Jurnal Teknik Informatika (Jutif)

Information Gain-Based Feature Selection and Machine Learning Classification for DDoS Attack Variant Detection in Cloud Computing Environment

Eko Arip Winanto — 2026-06-15

Cloud computing environments face significant security vulnerabilities from Distributed Denial of Service (DDoS) attacks, which can cause system failures and service disruptions. Despite various existing detection methods, challenges remain regarding high computational overhead and suboptimal accuracy due to redundant features in complex datasets. This study aims to identify the optimal feature subset and evaluate its impact on detection performance across multiple machine learning algorithms for multi-class DDoS variants. The research methodology employs a two-stage approach: feature selection using Information Gain (IG) to reduce 47 original features into subsets of 8, 10, 15, and 20, followed by classification using Decision Tree (DT), Random Forest (RF), and Naïve Bayes (NB) on the CICIoT2023 dataset. Experimental results demonstrate that the Decision Tree model with an optimized subset of only 8 features, primarily Inter-Arrival Time (IAT), Header_Length, and Tot_size, achieves a superior accuracy of 99.97%. While Naïve Bayes performs well in binary classification, its accuracy drops significantly to approximately 30% in multiclass settings. This study concludes that IG-based feature selection reduces computational complexity by 30-40% while maintaining high performance across 12 DDoS variants. These findings provide a practical framework for scalable and efficient intrusion detection systems suitable for real-time deployment in resource-constrained IoT-cloud environments.

Classification of Roronoa Zoro Anime, Cosplay, and Action Figure Images Using VGG16 and Inception V3 with Logistic Regression and Support Vector Machine to Improve Popular Culture Object Recognition

Denaldy Oktavian Noor Rizki — 2026-06-15

The diversity of visual representations of anime characters across anime scenes, cosplay photographs, and action figure images poses challenges for automated image classification due to variations in pose, lighting, background, and visual style. This study aims to develop a robust image classification system for the character Roronoa Zoro using deep learning–based feature extraction combined with classical classification algorithms. The method employs VGG16 and Inception V3 as feature extractors, followed by classification using Logistic Regression and Support Vector Machine. The dataset comprises three classes (anime, cosplay, and action figure), processed through image resizing, normalization, and data augmentation. Performance was evaluated using accuracy, F1-score, Area Under Curve (AUC), Matthews Correlation Coefficient (MCC), confusion matrix, silhouette plot, and multidimensional scaling. The experimental results show that Inception V3 combined with Logistic Regression achieved the best performance, with an AUC of 0.993, accuracy of 95.7%, F1-score of 0.957, and MCC of 0.935, outperforming VGG16 with Logistic Regression, which achieved 91.7% accuracy and an AUC of 0.986. Visualization-based evaluation indicates that Inception V3 produces more separable feature representations, particularly in distinguishing cosplay images from anime and action figure categories. This research demonstrates the effectiveness of multi-model feature extraction and classification for improving recognition performance in character-based image classification tasks and contributes empirically to the application of hybrid deep feature–machine learning approaches in computer vision.

Comparative Analysis of Baseline IndoBERT, Class-Weighted IndoBERT, and SMOTE with Support Vector Machine for Handling Imbalanced Sentiment Classification in Indonesian

Riya Widayanti — 2026-06-15

Imbalanced data distribution is a common issue in Indonesian sentiment classification and significantly affects the performance of classification models. This study investigates three approaches, namely SMOTE combined with Support Vector Machine (SMOTE + SVM), Baseline IndoBERT, and Class-Weighted IndoBERT. The dataset consists of Google Maps reviews, which are categorized into positive, neutral, and negative sentiments. Prior to model training, the data undergo preprocessing steps including cleaning, normalization, and tokenization. Model performance is evaluated using confusion matrix analysis and macro-averaged F1-score. The results show that Baseline IndoBERT achieves a macro F1-score of 0.598, followed by Class-Weighted IndoBERT with 0.582, while SMOTE + SVM obtains the lowest performance at 0.545. Despite having slightly lower overall performance, Class-Weighted IndoBERT demonstrates a more balanced capability in recognizing minority classes. These findings indicate that incorporating class-weighting mechanisms into transformer-based models can help mitigate bias toward majority classes and improve minority class recognition. From a scientific perspective, this study provides empirical evidence on how imbalance-aware learning strategies influence the behavior of transformer-based models in imbalanced text classification tasks. Furthermore, this study highlights the importance of using macro-averaged evaluation metrics to ensure a more comprehensive and fair assessment of model performance, particularly in low-resource and imbalanced language settings.

VGG-16 Transfer Learning for Accurate Classification of Three Local Durian Varieties Using Leaf Morphology Images

Ahmad Haikal Nuqqy Zahhar — 2026-06-15

Durian (Durio zibethinus Murr), recognized as the "king of fruits" in Southeast Asia, represents a significant genetic asset for Indonesian agriculture with high economic value. East Java leads national production, contributing 580.5 thousand tons (29.59%) of the total 19.6 million tons in 2024. However, local durian quality faces persistent challenges due to minimal maintenance practices and farmers' limited expertise in variety identification. Manual taxonomic identification based on leaf morphology requires specialized knowledge, is time-consuming, and prone to subjective errors, particularly for three popular Nganjuk varieties—local, montong, and lai—which exhibit similar leaf characteristics. Previous studies have addressed durian classification using fruit images or disease detection on leaves, but a research gap exists for variety classification specifically using leaf images with deep learning approaches. This study implements VGG-16 transfer learning architecture with ImageNet pre-trained weights to classify three durian varieties based on leaf morphology images. A dataset of 600 high-resolution images (2048×2048 pixels, 200 per class) was collected from Nganjuk orchards following standardized protocols and validated by three independent experts (two experienced farmers and one plant taxonomist), achieving substantial inter-annotator agreement (Fleiss' kappa = 0.87). Preprocessing included resizing to 224×224 pixels with bilinear interpolation, normalization to [0,1], and standardization using ImageNet statistics. Data augmentation through random rotation (±30°), horizontal flipping (48.8% probability), contrast adjustment (±50.1%), and width/height shifting (±12%) expanded the dataset fourfold to 2,400 images. Using a 90:10 train-test split (2,160:240), the VGG-16 model trained with Adam optimizer (learning rate 0.001, dropout 0.5, dense layer 256 units) achieved 97.08% accuracy after 4 epochs in 1.11 minutes. Performance metrics demonstrated high precision (0.93-1.00), recall (0.92-1.00), and F1-scores (0.95-0.99) across all classes. This research advances precision agriculture informatics by providing an automated, reliable tool for durian variety identification, supporting farmers in optimal cultivation decisions, quality control, and economic value enhancement while contributing to sustainable agricultural development and the Center for Plant Variety Protection and Agricultural Licensing (PVTPP) registration systems in Indonesia.

