using NDVI, NDWI, texture analysis through GLCM, and Principal Component Analysis (PCA). To classify land cover changes and assess model accuracy, two machine learning approaches were applied: Multilayer Perceptron (MLP) and Support Vector Machine (SVM). The findings reveal a considerable reduction in dense vegetation, indicated by the downward shift of NDVI values in 2024. MLP achieved an accuracy of 82%, outperforming SVM, which reached 71%. Furthermore, MLP obtained a higher F1-score for non-rice field detection (0.75 vs. 0.74) and produced more realistic delineations of rice field boundaries during spatial validation. These outcomes highlight the potential of MLP in monitoring land use conversion, supporting agricultural land conservation, and guiding sustainable urban planning. Moreover, the study contributes to computer science by advancing the use of machine learning for spatio-temporal analysis and reinforcing the role of non-linear models in satellite image classification.

Digital Landscape and Behavior in Indonesia 2024: A National Survey Analysis of Internet Penetration, Cybersecurity Risks, and User Segmentation Using K-Means Clustering and Logistic Regression

Nur Aminudin, Nurul Hidayat, Dwi Feriyanto, Dita Septasari, Ikna Awaliyani — Thu, 16 Oct 2025 00:00:00 +0000

Digital transformation in Indonesia reveals significant disparities in internet access, digital behavior, and cybersecurity vulnerabilities. This study analyzes the digital landscape using national survey data from 8,720 respondents across 38 provinces. This research employs a quantitative approach, utilizing chi-square tests, logistic regression for risk analysis, and K-Means clustering for user segmentation, supported by Principal Component Analysis (PCA) for dimensionality reduction. The results show a national internet penetration rate of 79.5%, with significant disparities across regions and socio-economic segments. Logistic regression analysis reveals that higher education, greater income, and the use of fixed broadband are negatively correlated with cybersecurity risks. Furthermore, K-Means clustering identifies three distinct user profiles: 'Digital Savvy', 'Pragmatic Users', and the 'Vulnerable Segment', each with unique characteristics regarding digital access and literacy. This research provides a critical empirical basis for understanding digital transformation in a developing nation. The findings underscore the necessity of data-driven, segmented policies to foster digital inclusion and enhance national cybersecurity, offering actionable insights for policymakers and service providers.

Sentiment Analysis and Topic Modeling for Discovering Knowledge in Indonesian Mobile Government Applications

Thu, 16 Oct 2025 00:00:00 +0000

The accelerated rate of government applications development in Indonesia has introduced new opportunities and challenges in delivering digital public services. While thousands of apps have been developed, systemic issues ranging from usability flaws to authentication failures persist, as reflected in user reviews on platforms like the Google Play Store. This study adopts a knowledge discovery approach to extract actionable insights from more than 17,000 user-generated reviews across three major government applications: Satusehat, Digital Korlantas, and M-Paspor. A hybrid methodology is applied, combining RoBERTa-based sentiment classification, BERTopic-based topic modeling, cosine similarity analysis, and qualitative user validation. The findings reveal recurring issues in authentication, interface design, and system responsiveness that span across organizational boundaries. Cross-app topic correlation highlights critical shared pain points such as login failures and unintuitive UI that undermine user trust in e-government services. Mapping these insights onto the SECI knowledge management model, this research contributes both practical recommendations and a replicable analytical framework for public agencies seeking to institutionalize user feedback. By transforming fragmented digital feedback into organizational knowledge, this study supports continuous service improvement and strengthens the foundation for user-centric e-government.

An Integrated Pipeline with Hierarchical Segmentation and CNN for Automated KTP-el Data Extraction on the e-Magang Platform

Nuansa Syafrie Rahardian, Eddy Maryanto, Devi Astri Nawangnugraeni — Thu, 16 Oct 2025 00:00:00 +0000

In alignment with Indonesia's digital transformation agenda, this research addresses the inefficiencies and error-prone nature of manual data entry on the Foreign Policy Strategy Agency's (BSKLN) e-magang platform. This study introduces a comprehensive, end-to-end Optical Character Recognition (OCR) pipeline, specifically designed for structured identity documents and real-world government platform integration. The proposed methodology features a robust workflow, including image preprocessing with histogram matching, hierarchical segmentation using vertical projection, and intelligent postprocessing to structure the output. To overcome the limitations of a small dataset, three specialized Convolutional Neural Network (CNN) models were rigorously trained and validated using a stratified 5-fold cross-validation technique. The final system was successfully integrated, connecting a Flask-based model engine with the existing Laravel and React platform. End-to-end testing demonstrated strong performance, achieving an average character-reading accuracy of 93.31% with a mean processing time of 14.48 seconds per image. The primary contribution of this research to the field of informatics is the development of a complete and deployable system architecture that ensures data interoperability and reliability, providing a practical blueprint for integrating intelligent automation into digital public services.

Brain Tumor Segmentation From MRI Images Using MLU-Net with Residual Connections

Eric Timothy Rompisa, Gede Putra Kusuma — Thu, 16 Oct 2025 00:00:00 +0000

Brain tumor segmentation plays an important role in medical imaging in assisting diagnosis and treatment planning. Although advances in deep learning such as Unet already perform image segmentation, many challenges exist in segmenting brain tumors with tumor spread boundaries. This paper proposes a model that combines CNN and MLP (MLU-Net) techniques enhanced by the addition of residual connections to improve segmentation accuracy called ResMLU-Net. This architecture combines 2D covolution layers, block MLP and residual connections to process MRI images with the dataset used is BraTS 2021. The residaul connection helps reduce gradient degradation which ensures smooth information flow and better feature learning. The performance of ResMLU-Net will be evaluated using Dice and IoU metrics and will also be compared with several models such as Unet, ResUnet and MLU-Net. The experimental scores obtained from ResMLU-Net for segmenting brain tumors are 83.43% for IoU and 89.94% for Dice. These results show that adding residual connections can improve the accuracy in segmenting brain tumors which can be seen that there is an increase in the Dice and Iou scores. The proposed ResMLU-Net model is a valuable contribution to medical imaging and health informatics. With its provision of a standard and computationally viable solution to brain tumor segmentation, it offers incorporation into Computer-Aided Diagnosis (CAD) systems and support to clinical decision-making protocols.