Artificial Intelligence-Based Aircraft Detection for Enhanced Aviation Safety and Air Traffic Management

Astika Ayuningtyas — 2026-06-15

The rapid growth of international air traffic has made maintaining aviation safety and managing air traffic efficiently increasingly complex, particularly in identifying aircraft in constantly changing airspace. Traditional monitoring systems such as radar and Automatic Dependent Surveillance-Broadcast (ADS-B) have limitations in operating at low altitudes, in adverse weather, and in overcrowded environments, which can reduce the ability to understand surrounding conditions. This research proposes an artificial intelligence-based visual detection system aimed at enhancing real-time aircraft identification and improving air traffic monitoring. The system uses a YOLO-based deep learning model enhanced with a special attention mechanism and data augmentation to increase accuracy, flexibility, and operational resilience. The dataset used covers various flight situations, such as variations in light, viewing angles, and background complexity, to train the model. The model's test results show that it can correctly identify 95.24% of passenger planes, 92.4% of blimps, and 90% of fighter planes. The average overall precision (mAP) is over 90%. This system is also capable of real-time inference with precision and recall consistently above 85% under various conditions. Compared with conventional vision-based detection methods, this system demonstrates superior localization capabilities and robustness, making it suitable for use in real-world flight surveillance and air traffic management. In conclusion, this AI-based framework provides a practical and scalable solution that can improve flight safety and promote smarter air traffic management.

Development and Comparative Evaluation of Machine Learning Models using Clinically Relevant Features for Predicting Newborn Patients’ Length of Stay

Gandung Triyono — 2026-06-15

The Length of Stay (LOS) of newborns is a crucial indicator for healthcare management and hospital resource allocation. However, prior research has yet to systematically compare machine learning models for newborn LOS prediction using clinically pertinent features in developing-country hospital contexts, creating an important methodological and contextual gap. Accurate prediction of LOS is urgently needed to support timely clinical decision-making and prevent overcrowding, inefficiencies, and unnecessary healthcare costs. This study aims to identify factors influencing LOS and develop a predictive model for newborn LOS using several machine learning algorithms. A comparison was conducted among Linear Regression, Random Forest Regression, Support Vector Regression (SVR), and Artificial Neural Networks (ANN). The dataset consisted of medical records of newborn patients from three private hospitals in Indonesia. The research included data collection and understanding, data preprocessing, modeling, and evaluation. Experimental results show that Random Forest Regression achieved the best predictive performance, with MAE = 0.019, MSE = 0.011, RMSE = 0.086, and R² = 0.987. Feature importance analysis revealed that gender, referral source, insurance type, and diagnosis were the most influential predictors of LOS. This study contributes to the advancement of machine learning applications in healthcare data analytics and provides evidence-based insights to support neonatal care planning and hospital resource optimization.

Optimization of Machine Learning Model using Grid and Random Search Algorithms for Predicting Student Dropout

M. Faris Al Hakim — 2026-06-15

Student dropout is a serious problem that can affect the quality of education and operational efficiency of higher education institutions. Early prediction of potential students who will dropout is essential to develop appropriate intervention strategies, so as to increase graduation rates and reduce the negative impact on academic continuity. A better model for student dropout prediction becomes an objective of this research. The method used in this research is to improve the performance of machine learning models through the selection of optimal hyperparameters. The research methodology consists of several stages, including data preprocessing, handling imbalanced data, model training, and performance evaluation. There are three machine learning models used in this research, namely XGBoost, AdaBoost, and Random Forest. The selection of optimal hyperparameter values is carried out using the Random Search and Grid Search methods. Model evaluation is conducted using k-fold cross-validation and multiple evaluation metrics, including accuracy, precision, recall, and F1-score. As part of the important results, the combination of XGBoost and Random Search produced the best performance with 91.18% accuracy, indicating that hyperparameter optimization significantly improves predictive performance. The findings of this research explicitly contribute to the field of informatics, particularly educational data mining, and provide insights for educational institutions to identify high-risk dropout students more accurately.

Improving Sentiment Classification of Kredit Pintar Reviews Using IndoBERT, SMOTE, and Stacking Ensemble

Ayu Safitri — 2026-06-15

Kredit Pintar is one of the most widely used fintech applications in Indonesia, generating millions of user reviews on the Google Play Store that reflect diverse user experiences. These reviews provide valuable insights into application performance; however, extracting sentiment from such unstructured and imbalanced textual data remains a challenging task. This study aims to improve sentiment classification of Kredit Pintar user reviews by proposing a hybrid approach that integrates IndoBERT, SMOTE (Synthetic Minority Over-Sampling Technique), and a stacking ensemble model. From 2020 to 2024, 2,278 user reviews were classified into positive, neutral, and negative categories based on star ratings. SMOTE was employed to rectify class imbalance, whereas IndoBERT gathered contextual representations of the Indonesian language. Furthermore, a stacking ensemble combining IndoBERT, Random Forest, and SVM (Support Vector Machine) was implemented to enhance classification performance. Experimental results show that IndoBERT without data balancing achieved an accuracy of 84%, whereas the proposed combination of IndoBERT, SMOTE, and stacking ensemble consistently produced superior performance, achieving 92% accuracy, precision, recall, and F1-score. The findings demonstrate that integrating language-specific transformer models with data balancing and ensemble techniques effectively improves sentiment classification. This study contributes to the advancement of Indonesian-language natural language processing in the fintech domain and provides practical insights for fintech developers in understanding user perceptions and improving digital financial services.

Implementing Proxmox VE-Based High Availability Clustering with Ceph Replication and Performance Testing for Resilient IT Infrastructure in High-Risk Disaster Areas

Muhammad Abdul Muin — 2026-06-15

IT infrastructure in disaster-prone areas, particularly along Java's southern coastal region within the Sunda Arc subduction zone, faces significant vulnerability to seismic events and tsunamis that cause critical system downtime, disrupting emergency coordination and exacerbating disaster impacts. This study aims to develop and validate an open-source High Availability (HA) solution using Proxmox Virtual Environment (PVE) ensuring service continuity with Recovery Time Objective (RTO) under 2 minutes and near-zero Recovery Point Objective (RPO). The methodology encompasses four systematic stages: needs analysis identifying infrastructure requirements and disaster risk assessment for Cilacap region; architecture design implementing three-node PVE cluster with Ceph distributed storage (replication factor 3) and Corosync quorum mechanism; system implementation including network bonding, VLAN segmentation, and dedicated 1Gbps Ceph replication network; and comprehensive performance testing through fault injection scenarios (power-off simulation, network partition, storage failure) measuring inter-node latency, disk I/O performance, and failover recovery metrics. Results demonstrate exceptional reliability with 99.92% availability over 72-hour monitoring, Mean Time Between Failures (MTBF) of 24.1 hours, and Mean Time To Recovery (MTTR) of 70 seconds with total downtime of 3.53 minutes across three failover simulations. Inter-node latency remains below 1ms (average 0.372-0.593ms), while disk I/O latency maintains sub-0.5ms performance during failover events. This research contributes to computer science and disaster informatics by providing a validated, replicable open-source blueprint for resilient IT infrastructure in Indonesia's disaster-prone regions, offering practical implementation pathways for integration with national emergency systems including BNPB coordination networks and BMKG early warning infrastructure.

Classification of Banana Leaf and Ornamental Plant Diseases Using Gray Level Co-occurrence Matrix (GLCM) and Hybrid Random Forest–Support Vector Machine (SVM)

Novia Urfiyati — 2026-06-15

Leaf diseases in banana plants and ornamental crops can significantly reduce productivity and product quality, highlighting the need for accurate early detection methods. This study proposes an image-based classification approach utilizing texture features extracted from the Gray Level Co-occurrence Matrix (GLCM) combined with a Hybrid Stacking model that integrates Random Forest (RF) and Support Vector Machine (SVM). The preprocessing stage involves image resizing and noise reduction, followed by feature extraction using energy, contrast, homogeneity, and correlation parameters. The dataset consists of eight classes of healthy and diseased leaves, collected from both field documentation and secondary sources. Model performance was evaluated using accuracy, precision, recall, and F1-score metrics under a cross-validation scheme. Experimental results show that SVM achieved 89.2% accuracy, RF 88.5%, while the stacking model yielded the best performance with 91.7% accuracy, effectively reducing misclassification among visually similar disease classes. This study demonstrates the effectiveness of combining GLCM features and hybrid stacking models for leaf disease classification, with potential applications in automated plant monitoring systems to support precision agriculture.

Predicting Mental Health Status using a Fine-Tuned CNN-LSTM Hybrid Model

Agustin Agustin — 2026-06-15

Mental health has become a critical global concern in the digital era, particularly as social media platforms increasingly serve as spaces where users express psychological conditions, emotions, and personal struggles. This study aims to predict mental health status from Twitter text using a fine-tuned hybrid CNN–LSTM deep learning model. A total of 12,214 tweets were collected, cleaned, and labeled into five categories: Normal, Stress, Anxiety, Depression, and High-Risk Condition. The dataset was split using stratified sampling into 70% training, 15% validation, and 15% testing portions. Text was transformed into numerical representations through tokenization, padding, and 100-dimensional word embeddings. The hybrid CNN–LSTM architecture combines the CNN’s ability to extract local linguistic features with the LSTM’s strength in capturing long-term contextual dependencies, supported by dropout, early stopping, and hyperparameter fine-tuning. Experimental results show that the hybrid model achieves superior performance compared to standalone CNN and LSTM architectures, obtaining an overall accuracy of 0.892, macro precision of 0.874, macro recall of 0.861, and a macro F1-score of 0.865. Class-wise evaluation indicates that the Normal category achieves the highest accuracy (0.960), followed by Anxiety (0.884) and High-Risk Condition (0.808). Meanwhile, Stress (0.751) and Depression (0.745) show lower accuracies due to semantic overlap in linguistic expressions commonly found on social media. The training process demonstrates stable convergence without significant overfitting, confirming the effectiveness of the selected architecture and training strategy. Overall, this study highlights the effectiveness of the hybrid CNN–LSTM model for early mental health detection based on text data. The findings provide a strong foundation for developing scalable and data-driven mental health monitoring systems in digital environments and contribute to advancing natural language processing approaches for mental health analysis.

Model-Driven Engineering for ARrupiah Cultural AR App: Kano Model and Qualitative User Experience Evaluation

Gerson Feoh — 2026-06-15

The use of Augmented Reality (AR) in learning is becoming increasingly widespread, especially for introducing cultural and historical material in a more interesting way. However, many AR applications are still built without a structured design, making them difficult to develop when content is added. This study uses a Model-Driven Design (MDD) approach to organize the design of the ARrupiah application to make it more modular and easier to expand. After the prototype was completed, testing was conducted through surveys and interviews. The Kano survey involved 50 students to evaluate the main features of the application, while semi-structured interviews were analyzed using NVivo software to explore response patterns and user experiences, with a code saturation level of 80%. The survey results showed that around 70% of the features fell into the Attractive category, with a System Usability Scale (SUS) score of 82/100, indicating ease of use. Qualitative analysis reinforced the quantitative results through a triangulation process, in which features categorized as Attractive also emerged as a dominant theme of visual engagement in the NVivo results. This combined approach strengthens the validity of the findings and provides a more comprehensive understanding of user perceptions and satisfaction. Overall, the application of MDD not only helps refine the technical design but also improves the quality of the learning experience through ARrupiah-based interactive media.

Comparative Evaluation Of Sparse, Dense, And Hybrid Retrieval Models On Indonesian Wikipedia

Tino Saputra — 2026-06-15

This study presents a comparative evaluation of Information Retrieval (IR) models on the Indonesian Wikipedia corpus, focusing on sparse, dense, and hybrid retrieval approaches. The evaluated methods include TF-IDF and BM25 as sparse models, SBERT (MiniLM) as a dense retrieval model, and hybrid retrieval implemented through score fusion. The dataset consists of 713,044 Wikipedia articles, with experiments conducted using 1,000 test queries. Performance is measured using Precision@10 (P@10) and Mean Reciprocal Rank (MRR). The results show that BM25 achieves the highest performance, with a P@10 of 0.973 and an MRR of 0.9174, significantly outperforming TF-IDF and SBERT. Hybrid retrieval provides a slight performance improvement, where the BM25 + SBERT combination reaches a P@10 of 0.979 and an MRR of 0.9253 at higher α values. These findings indicate that lexical matching remains dominant in encyclopedic corpora, while semantic representations provide complementary improvements. However, the performance gain of hybrid retrieval is relatively marginal compared to the additional computational cost introduced by dense embedding and score fusion processes, indicating a trade-off between effectiveness and efficiency. These results highlight that, for low-resource languages such as Indonesian, lexical-based retrieval remains highly reliable, while hybrid approaches provide incremental improvements. Therefore, this study provides practical guidelines for developing efficient, scalable, and reliable Information Retrieval systems for Indonesian Wikipedia and other low-resource language corpora.

Enhancing Flood Area Segmentation in Remote Sensing Images Using Hybrid Attention Mechanism on DeepLabV3+ with ResNet-50 Backbone

Annisa Syifaul Ummah — 2026-06-15

Flooding is caused by climate change and urbanization, so rapid and accurate monitoring is essential in supporting emergency response. However, flood segmentation still faces challenges in dense vegetation. This study aims to improve and analyze the performance of the Hybrid Attention Mechanism in the form of Point-wise spatial attention (PSA) and Squeeze-and-Excitation Block (SE Block) in the DeepLabV3+ architecture with the ResNet-50 backbone. The methods used include collecting a dataset of 600 training and 63 validation, data augmentation, model development and Hybrid Attention Mechanism design, hyperparameter optimization, ablation study, and performance evaluation. The ablation results obtained show the best performance with accuracy of 0.9624, F1-score of 0.9618, IoU (Non-Flood) of 0.9323, IoU (Flood) of 0.9208, and mIoU of 0.9265, surpassing previous studies that used Modified U-Net in detecting floods in dense vegetation. This research contributes to the development of a flood segmentation model based on a hybrid attention mechanism, which is more effective in detecting flooded areas in densely vegetated regions.