Enhanced U-Net Cnn For Multi-Class Segmentation And Classification Of Rice Leaf Diseases In Indonesian Rice Fields

Faturrohman, Odi Nurdiawan, Willy Prihartono, Rully Herdiana — Thu, 16 Oct 2025 00:00:00 +0000

Rice is a strategic food crop whose productivity is often threatened by leaf diseases and pests. This study aims to develop an Enhanced U-Net CNN model for multi-class segmentation and classification of rice leaf conditions from field images to support early detection and plant health management. The methodology includes direct field image acquisition of rice leaves, preprocessing for image quality enhancement, expert data labeling, segmentation using a U-Net architecture, and classification using CNN. The dataset was divided into training and testing data with balanced distribution across four classes: Healthy, BrownSpot, Hispa, and LeafBlast. Evaluation results show that the model can identify rice leaf conditions with high accuracy, although signs of overfitting were observed from the performance gap between training and validation data. The implementation of this model is expected to accelerate automatic disease detection in the field, reduce reliance on manual inspection, and support precision agriculture. This study achieved a testing accuracy of 76.36% with a macro-average F1-score of 0.34. While the results indicate limitations in generalization, the proposed Enhanced U-Net CNN demonstrates the feasibility of combining segmentation and classification in field conditions. This research contributes to agricultural informatics by supporting scalable deployment in precision agriculture systems, reducing reliance on manual inspection, and providing a foundation for further optimization studies.

Comparison of Information Technology Governance Maturity Levels Based on COBIT 2019 at PT Kereta Commuter Indonesia in 2023 and 2024

Purwadi, Handri Santoso — Thu, 16 Oct 2025 00:00:00 +0000

This study aims to analyze and compare the maturity level of Information Technology (IT) governance at PT Kereta Commuter Indonesia (KCI) between 2023 and 2024 using the COBIT 2019 framework. The background of this study is based on the operational complexity of KCI which serves a high daily passenger volume, so that the information system becomes the backbone of the smooth transportation service. The method used is a descriptive-comparative case study with a mixed approach, through interviews, Likert scale questionnaires, and internal document reviews such as IT audit reports and government regulations. The results of the analysis show a significant and consistent increase, where the level of IT governance maturity which was previously at level 2 (Managed) and 3 (Defined) in 2023, increased to level 3 and 4 (Quantitatively Managed) in 2024. The most prominent improvements were seen in the strategic domain EDM01 (Ensure Governance Framework Setting) and the operational domain DSS01 (Manage Operations), which successfully reached level 4. This success reflects top management's commitment and ongoing internal evaluation in strengthening IT governance strategically and operationally. The research findings confirm that annual evaluations serve as an objective benchmark for identifying governance gaps, developing digital strategies, and determining future IT investment priorities. Overall, this study confirms that regular assessments can improve the effectiveness of data-driven IT transformation and ensure alignment of IT implementation with the company's business objectives.

Air Quality Index Classification: Feature Selection for Improved Accuracy with Multinomial Logistic Regression

Rizky Caesar Irjayana, Abdul Fadlil, Rusydi Umar — Thu, 16 Oct 2025 00:00:00 +0000

Air pollution is a major public health concern, creating the need for accurate and interpretable Air Quality Index (AQI) classification models. This study aims to classify AQI into three categories—Good, Moderate, and Unhealthy—using Multinomial Logistic Regression (MLR) with feature selection. The dataset, obtained from public monitoring stations in Jakarta between 2021 and 2024, initially contained 4,620 daily records. After cleaning and outlier removal, 3,586 valid samples remained, from which 900 balanced records (300 per class) were selected for modeling. Key features included PM₁₀, PM₂.₅, SO₂, CO, O₃, and NO₂, which were standardized using Max Normalization to ensure uniform feature scaling. The classification process applied k-fold cross-validation (k = 2–5), and performance was assessed using accuracy and Macro F1-score. Results show that including PM₂.₅ improves performance by about 10%, with the best outcome at k = 5 (accuracy = 91.67%, Macro F1 = 91.45%). These findings confirm PM₂.₅ as a decisive feature for AQI prediction and demonstrate that MLR provides a lightweight, transparent, and computationally efficient solution. Beyond environmental health, the contribution of this work lies in advancing data-driven decision support systems in Informatics, particularly for real-time monitoring and policy applications.

Bayesian Optimized Pretrained CNNs for Mango Leaf Disease Classification: A Comparative Study

Sri Rahayu, Sayyid Faruk Romdoni — Thu, 16 Oct 2025 00:00:00 +0000

Mango leaf diseases pose a major threat to crop productivity, causing significant economic losses for farmers. Accurate and early detection is essential, yet manual diagnosis remains subjective and inefficient. This study aims to evaluate and compare the performance of five pretrained Convolutional Neural Network (CNN) architectures—DenseNet121, ResNet50V2, MobileNetV3 Small, MobileNetV3 Large, and InceptionV3—by systematically optimizing their hyperparameters to identify the most effective model for mango leaf disease classification. The public MangoLeafBD dataset, containing 4,000 images from eight balanced classes, was used. Bayesian Optimization was applied to fine-tune each model, and their performances were assessed before and after optimization. Results show that optimization substantially improved all models, with MobileNetV3 Large achieving the highest accuracy of 100% on the test set, followed by DenseNet121 (99.75%), ResNet50V2 (99.63%), MobileNetV3 Small (99.50%), and InceptionV3 (98.50%). The findings highlight that a well-tuned lightweight model can outperform more complex architectures, offering a practical and efficient solution for developing mobile-based diagnostic tools to support precision agriculture in resource-constrained settings.

Stacked Random Forest-LightGBM for Web Attack Classification

Fadli Dony Pradana, Farikhin, Budi Warsito — Thu, 16 Oct 2025 00:00:00 +0000

The rapid expansion of web services in the digital era has intensified exposure to increasingly complex and imbalanced cyber threats. This study proposes a stacking hybrid ensemble framework for web attack classification, integrating Random Forest as the base learner and LightGBM as the meta-learner, enhanced by the SMOTE technique for data balancing. The Web Attack subset of the CICIDS-2017 dataset serves as a case study, with a focus on detecting minority attacks such as SQL Injection, XSS, and Brute Force. The preprocessing pipeline includes data cleaning, removal of irrelevant features, normalization, extreme value imputation, and ANOVA F-test-based feature selection. Evaluation results indicate that the proposed model outperforms baseline models in both multiclass classification (98.7% accuracy, 0.634 macro F1-score) and binary classification (99.41% accuracy, 99.47% F1-score), while maintaining high sensitivity to minority classes. These results contribute to informatics and cybersecurity scholarship through a generalizable stacking baseline and well-specified evaluation procedures for web-attack detection, facilitating replicability, fair comparison, and dataset-agnostic insights.