Optimizing Breast Cancer Classification: SVM and Random Forest with Hybrid Hyperparameter Tuning and Feature Selection

Adil Setiawan — 2026-06-15

Breast cancer remains one of the leading causes of cancer-related mortality among women worldwide, underscoring the urgent need for early, accurate, and reliable diagnostic support systems. This study proposes an optimized breast cancer classification framework using Support Vector Machine (SVM) and Random Forest (RF) models enhanced through hybrid hyperparameter tuning and feature selection. The Breast Cancer Wisconsin (Diagnostic) dataset, comprising 569 samples with 30 numerical features derived from Fine Needle Aspirate (FNA) examinations, was utilized in this research. Feature selection was conducted using Random Forest feature importance to identify the most relevant diagnostic attributes and reduce dimensionality. Hybrid hyperparameter tuning was implemented using GridSearchCV combined with 5-fold cross-validation to obtain optimal model configurations. Model performance was evaluated using accuracy, malignant-class recall, confusion matrix analysis, and Receiver Operating Characteristic–Area Under the Curve (ROC–AUC). Experimental results show that the optimized SVM model achieved significant improvements in accuracy, recall, and ROC–AUC compared to baseline models, indicating enhanced sensitivity and discrimination capability, while the Random Forest model maintained stable performance with marginal gains after optimization. These findings highlight the critical importance of systematic optimization strategies in improving diagnostic safety and reducing false negatives, thereby contributing to the development of more reliable and clinically applicable machine learning-based medical decision support systems.

Comparison of SVR Parameter Optimization Using Particle Swarm Optimization (PSO) and Random Search for Rice Harvest Yield Prediction

Narlin Yumeivia — 2026-06-15

Rice yield is an important part in a precision agriculture system that can support farmers' decision-making in a more targeted manner. The author's research aims to help farmers and stakeholders in Bambang Village predict crop yields accurately to overcome production fluctuations. Through appropriate efforts and strategies, this technology is expected to improve food security and farmer welfare. The research method uses the Support Vector Regression (SVR) algorithm for the modeling process, with the help of Particle Swarm Optimization (PSO) and Random Search optimization in finding the best parameters. The research dataset includes 1,120 historical data of rice harvests in Bambang Village for the 2022–2023 period tested through 70:30 and 60:40 data sharing scenarios. Model performance is evaluated using the Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and coefficient of determination (R²) metrics. The MAPE metric is used as the main indicator of relative accuracy by measuring the average percentage deviation between predicted values and actual values; a low MAPE value is very significant because it reflects the model has a minimal error rate on a percentage scale, thus providing more precise estimates for farmers. The results showed that both optimization methods successfully identified SVR parameters (C, gamma, epsilon) that followed the data trend. Random Search produced slightly superior R² performance (reaching 82.20% at a 60:40 ratio), while PSO showed more consistent parameter exploration stability. These findings demonstrate that the integration of machine learning and optimization techniques has great potential in strengthening data-driven agricultural systems to improve food security and farmer welfare.

Peningkatkan Keamanan ElGamal Menggunakan CNN dan Rolling Hash untuk Generasi Kunci dalam Enkripsi Gambar

Achmad Fauzi — 2026-06-15

The large scale exchange of digital images requires security mechanisms that are robust not only at the cryptographic algorithm level but also in the key generation process, which is often the weakest component of the system. In conventional ElGamal schemes, security may degrade due to static entropy sources and predictable key patterns. This study proposes an ElGamal key generation model based on a pipeline of Convolutional Neural Networks (CNNs) and a rolling hash function, utilizing visual image content as an adaptive entropy source. The CNN extracts latent features through a fully connected layer, while the rolling hash enhances diffusion and key sensitivity to minor image variations. The model was evaluated using the CIFAR-10 dataset in PNG, WEBP, and JPG formats. Experimental results show stable key generation times ranging from 0.426 to 0.444 ms, with high entropy values between 7.98 and 7.99 bits, indicating strong randomness and resistance to prediction. Strong diffusion characteristics were also observed (PSNR 5.94 dB, SSIM −0.24, MAE 0.43). During encryption, WEBP achieved the fastest processing time (0.48 ms), followed by PNG (1.01 ms) and JPG (15.39 ms), while PNG demonstrated the highest size efficiency with a reduction of up to 70.6%. Decryption remained highly reliable, with success rates exceeding 97% across all formats. Overall, the results confirm that integrating CNNs and rolling hash significantly enhances ElGamal key generation security without compromising decryption reliability or image quality.

Integrating Whale Transaction Flow Scoring with LSTM for Bitcoin Trend Forecasting

Muhammad Ridhwan Hakiki — 2026-06-15

Bitcoin price prediction faces significant challenges due to high volatility and the influence of large holders, known as whales, whose transactions exceeding 500 BTC can affect market behavior. This study develops an LSTM model combining whale transaction sentiment scores with historical Bitcoin OHLC prices to forecast 7-day ahead price movements. The dataset comprises 2,069 whale transactions and 8,761 hourly price observations from April 20, 2024 to April 20, 2025. The scoring mechanism assigns +1 to exchange outflows, -1 to inflows, and 0 to neutral transfers, multiplied by logarithmically normalized transaction amounts. The LSTM architecture consists of two recurrent layers with 128 and 64 memory units, processing 720-hour input sequences to generate 168-hour OHLC forecasts. Training evaluation yielded R² of 0.9386, RMSE of 0.0686, and MAE of 0.0498. Test evaluation produced Mean Absolute Errors ranging from 871.72 USD to 3,482.27 USD across OHLC components. The model correctly predicted upward directional trends but systematically underestimated prices by 2,000-3,000 USD initially and failed to anticipate a 6,422 USD intraday surge on April 22, 2025. Results demonstrate that whale sentiment features enhance directional trend identification but do not enable precise multi-day price point prediction due to sudden market regime changes. These findings contribute empirical evidence that directional sentiment scoring of large-holder transactions provides complementary predictive value beyond conventional price-volume indicators, establishing a methodological foundation for integrating blockchain-native behavioral signals into cryptocurrency forecasting frameworks.

Optimization of ShuffleNetV2 Using Self-Knowledge Distillation for Cocoa Fruit Disease Classification

H.R Merdu Wira Jasa — 2026-06-15

Timely cocoa fruit disease diagnosis is critical for field management, yet manual inspection is subjective and inconsistent, while many accurate deep learning models remain too computationally demanding for practical on-device use. This study aims to optimize cocoa fruit disease classification by applying self-knowledge distillation (Self-KD) to a lightweight ShuffleNetV2 architecture without increasing inference complexity. Using a three-class dataset (healthy, pod borer, and black pod rot) with preprocessing and class balancing, ShuffleNetV2 was selected as the baseline and trained with Self-KD, improving accuracy from 96.84% to 98.34% along with consistent gains in precision, recall, and F1-score. These results indicate that Self-KD provides a learning-level optimization that enhances robustness and prediction stability in lightweight CNNs, which is especially relevant for edge AI deployment in agricultural environments. Therefore, the proposed approach supports efficient, scalable, and sustainability-oriented AI (Green/Sustainable AI) for smart farming, with potential transferability to other crops that exhibit similar visual symptom patterns.