RNN-Based Intrusion Detection System for Internet of Vehicles with IG, PCA, and RF Feature Selection

Thu, 16 Oct 2025 00:00:00 +0000

Cyberattacks in the Internet of Vehicles (IoV) threaten road safety and data integrity, requiring intrusion detection systems (IDS) that capture temporal patterns in vehicular traffic. This study develops a Recurrent Neural Network (RNN)-based IDS and evaluates three feature-selection strategies—Information Gain (IG), Principal Component Analysis (PCA), and Random Forest (RF)—on the CICIoV2024 dataset. Features are normalized using Min–Max scaling before being fed into the RNN classifier. The models achieve perfect classification on held-out tests (accuracy/precision/recall/F1 = 1.00). However, probabilistic evaluation reveals low ROC–AUC scores (IG: 0.572, PCA: 0.429, RF: 0.415), indicating limited discriminative margins and potential overfitting or calibration issues despite flawless confusion matrices. PCA and RF further reduce computational overhead during inference compared to IG. These findings highlight that relying solely on accuracy can be misleading for IDS evaluation; temporal RNNs should be complemented with probability-aware training, calibration, or hybrid architectures. This work contributes a temporal-aware IDS framework for IoV and motivates future research on real-time deployment, hybrid RNN-CNN/LSTM models, and adversarial robustness to improve generalization and safety of connected vehicles

Comparative Evaluation of Decision Tree and Random Forest for Lung Cancer Prediction Based on Computational Efficiency and Predictive Accuracy

Muhammad Yashlan Iskandar, Handoyo Widi Nugroho — Thu, 16 Oct 2025 00:00:00 +0000

Early detection of lung cancer is essential for improving treatment outcomes and patient survival rates. This paper presents a comparative evaluation of two classification algorithms: Decision Tree and Random Forest, focusing on both predictive performance and computational efficiency. The models were tested using 10-fold cross-validation to ensure robustness. Both algorithms achieved the same accuracy of 93.3%. However, Random Forest slightly outperformed Decision Tree in recall (88.8% vs. 87.9%), F1-score (92.2% vs. 92.1%), and AUC (0.94 vs. 0.91), while Decision Tree obtained higher precision (97% vs. 95.9%). In terms of computational efficiency, Decision Tree demonstrated faster training and testing times, lower memory usage, and reduced energy consumption compared to Random Forest. The results reveal a clear trade-off between prediction quality and resource usage, highlighting the importance of selecting algorithms not only for their accuracy but also for their practicality in real-world healthcare scenarios. This comprehensive evaluation provides valuable insights for developing intelligent decision support systems that are both effective and resource-efficient, especially in environments with limited computing capacity. These findings contribute to the advancement of resource-aware intelligent systems in the field of medical informatics.

Comparison of Transfer Learning Strategies Using MobileNetV2 and ResNet50 for Ecoprint Leaf Classification

Siti Hajar, Murinto, Anton Yudhana — Thu, 16 Oct 2025 00:00:00 +0000

This research focuses on the classification of leaf types used in ecoprint production through the steaming technique by applying transfer learning on two widely recognized convolutional neural network (CNN) architectures, MobileNetV2 and ResNet50. Leaves have diverse applications in various sectors such as medicine, nutrition, and handicrafts. The study utilized a total of 600 leaf images from 15 species were collected from the surrounding environment and divided into 80% training and 20% testing sets. The aim of this study is to classify leaf types suitable for ecoprint quickly and efficiently, based on transfer learning with two CNN architectures, while incorporating fine-tuning. MobileNetV2 was selected for its computational efficiency, while ResNet50 was chosen for its ability to address the vanishing gradient problem and deliver high accuracy. Fine-tuning was employed to optimize model performance. Experimental results demonstrate that both architectures achieved strong performance, with MobileNetV2 reaching 94.12% accuracy and ResNet50 slightly outperforming it at 94.96%. Confusion matrix evaluation further confirmed these results, yielding accuracy, precision, recall, and F1-score values of 0.94, 0.95, 0.95, and 0.94, respectively. These findings highlight ResNet50’s superior performance over MobileNetV2 while affirming the effectiveness of both models in ecoprint leaf classification.

Single-Image Face Recognition For Student Identification Using Facenet512 And Yolov8 In Academic Environtment With Limited Dataset

Thu, 16 Oct 2025 00:00:00 +0000

Face recognition has become one of the most significant research areas in image processing and computer vision, mainly due to its wide applications in security, identity verification, and human and machine interaction. In this study, FaceNet512 and YOLOv8 models are used to overcome the challenges in face recognition with a limited dataset, which is only one formal photo per individual. The application of image augmentation to the model achieved 90% accuracy and ROC curve of 0.82, while the model without augmentation achieved 89% accuracy and ROC curve of 0.79. FaceNet512 showed superiority in producing more accurate and detailed facial representations compared to other models, such as ArcFace and FaceNet, especially in handling minimal facial variations. Meanwhile, YOLOv8 provides efficient face detection across various lighting conditions and viewing angles. The main challenge in this research is the low quality of the original image, which can reduce the accuracy of face recognition. These results show the great potential of using deep learning-based face recognition systems in the real world, especially for automatic attendance applications in academic environments.

Automated Property Valuation with Multi-Hazard Risk: Jakarta Metropolitan Area Study

Fachrurrozi, Hanna Arini Parhusip, Suryasatriya Trihandaru — Thu, 16 Oct 2025 00:00:00 +0000

This study crafts a machine learning framework that systematically integrates multi-hazard disaster risk assessments into automated property valuation for the Jakarta Metropolitan Area. The framework addresses 25–30% MAPE typically observed in disaster-prone regions, providing more reliable valuation results. We made 114 prediction features from 42 input variables by using 14,284 property data from Indonesian markets, physical risk data from the Think Hazard platform, and socio-economic data from Central Bureau of Statistics. Elastic Net model performed superior compared to other models which had R² = 0.7922 and a MAPE of 28.27%. We found that some disaster risks had unexpected beneficial effects on property prices. We expected that risks related to the earth (+40.5%) and water (+19.2%) would have positive effects, while risks related to the weather (-66.9%) would have negative effects. These conflicting results suggest that in complex urban markets, the quality of infrastructure, location premiums, and differences in risk perception may outweigh simple risk penalties. The idea gives realistic ideas for property valuation that takes risks into account, but it also points out big problems with how the market judges how likely a disaster is to happen.

Forecasting Bitcoin Price Prediction with Long Short-Term Memory Networks: Implementation and Applications Using Streamlit

Thu, 16 Oct 2025 00:00:00 +0000

The rapid growth of cryptocurrency markets, particularly Bitcoin, has highlighted the need for accurate price prediction models to support informed decision-making. While existing studies primarily evaluate machine learning models for price forecasting, few have implemented these models in real-world applications. This paper addresses this gap by developing a Bitcoin price prediction system using Long Short-Term Memory (LSTM) networks, integrated into a user-friendly web-based application powered by Streamlit. The model forecasts Bitcoin prices at 5-minute, 1-hour, and 1-day intervals, demonstrating strong predictive performance. For the 5-minute interval, the model achieved a Mean Squared Error (MSE) of 53,479.86, Mean Absolute Error (MAE) of 150.58, Root Mean Squared Error (RMSE) of 231.26, and Mean Absolute Percentage Error (MAPE) of 0.144%. At the 1-hour interval, errors increased moderately with an MSE of 423,198.24, MAE of 499.93, RMSE of 650.54, and MAPE of 0.505%. For the 1-day interval, the model faced greater variability, reflected in an MSE of 3,089,699.07, MAE of 1,058.88, RMSE of 1,757.75, and MAPE of 2.027%. These results indicate that while predictive precision decreases over longer horizons, the model maintains strong performance across all timeframes. By embedding LSTM predictions into an interactive, real-time forecasting platform, this study demonstrates the practical integration of deep learning into complex financial systems. Beyond cryptocurrency, the approach highlights the potential of intelligent computational models to enhance decision-making processes in data-intensive domains, reinforcing the role of informatics in bridging advanced algorithms with usable technological solutions.

Brain Tumor Auto Segmentation On 3D MRI Using Deep Neural Network

Melda Agarina, Muh Royan Fauzi Maulana, Sutedi, Arman Suryadi Karim — Thu, 16 Oct 2025 00:00:00 +0000

Accurate and automated segmentation of brain tumours from Magnetic Resonance Imaging (MRI) is crucial for clinical diagnosis and treatment planning, yet it remains a significant challenge due to tumour heterogeneity and data imbalance. This research investigation examines the effectiveness of a 3D UNet architecture for the segmentation of brain tumours utilizing MRI imaging modalities. The research employs the BRATS 2021 dataset, which consists of 675 MRI datasets across four distinct imaging modalities (FLAIR, T1-Weighted, T1-Contrast, and T2-Weighted) and encompasses four distinct segmentation label classes. The employed model integrated soft dice loss and dice coefficient as its loss functions, with the objective of achieving convergence despite the presence of imbalanced data. While constraints related to resources limited the training process, the model yielded promising outcomes, exhibiting high accuracy (99.43%) and specificity (99.5%), The model aids medical professionals in understanding tumor growth and enhances treatment planning via segmentation predictions in surgery. Nevertheless, the sensitivity, particularly concerning non-enhancing tumour classes, persists as a significant challenge, underscoring the necessity for future research to concentrate on data-centric methodologies and enhanced pre-processing techniques to improve model efficacy in critical medical applications such as the segmentation of brain tumours.

Development of a Church Information Management System Using Scrum at HKBP Sola Gratia Kayu Mas Jakarta

Master Edison Siregar, Hendra Mayatopani, Rido Dwi Kurniawan, Deasy Olivia — Thu, 16 Oct 2025 00:00:00 +0000

The rapid growth of the congregation at HKBP Sola gratia Kayu Mas Church in Jakarta has posed challenges in managing member data efficiently and effectively. The previous data management system, which relied on Microsoft Excel, showed significant limitations in data retrieval, family grouping, and presenting birthday or elderly member information. This study aims to develop a web-based church congregation management information system using the Scrum methodology as an iterative and flexible software development approach. The research methodology includes observation, interviews, literature review, and black box testing. The results indicate that the developed system successfully meets user needs, simplifies congregation data management, and enhances the effectiveness of church administrative services. The implementation of Scrum has proven to be effective in accelerating development processes, accommodating changing requirements, and increasing user involvement through continuous evaluation. This system is expected to be replicable in other churches with similar needs as an integrated digital solution for congregation management.

Predicting Smartphone Addiction Levels with K-Nearest Neighbors Using User Behavior Patterns

M. Rhifky Wayahdi, Fahmi Ruziq — Thu, 16 Oct 2025 00:00:00 +0000

Smartphones have become an integral part of everyday life, but their ever-increasing popularity has raised growing global concerns about excessive use (nomophobia), which impacts quality of life, mental health, and academic performance. Existing research often relies on subjective questionnaires, limiting scalability and objectivity. This study addresses this gap by developing a machine learning model to predict smartphone addiction levels through an objective analysis of user behavior patterns. This research evaluates the effectiveness of the K-Nearest Neighbor (KNN) algorithm, identifies the most influential behavioral features, and assesses the model's classification performance. Using a dataset of 3,300 user behavior entries with 11 features, a waterfall-based framework was employed for data preprocessing, model design, and evaluation. The KNN model achieved 95% accuracy in classifying addiction levels. Permutation Feature Importance analysis confirmed ‘App Usage Time’ and ‘Battery Drain’ as the two most influential predictive features. This study demonstrates that KNN is a powerful and viable method for objectively classifying smartphone addiction. The findings provide a strong foundation for developing scalable, AI-driven early detection and intervention systems, offering significant contributions to the fields of computer science and digital well-being.

Web-Based Diabetes Risk Prediction System Using K-NN on Kaggle Early Stage Diabetes Dataset

Fahmi Ruziq, M. Rhifky Wayahdi — Thu, 16 Oct 2025 00:00:00 +0000

Diabetes mellitus affects approximately 537 million adults globally, and its rising prevalence poses serious health and economic burdens. Early detection is crucial to reduce risks of complications and improve patient outcomes. This study aims to design and implement a web-based diabetes risk prediction system using the K-Nearest Neighbors (K-NN) algorithm to support early detection based on symptoms. The system utilizes the Kaggle Early Stage Diabetes Risk Prediction Dataset containing 520 records with 17 symptom attributes and one class label. Data preprocessing includes converting categorical data into numerical values, discretizing age into predefined ranges, and applying min-max scaling to normalize feature values. K-NN classification was conducted with K values of 1, 3, and 5, using the PHP Machine Learning (PHP-ML) library and MySQL database integration. The system achieved its highest accuracy of 93.46% at K = 1. Manual testing confirmed that the system processes symptom inputs correctly and provides predictions consistent with training data. This web-based tool offers an accessible platform for early diabetes risk screening, supporting self-assessment and triage. It demonstrates that PHP-ML can effectively implement machine learning in a web environment and can be further enhanced through parameter optimization and integration with larger, more diverse datasets to strengthen generalization.