Comparative Analysis of Hyperparameter Optimization Methods for LSTM in Cryptocurrency Price Prediction: An Application to TRX–USD

Dasril Aldo — 2026-06-15

The rapid growth of cryptocurrencies increases the demand for accurate forecasting models to support investment decisions and automated trading systems. This study analyzes and compares the performance of several hyperparameter optimization methods applied to a Long Short-Term Memory (LSTM) model for predicting the price of TRX–USD. The dataset consists of 2,096 daily historical records obtained from the Binance platform, including open, high, low, close, volume, and percentage change, with the closing price selected as the forecasting target. A baseline LSTM model was evaluated against six optimization techniques: Grid Search, Random Search, Bayesian Optimization (Hyperopt), Optuna, Particle Swarm Optimization (PSO), and Genetic Algorithm (GA). Experimental results show that GA provides the best performance with an R² score of 0.88, MAE of 0.0123, RMSE of 0.0189, and a validation loss of 0.069. In contrast, Random Search yields the lowest performance, achieving an R² of only 0.2979. These findings highlight significant performance gaps among optimization strategies and demonstrate the superiority of metaheuristic-based approaches over conventional tuning methods. This research contributes to the advancement of computational intelligence by providing empirical evidence on the effectiveness of hyperparameter optimization techniques for deep learning–based time series forecasting, particularly in high-volatility financial environments.

Forecasting Nutrient Concentration Dynamics in Hydroponic Lettuce Cultivation Using a Hybrid Fuzzy Time Series and Long Short-Term Memory Approach for Internet of Things–Based Systems

Muh. Agus — 2026-06-15

Proper nutrient management is crucial for the optimal growth and yield of hydroponically cultivated lettuce. This study proposes a hybrid time-series forecasting model that integrates Fuzzy Time Series (FTS) and Long Short-Term Memory (LSTM) networks to predict nutrient concentration dynamics in hydroponic lettuce cultivation within an Internet of Things–based environment. Experimental data from four lettuce plant samples with different nutrient treatments (control, 400 PPM, 600 PPM, and 1000 PPM) were analyzed for 26 days, with the prediction extended to 40 days, representing the complete growth cycle using a TDS Sensor as a PPM value reader and a Solenoid Valve to accurately control the PPM value via ESP32 with Internet of Things (IoT) communication. This hybrid model incorporates growth-stage awareness through an adaptive weighting mechanism, resulting in a superior forecasting accuracy. The results showed that the ensemble approach achieved a Mean Absolute Percentage Error (MAPE) of 2.43% for the control, 3.12% for the 400 PPM, 3.45% for the 600 PPM, and 3.78% for the 1000 PPM sample. The 600 PPM treatment showed optimal development with 82% compliance with the recommended PPM range (560-840 ppm). The proposed model provides actionable insights for precision nutrient management, potentially reducing fertilizer use by 23-35% while maintaining crop quality. This study contributes to hybrid intelligent systems and time-series forecasting by demonstrating an effective integration of rule-based fuzzy modeling and deep recurrent neural networks in Internet of Things–driven environments for hydroponic systems, supporting efficient resource utilization and increased crop productivity.

Hybrid Cryptography-Steganography Scheme Based on Camellia-256 and LSB for Enhanced Security and Imperceptibility of Secret Messages

Imam Prayogo Pujiono — 2026-06-15

The development of digital communications has increased the risk of message interception and manipulation, necessitating robust and multi-layered security solutions. This research designs, implements, and evaluates a multi-layered security scheme that integrates cryptography and steganography. The proposed method first encrypts the secret message using the Camellia-256 algorithm in Electronic Codebook (ECB) mode with PKCS#7 padding. The resulting ciphertext is then embedded into the cover image using the Least Significant Bit (LSB) steganography technique. From a practical standpoint, this design provides defense-in-depth for covert communication: encryption preserves confidentiality even if the hidden payload is detected, while steganography reduces the likelihood that the encrypted content is flagged during transmission. This combination mitigates LSB’s weakness against statistical steganalysis by encrypting the payload into ciphertext, thereby reducing structured bit patterns that may otherwise facilitate statistical detection. System performance is quantitatively evaluated using two primary metrics: the Avalanche Effect to measure cryptographic strength and the Peak Signal-to-Noise Ratio (PSNR) to measure the visual imperceptibility of the stego-image. The experimental results demonstrate excellent cryptographic strength, evidenced by an average Avalanche Rate of 54.37%, indicating that minimal changes to the input result in significant changes to the output. Furthermore, the scheme exhibits excellent visual imperceptibility with an average PSNR of 75 dB, making the stego-image visually indistinguishable from the original cover image. It is concluded that the proposed hybrid scheme offers a robust and validated solution for secure message communication, combining content confidentiality through cryptography and message obfuscation through steganography, thus providing dual protection against cybersecurity threats.

Comparative Analysis of Temporal Fusion Transformer and Long Short-Term Memory Architecture Resilience in Predicting Solana Price Volatility Across Different Market Phases

Mahdy Eka Putra — 2026-06-15

Abstract must be written in English. The high volatility of cryptocurrency markets, particularly for altcoins like Solana (SOL), presents a significant challenge for predictive modeling. Traditional deep learning architectures often struggle to adapt to sudden market regime shifts. Therefore, this study aims to provide a comparative analysis of the resilience between the Temporal Fusion Transformer and Long Short-Term Memory architectures in predicting Solana price volatility across three distinct market phases: the bull market of 2024, the bear market of 2025, and the recovery phase of 2026. We utilized hourly historical price and volume data combined with technical indicators such as Relative Strength Index (RSI). The models were evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and a specific performance degradation rate formula. The results demonstrate that while LSTM performs adequately during stable trends, its accuracy degrades massively by 1575.69% during high-volatility regime changes due to memory inertia causing a severe lagging effect. Conversely, the TFT model exhibited superior resilience, limiting its performance degradation to only 218.53% during the extreme bear market phase. The inherent attention mechanism and skip connections in TFT allow it to dynamically adapt to sudden structural breaks in real-time without delay. Furthermore, the implementation of the TFT architecture proved to be 62% more computationally efficient than LSTM. This research significantly contributes to the field of computer science and informatics, specifically in adaptive time-series forecasting, by proving that attention mechanisms and skip connections can efficiently solve the memory inertia problem in recurrent networks during real-time structural breaks.