Prediction of Life Expectancy of Lung Cancer Patients After Thoracic Surgery Using Decision Tree Algorithm and Adaptive Synthetic Sampling

Thu, 16 Oct 2025 00:00:00 +0000

This research focuses on predicting the life expectancy of lung cancer patients after undergoing thoracic surgery, using a decision tree classification algorithm (C4.5) combined with adaptive synthetic sampling to handle data imbalance. Data imbalance in the lung cancer patient dataset is a major obstacle in obtaining accurate prediction results, especially in identifying minority classes. Data imbalance in the lung cancer patient dataset is a major obstacle in obtaining accurate prediction results, especially in identifying minority classes. By applying ADASYN, the data distribution becomes more even, thus improving the performance of the C4.5 model. The results showed that combining these methods increased the prediction accuracy from 67% to 87%. In addition, the precision, recall, and f1-score for minority classes have significantly improved, which were previously difficult to identify by the model. Thus, combining the C4.5 algorithm and the ADASYN technique proved effective in dealing with the challenge of data imbalance and resulted in better prediction in the case of lung cancer. This study is expected to contribute to the field of medical classification and serve as a reference for further research on similar cases.

IoT-Enabled Real-Time Monitoring and Tsukamoto Fuzzy Classification of Mandar River Water Quality via Web Integration for Sustainable Resource Management

Chairi Nur Insani, Nurhikma Arifin — Thu, 16 Oct 2025 00:00:00 +0000

This study presents the design and implementation of a real-time water quality monitoring system that utilizes pH, Total Dissolved Solids (TDS), and turbidity sensors, integrated with an ESP32 microcontroller. Sensor data are processed using the Tsukamoto fuzzy logic method to classify river water suitability into two categories: Suitable and Not Suitable. This approach effectively addresses imprecise and uncertain data, thereby producing more reliable classifications compared to conventional threshold-based methods. System validation was conducted through field testing over seven consecutive days at four different times of the day (morning, midday, afternoon, and evening), with results demonstrating stable performance. Recorded pH values ranged from 7.02 to 9.96, TDS values from 140 to 176 ppm, and turbidity levels between 4.00 and 5.15 NTU, indicating that the Mandar River remains within safe limits for daily use. The novelty of this study lies in the direct implementation of the Tsukamoto fuzzy logic method on a resource-constrained IoT device (ESP32), enabling edge-level classification with low latency and without full reliance on cloud computing. The system is designed to maintain decision reliability even under fluctuating sensor data, thus offering a practical and integrated solution for real-time monitoring. The main contribution of this work to computer science is the demonstration of lightweight embedded intelligent algorithms capable of running on constrained devices, the reinforcement of Explainable AI through transparent linguistic rules, and the integration of IoT with edge computing to support sustainable resource management in real-time.

A BiLSTM-Based Approach For Speech Emotion Recognition In Conversational Indonesian Audio using SMOTE

Nariswari Nur Shabrina, Fatan Kasyidi, Ridwan Ilyas — Thu, 16 Oct 2025 00:00:00 +0000

Speech Emotion Recognition (SER) identifies human emotions through voice signal analysis, focusing on pitch, intonation, and tempo. This study determines the optimal sampling rate of 48,000 Hz, following the Nyquist-Shannon theorem, ensuring accurate signal reconstruction. Audio features are extracted using Mel-Frequency Cepstral Coefficients (MFCC) to capture frequency and rhythm changes in temporal signals. To address data imbalance, Synthetic Minority Over-sampling Technique (SMOTE) generates synthetic data for the minority class, enabling more balanced model training. A One-vs-All (OvA) approach is applied in emotion classification, constructing separate models for each emotion to enhance detection. The model is trained using Bidirectional Long Short-Term Memory (BiLSTM), capturing contextual information from both directions, improving understanding of complex speech patterns. To optimize the model, Nadam (Nesterov-accelerated Adaptive Moment Estimation) is used to accelerate convergence and stabilize weight updates. Bagging (Bootstrap Aggregating) techniques are implemented to reduce overfitting and improve prediction accuracy. The results show that this combination of techniques achieves 78% accuracy in classifying voice emotions, contributing significantly to improving emotion detection systems, especially for under-resourced languages.

Hybrid Model for Speech Emotion Recognition using Mel-Frequency Cepstral Coefficients and Machine Learning Algorithms

Odi Nurdiawan, Dian Ade Kurnia, Dadang Sudrajat, Irfan Pratama — Thu, 16 Oct 2025 00:00:00 +0000

Speech Emotion Recognition (SER) is a subfield of affective computing that focuses on identifying human emotions through voice signals. Accurate emotion classification is essential for developing intelligent systems capable of interacting naturally with users. However, challenges such as background noise, overlapping emotional features, and speaker variability often reduce model performance. This study aims to develop a lightweight hybrid SER model by combining Mel-Frequency Cepstral Coefficients (MFCC) as feature representations with three machine learning algorithms: Support Vector Machine (SVM), Decision Tree (DT), and K-Nearest Neighbors (KNN). The methodology involves audio data preprocessing, MFCC-based feature extraction, and classification using the selected algorithms. The RAVDESS dataset, consisting of 1,440 English-language audio samples across four emotions (happy, angry, sad, neutral), was used with an 80/20 train-test split to ensure class balance.. Experimental results show that the KNN model achieved the highest performance, with an accuracy of 78.26%, precision of 85.09%, recall of 78.26%, and F1-score of 77.06%. The Decision Tree model produced comparable results, while the SVM model performed poorly across all metrics. These findings demonstrate that the proposed hybrid approach is effective for recognizing emotions in speech and offers a computationally efficient alternative to deep learning models. The integration of MFCC features with multiple machine learning classifiers provides a robust framework for real-time emotion recognition applications, especially in environments with limited computing resources.

Sentiment Analysis Of Indihome Service Based On Geo Location Using The Bert Model On Platform X

Robiatul Adawiyah Siregar, Fitriyani, Lazuardy Syahrul Darfiansa — Thu, 16 Oct 2025 00:00:00 +0000

The rapid growth of internet usage in Indonesia has led more people to express their feelings, whether positive or negative, about online services, including IndiHome, through social media platforms such as X (formerly Twitter). This study aims to analyze public sentiment toward IndiHome services based on geographic location using the IndoBERT natural language processing model. The data consists of 3.307 Indonesian tweets that are geo-tagged and categorized into three sentiment groups: good, okay, and bad. The research process involves collecting the data, cleaning it (organizing and splitting words), and testing the IndoBERT model with a confusion matrix and classification scores. The findings reveal that negative feelings are more prevalent in most locations, especially in Java. The IndoBERT model achieved its highest accuracy of 80% in detecting negative sentiment. However, there is still room for improvement in distinguishing between positive and neutral sentiments, possibly due to data imbalance. The study shows how combining sentiment analysis with geo-location can provide strategic insights to service providers. In practical terms, these insights can help IndiHome prioritize infrastructure upgrades, improve customer support in areas with high dissatisfaction, and assist policymakers in promoting fairer digital access across regions. Beyond these practical implications, this study also contributes to the field of informatics by demonstrating the application of a transformer-based NLP model (IndoBERT) combined with geo-tagged data for regional sentiment mapping- a relatively unexplored approach in the Indonesian context. The integration of geospatial analysis with sentiment classification offers methodological advances for NLP-based service evaluation beyond business applications.

Analysis of Technology Adoption Factors in Learning among Vocational Students using UTAUT2 Model

Bambang Harimanto, Berlilana, Azhari Shouni Barkah — Thu, 16 Oct 2025 00:00:00 +0000

Technology acceptance in vocational education is a key factor in supporting the effectiveness of teaching and learning processes in the digital era. This study aims to analyze the factors influencing technology acceptance among students of the Computer and Network Engineering (TKJ) Department at SMK Ma'arif 1 Kroya using the Unified Theory of Acceptance and Use of Technology 2 (UTAUT2) framework. The model includes the variables Performance Expectancy, Effort Expectancy, Social Influence, Facilitating Conditions, Hedonic Motivation, Price Value, Habit, Behavioral Intention, and Actual Usage. The results reveal that five key variables—Performance Expectancy, Effort Expectancy, Social Influence, Hedonic Motivation, and Price Value—significantly influence Behavioral Intention, while Habit, Facilitating Conditions, and Behavioral Intention directly affect Actual Usage. All constructs in the model meet validity and reliability criteria, and no multicollinearity was detected (VIF < 3.3). The coefficient of determination (R²) values of 0.612 for Behavioral Intention and 0.673 for Actual Usage indicate strong predictive power of the model. These findings confirm the relevance of the UTAUT2 framework for understanding and enhancing technology acceptance in vocational education settings and provide valuable insights for improving technology integration in technical learning environments.

Implementation of Clustering on Packaged Coffee Sales Data for Simulating Goods Entry in Sole Proprietorship Businesses

Ayu Anjar Paramestuti, Bangun Wijayanto, Mochammad Agri Triansyah — Thu, 16 Oct 2025 00:00:00 +0000

In retail businesses operating under the sole proprietorship structure, decision-making regarding partnerships with beverage distributors—especially those offering packaged coffee—remains a challenge. Store owners often face uncertainty about the profitability of accepting product offerings, which can lead to suboptimal inventory decisions. This study addresses that issue by simulating goods entry scenarios and applying clustering techniques to historical packaged coffee sales data, enabling data-driven insights into product performance and distributor value. Studies focusing on clustering within retail include segmenting customer behavior and stock management strategies, yet many lacked specific application to single owner businesses and product-centric simulations. This research is novel in its contextual focus on packaged coffee distribution within sole proprietorship environments, integrating real sales metrics and clustering algorithms to empower store owners with actionable evaluation tools. Results demonstrate that clustering reveals patterns of profitable product categories and distributor consistency, offering scalable insights for micro-retail optimization. The findings provide a framework that differs from prior studies by emphasizing the intersection between small business dynamics and algorithmic decision support.

Comparing BERTBase, DistilBERT and RoBERTa in Sentiment Analysis for Disaster Response

Hafiz Budi Firmansyah, Aidil Afriansyah, Valerio Lorini — Thu, 16 Oct 2025 00:00:00 +0000

Social media platforms are vital for real-time communication during disasters, providing insights into public emotions and urgent needs. This study evaluates the performance of three transformer-based models—BERTBase, DistilBERT, and RoBERTa—for sentiment analysis on disaster-related social media data. Using a multilingual dataset sourced from the Social Media for Disaster Risk Management (SMDRM) platform, the models were assessed on classification metrics including accuracy, precision, recall, and weighted F1-score. The results show that RoBERTa consistently outperforms the others in classification performance, while DistilBERT offers superior computational efficiency. The analysis highlights the trade-offs between model accuracy and runtime, emphasizing RoBERTa's suitability for scenarios prioritizing accuracy, and DistilBERT's potential in time-sensitive or resource-constrained applications. These findings support the integration of sentiment analysis into disaster response systems to enhance situational awareness and decision-making.

Classification of Worker Productivity and Resource Allocation Optimization with Machine Learning: Garment Industry

Thu, 16 Oct 2025 00:00:00 +0000

This study presents an integrated predictive–prescriptive framework for improving workforce management in the garment industry by combining machine learning classification with linear programming optimization. Using a publicly available dataset of 1,197 production records, productivity levels were categorized into low, medium, and high classes. Data preprocessing included handling missing values, one-hot encoding of categorical variables, and class balancing using SMOTE. Eleven classification algorithms were evaluated, with LightGBM achieving the highest performance (accuracy 78.3%, weighted F1-score 78.3%, Cohen’s Kappa 63.4%) after hyperparameter tuning via Bayesian Optimization. The optimized model’s predictions were then incorporated into a linear programming model, implemented with PuLP, to maximize the allocation of high-productivity workers across production departments under capacity constraints. The results yielded an allocation plan assigning 117 high-productivity workers, significantly enhancing potential operational efficiency. The novelty of this work lies in integrating an optimized ensemble learning model with mathematical programming for end-to-end productivity classification and resource allocation, a combination rarely explored in labor-intensive manufacturing contexts. This framework offers a scalable decision-support tool for data-driven workforce planning and could be adapted to other manufacturing domains with similar operational structures.

IT Governance through Mathematical Modeling: A Quantitative Assessment of Maturity Using Factor Analysis and Structural Equation Modeling

Richardus Eko Indrajit, Erick Dazki, Rido Dwi Kurniawan, Januponsa Dio F — Thu, 16 Oct 2025 00:00:00 +0000

IT Governance (ITG) ensures an organization's technological capabilities align with its business strategy. Although frameworks like COBIT 2019 offer structured guidelines, many assessment techniques rely on qualitative measures, which can compromise objectivity. This paper proposes a novel quantitative approach that integrates Factor Analysis (FA) and Structural Equation Modeling (SEM) to measure IT Governance maturity. By mapping each COBIT 2019 domain—EDM, APO, BAI, DSS, and MEA—onto a latent construct, organizations gain empirical insights into their governance status. Exploratory and confirmatory factor analyses validate these domains, while SEM reveals the magnitude and significance of each domain's impact on overall IT Governance maturity. A real-world example from a financial services company, "FinServEU," demonstrates how this framework can prioritize improvements, enhance regulatory compliance, and promote continuous monitoring. The results highlight that quantitative ITG modeling provides a reliable basis for informed decision-making and optimal resource allocation, bridging the gap between broad qualitative assessments and actionable strategies. This approach is crucial for the field of informatics and computer science, as it offers a robust, reproducible, and objective framework for evaluating a key aspect of digital transformation, ensuring that technological progress is guided by sound, data-driven principles.