Comparative Evaluation of ARIMA, LSTM, Hybrid ARIMA-GARCH, and Hybrid GARCH-LSTM Models for Daily Bitcoin and Gold Price Forecasting

Isna Nurul Fatatik — 2026-06-15

The volatile nature of digital financial markets poses major challenges for predictive modelling, particularly in developing accurate forecasting models that can address diverse asset characteristics such as Bitcoin, with its extreme fluctuations, and Gold, which is known for its stable movements. This study addresses this challenge by evaluating the robustness of linear, deep learning, and hybrid architectures in both high-volatility and stable asset environments. Utilizing Bitcoin and Gold closing price data from 2022 to 2025, the methodology adopts a comparative workflow that involves ARIMA, ARIMA-GARCH, LSTM, and LSTM-GARCH Hybrid models. Stationarity (ADF) and heteroskedasticity (ARCH-LM) diagnostics alongside AIC/BIC selection criteria were applied, followed by a walk-forward validation scheme to assess the model's performance. Results confirmed that the hybrid GARCH-LSTM model delivered the lowest Root Mean Squared Error (RMSE), significantly outperforming single models by integrating statistical variance and temporal neural learning. Therefore, this study contributes to the field of computational intelligence by validating an accurate Artificial Intelligence (AI) framework for volatility-based forecasting and proposing a scalable blueprint for engineers to develop models that are capable of capturing the dynamics of financial time series data.

Interpretable and Statistically Validated Comparative Evaluation of EfficientNetB0, MobileNetV2, and ResNet50 for Bold and Natural Makeup Classification on CelebA

Aurelia Chiara Suryabangun — 2026-06-15

Facial makeup classificationplays a critical role in beauty technology, visual style analysis, and intelligent web-based image inference. Distinguishing bold makeup from natural makeup is challenging due to subtle visual overlap, borderline facial appearance, and inconsistent makeup intensity across images. While numerous prior studies have applied deep learning for facial analysis, most focus solely on conventional performance metrics without addressing statistical validation, probability calibration, or interpretability — a critical gap that limits reliable model selection in visually subtle classification tasks. This study presents an interpretable and statistically validated comparative evaluation of three transfer learning architectures — EfficientNetB0, MobileNetV2, and ResNet50 — for binary makeup classification using a curated CelebA-based dataset. The final dataset comprises 12,000 facial images equally divided into natural_makeup and bold_makeup classes, with separate training, validation, and clean test subsets. Models were evaluated using holdout testing, 10-fold cross-validation, McNemar statistical testing, calibration analysis, confidence intervals, ROC and PR curves, and Grad-CAM visualization. Experimental results show that EfficientNetB0 achieved the best overall performance, with 0.7900 Accuracy, 0.7898 Macro-F1, 0.8829 ROC-AUC, and 0.8461 PR-AUC on the clean holdout test set. Across ten-fold cross-validation, EfficientNetB0 further achieved 0.7801 ± 0.0093 Accuracy and 0.8780 ± 0.0090 ROC-AUC. It also demonstrated the strongest calibration performance, with the lowest Expected Calibration Error (ECE = 0.0558) and Brier Score (0.1449) among all compared models. The selected model was further implemented in a FastAPI-based backend system for web-based prediction. From a broader Informatics and Computer Science perspective, this study contributes a rigorous and reproducible evaluation framework that integrates statistical validation, calibration assessment, and interpretability, enabling more reliable model selection in visually subtle facial analysis tasks and supporting practical deployment in intelligent systems.

A Novel Hybrid CNN Model Integrating Resnet and Inception for Precision Classification of Coffee Beans

Rahmat Zulpani — 2026-06-15

Coffee is one of Indonesia’s key strategic commodities with substantial economic value for farmers and exporters. However, inconsistencies in post-harvest coffee bean quality remain a major challenge due to manual, subjective, and expertise-dependent classification. This study addresses this issue by developing an automated and objective computer vision–based classification system using a hybrid deep learning architecture. The proposed model, named RI-Net, integrates the residual learning capability of ResNet with the multi-scale feature extraction of the Inception module to improve the precision and robustness of coffee bean classification across four roasting levels: Green, Light, Medium, and Dark. The model was trained and evaluated on a locally collected dataset and benchmarked against three standard architectures—ResNet50, InceptionV3, and a Fully Convolutional Neural Network (FCNN). Experimental results show that RI-Net outperforms all baseline models, achieving perfect scores of 100% in accuracy, precision, recall, and F1-score. These findings confirm the effectiveness of combining residual and multi-scale features in capturing subtle visual differences across roasting levels. The study demonstrates the potential of advanced hybrid CNN architectures to enhance post-harvest quality control, supporting faster, more consistent, and standardized classification processes that strengthen the competitiveness of Indonesia’s coffee industry.

Evaluation of Image Transmission Strategies on Edge Server-Based Centralized Object Detection Systems

Firmansyah Achmad Adam — 2026-06-15

Urban waste management in smart city development requires efficient and stable visual monitoring systems. Utilization of edge devices such as Raspberry Pi is often constrained by limited computational power for complex computer vision models, making edge server architecture a relevant solution. This study evaluates the performance of image transmission from a Raspberry Pi to a centralized server for YOLOv8 object detection by comparing MJPEG streaming and HTTP POST-based periodic snapshot methods. Evaluation metrics included median latency (p50), jitter, and tail latency (p95 and p99). The results indicate that MJPEG streaming provides more stable latency compared to snapshots, particularly at tight transmission intervals. The transmission interval proved to have a significant effect on inference pipeline stability, while image resolution showed no observable impact on latency distribution under the evaluated conditions. This research recommends selecting appropriate transmission strategies to maintain the reliability of visual monitoring systems. These findings provide practical guidance for designing reliable centralized visual monitoring systems in resource-constrained edge environments.

Transformer-Based Multi-Class Intrusion Detection Using CICIoMT2024 Dataset for Secure IoMT Networks

Eko Arip Winanto — 2026-06-15

Internet of Medical Things (IoMT) ecosystems significantly enhance healthcare services but simultaneously expand the attack surface, exposing medical networks to diverse cyber threats such as distributed denial-of-service and spoofing attacks. Existing intrusion detection systems for IoMT are often limited to binary classification and struggle to capture complex multi-class attack behaviors, particularly under highly imbalanced data distributions. This study proposes a deep Transformer-based intrusion detection model as a reproducible baseline for multi-class intrusion detection in IoMT environments. The model is evaluated on the CICIoMT2024 dataset, which comprises 19 traffic classes including benign and multiple attack categories. Data preprocessing involves stratified data splitting, feature normalization, and label encoding to ensure fair evaluation. The proposed baseline employs a six-layer Transformer encoder with eight attention heads and is trained using the AdamW optimizer. Experimental results demonstrate an overall accuracy of 98.76% and a macro F1-score of 0.92, indicating strong detection capability across most attack classes. The model achieves excellent performance on benign traffic and high-volume attacks such as DDoS and DoS, while performance degradation is observed on minority classes, including ARP spoofing, highlighting the impact of class imbalance. These findings establish the proposed Transformer model as a transparent and robust baseline for IoMT intrusion detection research. By providing reproducible performance benchmarks, this work supports future development of hybrid and imbalance-aware detection mechanisms aimed at enhancing real-time security in medical cyber-physical systems.