Automated Video Recognition of Traditional Indonesian Dance Using Hyperparameter-Tuned Convolutional Neural Network

Santi Purwaningrum, Agus Susanto, Hera Susanti, Mohammed Ayad Alkhafaji — Thu, 16 Oct 2025 00:00:00 +0000

Traditional Indonesian dances serve as a vital expression of cultural identity and regional heritage, yet their preservation through intelligent video recognition remains limited due to technical challenges in motion complexity, costume variation, and the lack of annotated datasets. Prior research commonly employed Convolutional Neural Networks (CNNs) with manually defined hyperparameters, which often resulted in overfitting and poor adaptability when applied to dynamic and real-world video inputs. To overcome these limitations, this study proposes a robust and adaptive classification framework utilizing a hyperparameter-tuned CNN model. The approach automatically optimizes key training parameters such as learning rate, batch size, optimizer type, and epoch count through iterative experimentation, thereby maximizing the model’s ability to generalize across both static and temporal data domains. The model was trained using image datasets representing three traditional dances (Gambyong, Remo, and Topeng), and subsequently tested on segmented frames extracted from YouTube videos. Results indicate strong model performance, achieving 99.67% accuracy on the training set and 100% accuracy, precision, recall, and F1-score across all testing videos. The proposed method successfully bridges the gap between still-image learning and real-world motion recognition, making it suitable for practical applications in digital archiving and cultural documentation. This study’s contribution lies not only in the model’s technical effectiveness but also in its support for preserving intangible cultural assets through intelligent and automated video-based recognition. Future work may incorporate temporal modelling or multi-camera perspectives to further enrich motion understanding and extend the system to broader performance domains.

Comparative Analysis of LSTM and GRU for River Water Level Prediction

Thu, 16 Oct 2025 00:00:00 +0000

Accurate river water level prediction is essential for flood management, especially in tropical areas like Palembang. This study systematically analyzes the performance of two deep learning models, Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), for real-time water level forecasting using hourly rainfall and water level data collected from automatic sensors. A series of experiments were conducted by varying window sizes (10, 20, 30) and the number of layers (1, 2, 3) for both models, with model performance assessed using RMSE, MAE, MAPE, and NSE. The results demonstrate that both window size and network depth significantly influence prediction accuracy and computational efficiency. The LSTM model achieved its highest accuracy with a window size of 30 and a single layer, while the GRU model performed best with a window size of 20 and two layers. This work contributes by systematically analyzing hyperparameter configurations of LSTM and GRU models on hourly rainfall and water level time series for flood-prone regions, offering empirical insight into parameter tuning in recurrent neural architectures for hydrological forecasting. These findings highlight the importance of careful parameter selection in developing reliable early warning systems for flood risk management.

Enhancing Customer Purchase Behavior Prediction Using PSO-Tuned Ensemble Machine Learning Models

Princess Iqlima Kafilla, Fandy Setyo Utomo, Giat Karyono — Thu, 16 Oct 2025 00:00:00 +0000

Predicting customer purchase behavior remains a significant challenge in e-commerce and marketing analytics due to its complex and nonlinear patterns. This study introduces a machine learning framework that integrates ensemble learning models with Particle Swarm Optimization (PSO) for hyperparameter tuning to improve classification accuracy and class discrimination. Several ensemble algorithms, including CatBoost, XGBoost, LightGBM, AdaBoost, and Gradient Boosting, were compared against a baseline Logistic Regression model, both with default and PSO-optimized configurations. Experiments on a real-world e-commerce dataset containing behavioral and demographic variables showed that ensemble methods substantially outperformed traditional models across accuracy, F1-score, and ROC AUC metrics. Notably, the PSO-tuned Gradient Boosting model achieved the highest ROC AUC of 0.9547, improving the AUC by approximately 0.0076 compared to its default configuration, while CatBoost obtained the highest overall accuracy and F1-score. PSO optimization was especially effective in enhancing simpler models such as Logistic Regression but showed marginal gains and some convergence instability in more complex ensemble models. Feature importance analyses consistently identified variables such as time spent on the website, discounts availed, age, and income as key drivers of purchase intent. These findings demonstrate the benefit of combining ensemble learning with metaheuristic optimization, offering actionable insights for developing robust, data-driven marketing strategies.

PROTEGO: Improving Breast Cancer Diagnosis with Prototype-Contrastive Autoencoder and Conformal Prediction on the WDBC Dataset

Marselina Endah Hiswati, Mohammad Diqi — Thu, 16 Oct 2025 00:00:00 +0000

Breast cancer remains one of the leading causes of mortality among women, making accurate and trustworthy early detection a critical challenge in healthcare. To address this, we propose PROTEGO, a Prototype-Contrastive Autoencoder with integrated Conformal Prediction, designed to achieve both high diagnostic accuracy and reliable uncertainty quantification. The framework combines dual-head autoencoding, supervised contrastive learning, prototype-based regularization, and conformal calibration to generate discriminative yet interpretable representations. Using the Wisconsin Diagnostic Breast Cancer (WDBC) dataset, PROTEGO was trained and evaluated through stratified data splits, with performance measured by AUROC, AUPRC, F1-score, Balanced Accuracy, Brier score, calibration error, and conformal coverage metrics. The results show that PROTEGO achieves highly competitive performance with an AUROC of 0.992 and an AUPRC of 0.995, while uniquely providing conformal coverage guarantees with an average set size close to one and more than 92% decisive predictions. Ablation studies confirm the complementary role of each component in enhancing both accuracy and calibration. These findings demonstrate that integrating prototype-guided representation learning with conformal prediction establishes a clinically meaningful diagnostic framework. PROTEGO highlights the importance of unifying precision and reliability in medical AI, offering a step toward more interpretable, safe, and clinically trustworthy systems for breast cancer detection.

Hyperparameter Optimization Of IndoBERT Using Grid Search, Random Search, And Bayesian Optimization In Sentiment Analysis Of E-Government Application Reviews

Angga Iskoko, Imam Tahyudin, Purwadi — Thu, 16 Oct 2025 00:00:00 +0000

User reviews on Google Play Store reflect satisfaction and expectations regarding digital services, including E-Government applications. This study aims to optimize IndoBERT performance in sentiment classification through fine-tuning and hyperparameter exploration using three methods: Grid Search, Random Search, and Bayesian Optimization. Experiments were conducted on Sinaga Mobile app reviews, evaluated using accuracy, precision, recall, F1-score, learning curve, and confusion matrix. The results show that Grid Search with a learning rate of 5e-5 and a batch size of 16 provides the best results, with an accuracy of 90.55%, precision of 91.16%, recall of 90.55%, and F1-score of 89.75%. The learning curve indicates stable training without overfitting. This study provides practical contributions as a guide for improving IndoBERT in Indonesian sentiment analysis and as a foundation for developing NLP-based review monitoring systems to enhance public digital services.

Enhancing Chronic Kidney Disease Classification Using Decision Tree And Bootstrap Aggregating: Uci Dataset Study With Improved Accuracy And Auc-Roc

Thu, 16 Oct 2025 00:00:00 +0000

Chronic Kidney Disease (CKD) is a progressive medical disorder that requires timely and precise identification to avoid permanent impairment of kidney function. However, Decision Tree models, although widely used in clinical applications due to their transparency, ease of implementation, and ability to handle both categorical and numerical data, are prone to overfitting and instability when applied to small or imbalanced datasets. The purpose of this study is to optimize CKD classification by integrating Bootstrap Aggregating (Bagging) with Decision Tree to enhance accuracy and robustness. The methodology involves testing two model variants a standalone Decision Tree and a Bagging-supported Decision Tree using 10-fold cross-validation and evaluating performance with accuracy, precision, recall, F1-score, and the area under the ROC curve (AUC-ROC). Findings reveal that Bagging enhances model accuracy from 0.980 to 0.987, raises precision from 0.976 to 1.000, and improves recall from 0.954 to 0.954, and increases F1-score from 0.965 to 0.976. These results demonstrate that Bagging significantly improves the reliability and generalizability of Decision Tree classifiers, making them more effective for CKD prediction.

Implementation of Extra Trees Classifier and Chi-Square Feature Selection for Early Detection of Liver Disease

Thu, 16 Oct 2025 00:00:00 +0000

The imbalanced distribution of medical data poses challenges in accurately detecting liver disease, which is crucial as symptoms often remain unnoticed until advanced stages. This study examines the application of the Extra Trees Classifier algorithm and chi-square feature selection for early detection of liver disease. Compared to traditional methods like Random Forest and SVM, the Extra Trees Classifier offers enhanced computational efficiency and better handling of imbalanced datasets, while chi-square feature selection helps identify the most relevant medical indicators. The data consists of five medical variables likely to be laboratory test results from patient samples, with labels indicating classes A and B. The data is randomly divided with a ratio of 80% for each class. To address data imbalance, SMOTE technique was applied before the data was randomly split into a ratio of 80% for training and 20% for testing to ensure effective learning and testing of the model's performance. The results showed that with the help of chi-square feature selection, the Extra Trees Classifier algorithm could provide fairly accurate predictions in liver disease classification, with an accuracy of 82.6%, sensitivity of 85.5%, precision of 78.3%, and F1-Score of 81.7%. These results demonstrate significant improvement over existing methods, and the proposed approach can aid healthcare practitioners in making timely diagnostic decisions, potentially reducing mortality rates through early intervention in liver disease cases.

Improving the Performance of Machine Learning Classifiers in Sentiment Analysis of Jenius Application Using Latent Dirichlet Allocation in Text Preprocessing

Vincentius Riandaru Prasetyo, Njoto Benarkah, Bayu Aji Hamengku Rahmad — Thu, 16 Oct 2025 00:00:00 +0000

Sentiment analysis aims to classify a person’s opinion into a specific sentiment, such as positive or negative. The choice of preprocessing used can influence the performance of a sentiment analysis model. The Latent Dirichlet Allocation (LDA) method, commonly used for topic modelling, can be employed as an additional preprocessing step to identify relevant words associated with a particular sentiment label. This study aims to assess whether the LDA method, implemented in the preprocessing stage, can enhance the performance of machine learning models, including Naïve Bayes, Decision Tree, KNN, Logistic Regression, and SVM. This study utilized a dataset comprising 1,800 reviews, with 900 labelled as positive and 900 as negative. Words with an LDA score of at least 0.15 were given additional weight in the TF-IDF stage before model training. After the model was developed, evaluation was carried out by calculating accuracy, precision, recall, and F1-score. The use of LDA in preprocessing improved the performance of all classification models by 1-3% across most evaluation metrics. Specifically, the Logistic Regression model achieved the best performance, followed by SVM and KNN. This performance improvement is aligned with the use of LDA to reduce semantic noise and improve feature representation. Furthermore, this research is also helpful for monitoring customer opinions in the digital banking sector, enabling the rapid and accurate identification of priority issues. Further research could explore the comparison of performance with other topic modelling and feature extraction methods, as well as expanding the dataset and utilizing multiclass models.

Comparison of IndoNanoT5 and IndoGPT for Advancing Indonesian Text Formalization in Low-Resource Settings

Thu, 16 Oct 2025 00:00:00 +0000

The rapid growth of digital communication in Indonesia has led to a distinct informal linguistic style that poses significant challenges for Natural Language Processing (NLP) systems trained on formal text. This discrepancy often degrades the performance of downstream tasks like machine translation and sentiment analysis. This study aims to provide the first systematic comparison of IndoNanoT5 (encoder-decoder) and IndoGPT (decoder-only) architectures for Indonesian informal-to-formal text style transfer. We conduct comprehensive experiments using the STIF-INDONESIA dataset through rigorous hyperparameter optimization, multiple evaluation metrics, and statistical significance testing. The results demonstrate clear superiority of the encoder-decoder architecture, with IndoNanoT5-base achieving a peak BLEU score of 55.99, significantly outperforming IndoGPT's highest score of 51.13 by 4.86 points—a statistically significant improvement (p<0.001) with large effect size (Cohen's d = 0.847). This establishes new performance benchmarks with 28.49 BLEU points improvement over previous methods, representing a 103.6% relative gain. Architectural analysis reveals that bidirectional context processing, explicit input-output separation, and cross-attention mechanisms provide critical advantages for handling Indonesian morphological complexity. Computational efficiency analysis shows important trade-offs between inference speed and output quality. This research advances Indonesian text normalization capabilities and provides empirical evidence for architectural selection in sequence-to-sequence tasks for morphologically rich, low-resource languages.