Impact of Contrast Limited Adaptive Histogram Equalization and Image Upscaling on Cataract Classification Using Deep Learning Models: Inception-ResNetV2, EfficientNetB0, and ResNet-50

Ismi Dwi Junianti — 2026-06-15

Cataract is one of the leading causes of visual impairment worldwide, and its detection using retinal images remains a critical challenge in medical image analysis due to variations in image quality and subjectivity in clinical assessment. This study aims to evaluate the impact of image preprocessing techniques, namely Contrast Limited Adaptive Histogram Equalization (CLAHE) and image upscaling, on the performance and interpretability of deep learning–based cataract classification models. Three convolutional neural network architectures—Inception-ResNetV2, EfficientNetB0, and ResNet-50—were assessed using a balanced dataset of 2,000 retinal images under two experimental settings: raw images and enhanced images. The models were evaluated using accuracy, precision, recall, and F1-score, while Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to analyze model interpretability. Experimental results show that EfficientNetB0 achieved the highest accuracy on raw images (96%), followed by ResNet-50 (94%) and Inception-ResNetV2 (92%). After applying CLAHE and upscaling, ResNet-50 exhibited improved performance, reaching 95% accuracy, whereas EfficientNetB0 and InceptionResNetV2 experienced a decrease in accuracy to 83%. Grad-CAM visualizations indicate that all models consistently focused on clinically relevant regions associated with cataract characteristics. These findings demonstrate that image enhancement techniques do not universally improve classification performance and that their effectiveness is highly dependent on the underlying CNN architecture. The study provides practical insights for selecting appropriate preprocessing–model combinations to develop accurate, interpretable, and robust deep learning–based cataract classification systems for medical decision-support applications.

Explainable Artificial Intelligence Using SHAP and Multilayer Perceptron for Transparent Stunting Risk Prediction in Sukoharjo, Indonesia

Nimas Ratna Sari — 2026-06-15

Childhood stunting remains a critical public health challenge in Indonesia, with national prevalence at 19.8% in 2024 per SSGI data, hindering human capital development toward Indonesia Emas 2045. This study addresses the opacity of AI models in stunting prediction by integrating machine learning with Explainable AI (XAI) to enhance transparency for non-technical stakeholders. Using a survey dataset of 273 children from Sukoharjo Regency, risk factors encompassing key stunting determinants consist of maternal characteristics, household socioeconomic conditions, sanitation practices, and sociodemographic, were preprocessed via cleaning, label encoding, min-max scaling, and train-test split. Three classifiers; Logistic Regression (LR), Naïve Bayes (NB), and a Multilayer Perceptron (MLP) with ReLU/softmax were trained and evaluated on accuracy, precision, recall, and F1–score. MLP with 16 hidden nodes, achieved the highest performance: 82% accuracy, 87% precision, 82% recall, and 82% F1-score, outperforming baselines. Kernel SHAP was applied to decompose predictions, revealing mother's education, age, number of children, birth length, household size, and income as top influencers. This XAI enhanced framework promotes trust and actionability in public health interventions, advancing informatics by bridging high accuracy neural networks models with interpretable insights for targeted stunting reduction in resource–limited settings.

Development of a Hybrid Machine Learning-Based E-Commerce Chatbot Using Jaccard Similarity and K-Nearest Neighbor for Accurate Intent Classification

Andrian Sah — 2026-06-15

The advancement of technology in the e-commerce industry requires fast and accurate information services, particularly through the use of Natural Language Processing (NLP)-based chatbots. However, many existing chatbots rely on a single method, which often limits their ability to understand user question contexts effectively. This study proposes a hybrid approach integrating Jaccard Similarity and K-Nearest Neighbor (K-NN) to improve answer retrieval accuracy and intent classification in e-commerce chatbot systems. Jaccard Similarity is employed to measure the similarity between user queries and Frequently Asked Questions (FAQ) data, while K-NN is used to determine intent based on the nearest neighbor with the highest similarity values. The dataset, consisting of FAQ questions and answers, is preprocessed through case folding, tokenization, stopword removal, and stemming. System performance is evaluated using accuracy, precision, recall, and F1-score metrics. The experimental results show that Jaccard Similarity effectively selects relevant answer candidates, achieving similarity values of up to 66%, while K-NN produces stable intent classification results. The proposed hybrid model achieved an accuracy of 87%, precision of 86%, recall of 85%, and an F1-score of 85%, outperforming single-method implementations. Furthermore, confidence score analysis indicates that most chatbot responses fall into the high confidence category (>0.70). Rule-based NLP evaluation also provides insights into unclassified inputs, which can be used as a basis for future dataset development. The implementation results demonstrate that the chatbot system can be operated effectively on both customer and admin sides and monitored through analytical features. Overall, the proposed hybrid approach enhances the reliability, relevance, and stability of chatbot responses, making it a practical and effective solution for real-time intent classification and FAQ retrieval in e-commerce customer service environments.

Certainty Factor Algorithm Approach for Early Stage of Cattle Disease Diagnosis Using Mobile-Based Expert System

Budy Satria — 2026-06-15

Cattle are one of the livestock that play a crucial role in meeting the demand for meat and milk, as well as providing a source of income for farmers, particularly in various regions of Indonesia. Diseases in cattle pose a serious problem due to the lack of knowledge about accessing veterinary services, a lack of understanding among farmers, and the high cost and time required for consultations, which are significant obstacles for farmers in identifying diseases in cattle early, potentially leading to death. Limitations in accessing veterinary services, a lack of understanding among farmers, and the high cost and time required for consultations are significant obstacles to treating diseases in cattle. This study aims to assist farmers in diagnosing cattle diseases using an expert system based on the observed symptoms. The application of the expert system employs a certainty factor algorithm approach, utilizing the knowledge base of animal experts in the diagnosis process. This study used 6 types of diseases and 34 lists of symptoms in cattle. Based on the results of implementing the Certainty Factor method, it was concluded that the expert system was able to diagnose cattle diseases, specifically worms, with a confidence level of 90.1504%. This is certainly influenced by the selection of symptoms, the user's confidence value for each symptom, and the combination of the confidence values from experts. In addition, testing was also carried out on the functionality of the expert system built; the results obtained showed that all functionalities run well and as expected. Thus, the final conclusion is that expert systems can be a solution and help farmers diagnose cattle diseases. Suggestions for further research include comparing algorithms to achieve better accuracy and disease identification in specific cattle species.

Performance Comparison Of K-Nearest Neighbors And Decision Tree Algorithms With Random Oversampling For Imbalanced Heart Disease Classification

Dita Yustianisa — 2026-06-15

Heart disease remains one of the leading causes of mortality worldwide, including in Indonesia, where delayed detection continues to be a serious challenge. In medical data mining, class imbalance often degrades classification performance by reducing sensitivity toward minority class cases. This study aims to compare the performance of the K-Nearest Neighbors (KNN) and Decision Tree algorithms for heart disease classification and to evaluate the effectiveness of random oversampling in handling imbalanced data. This research uses a heart disease dataset consisting of 10,000 medical records obtained from Kaggle. Data preprocessing includes categorical transformation, missing value imputation using KNN Imputer, and Min–Max normalization. Random oversampling is applied to increase minority class representation. Model evaluation is performed using stratified 10-fold cross-validation with accuracy, precision, recall, F1-score, and Receiver Operating Characteristic–Area Under the Curve (ROC–AUC) as performance metrics. Experimental results show that after random oversampling, the KNN model achieves the best performance with an accuracy of 94%, precision of 96%, recall of 90%, F1-score of 92%, and ROC–AUC of 90.2%. In comparison, the Decision Tree model records an accuracy of 80%, precision of 78%, recall of 81%, F1-score of 79%, and ROC–AUC of 81.5%. These findings demonstrate that random oversampling significantly improves minority class detection, particularly for KNN. This study contributes to Informatics by providing empirical evidence that simple and efficient data mining strategies can effectively address class imbalance in large-scale medical datasets, supporting the development of accurate, interpretable, and accessible AI-based diagnostic systems for early heart disease detection.

Detection of Coffee Leaf Diseases Using Deep Learning to Support Digitalization and Smart Agriculture

Siti Mutmainah — 2026-06-15

Coffee is one of Indonesia's main commodities and an important agricultural sector for the economy. However, one of the challenges in coffee cultivation is disease. Improper and delayed treatment can cause leaf damage and death of coffee plants. This study aims to detect disease types on coffee leaves using a deep learning CNN approach and lightweight CNN architectures such as MobileNet and EfficientNet variants. This study also applies traditional image augmentation and OpenCV. The results of EfficientNetV2S and EfficientNetV2L achieve 95–98% accuracy with stable precision and recall in almost all classes, although minority classes remain challenging. The MobileNetV3 architecture showed optimal results with 99% accuracy in all variants (Small, Large, and Small with OpenCV augmentation). The research model was verified using local coffee leaf images Bumi Pajo. These findings confirm that MobileNetV3 not only excels in terms of accuracy but also has the potential to be applied to mobile device-based or Internet of Things (IoT) coffee leaf disease monitoring systems. With high accuracy and low computational requirements, this model can support real-time disease detection in the field, helping farmers and agricultural practitioners make quick and accurate decisions in disease control.

Historical Image Restoration Using GFPGAN-Based Face-Centered Enhancement Mechanism to Address Blur and Low-Light Degradation

Ardi Wijaya — 2026-06-15

Archaic image restoration faces significant challenges due to complex degradation in the form of blurring and attenuation of extreme luminance (low-light) that obscure the identity of historical subjects. This study constructs a new paradigm through the Face-Centered Enhancement mechanism based on GFPGAN to reconstruct high-fidelity facial features in visual archives from the Bengkulu Museum, Bung Karno's Exile House, and Fort Marlborough. The novelty of this study lies in the integration of a feature enhancement module capable of performing adaptive deconvolution specifically on the face area to mitigate stochastic hallucinations in the GAN latent space, thus balancing lighting restoration without distorting the authenticity of the original character of historical figures. Quantitative evaluation of 50 images using a synthetic degradation scheme shows superior performance, where 95% of the data achieves SSIM ≥ 0.95 and MSE ≤ 0.01. This improvement in visual quality has direct implications for the efficiency of the OCR system in extracting document text and optimizing classification in digital archival information systems. Despite its dependence on high-performance computing, this method has proven effective in bridging the disparity between improving pixel quality and preserving historical integrity in national digital preservation efforts.

Bandwidth Prediction in Zoom Meetings: A Mathematical Model Based on Feature Configuration Analysis

Andi Cahyono — 2026-06-15

Video conferencing applications such as Zoom Meeting require sufficient and stable bandwidth to maintain communication quality. However, bandwidth needs often vary depending on user configurations, including video resolution, audio bitrate, and content-sharing activity. This study aims to develop a mathematical formula capable of accurately estimating bandwidth requirements for Zoom Meeting sessions. The methodology combines quantitative experiments and numerical simulations by collecting throughput data using Wireshark, analysing feature-based parameter variations, and validating the proposed formula through MATLAB simulation. Data were obtained from multiple Zoom sessions executed under controlled conditions with different feature combinations and replicated twenty times to ensure accuracy. The validation results show that the formula consistently provides realistic and stable estimations when compared with actual throughput measurements and simulation outcomes. The proposed model offers a simple yet effective tool for predicting bandwidth requirements, supporting efficient network capacity planning, and enhancing the overall performance of video conferencing environments.

Optimization Performance of Extreme Gradient Boosting and Random Forest for Child Stunting Classification Based on Economic Factors

Yuyun Yusnida Lase — 2026-06-15

Stunting remains a major health concern in Indonesia due to its impact on children’s physical growth and cognitive development. One of the factors influencing the incidence of stunting is family economic status, which is linked to access to nutrition, sanitation, and a healthy environment. This study aims to optimize the performance of the XGBoost and Random Forest algorithms in classifying stunting in children based on economic factors and to compare the performance of the two models. The methods used in this study involve a machine learning approach, including data preprocessing, model training, hyperparameter optimization, and performance evaluation using a confusion matrix, accuracy, precision, recall, F1-score, and ROC-AUC curves. The results indicate that both algorithms perform well in classification, with an accuracy rate of approximately 70%. The Random Forest model demonstrated better performance than XGBoost with an AUC value of 0.7655, while XGBoost had an AUC value of 0.75. Additionally, the feature importance results indicated that economic and environmental factors, such as housing conditions and sanitation, have a significant influence on the incidence of stunting.

CCTV-Based River Waste Detection Using a Hybrid CNN–Graph Attention Network with Spatial–Contextual Feature Learning

Asep Surahmat — 2026-06-15

River waste accumulation has become a serious environmental problem in urban areas, particularly in highly polluted rivers such as the Angke River in Tangerang, where floating waste disrupts ecological balance and increases flood risk. Conventional computer vision–based detection methods often fail under dynamic river conditions due to water surface reflections, turbulence, occlusion, and visually ambiguous debris. This study aims to improve the accuracy and robustness of river waste detection by proposing a hybrid deep learning framework that integrates convolutional and graph-based spatial–contextual reasoning. The proposed method utilizes a ResNet50 backbone for feature extraction from CCTV imagery, followed by spatial graph construction that models adjacency relationships between image regions. A Graph Attention Network (GAT) is then applied to capture contextual dependencies and refine feature representations prior to classification. Unlike conventional CNN-only or YOLO-based detectors that rely primarily on local visual cues and bounding-box representations, the proposed approach explicitly models spatial–contextual relationships between image regions through graph-based attention mechanisms. Experiments were conducted on 4,200 CCTV image frames collected from the Angke River under varying environmental conditions. The proposed model achieved an accuracy of 92.4%, precision of 91.1%, recall of 93.2%, F1-score of 91.9%, and a mean Average Precision (mAP) of 0.78, outperforming CNN-only and YOLO-based baseline models. These findings highlight the contribution of graph-enhanced visual reasoning to the fields of Computer Vision and Intelligent Surveillance, particularly for real-time environmental monitoring systems operating in complex and dynamic visual environments.

Implementation of Moving Average and Weighted Moving Average for Forecasting Palm Oil Harvest and Income in a Web-Based GIS System

Elvia Andriyani — 2026-06-15

Independent palm oil farmers face significant challenges in financial management due to inefficient manual recording, fluctuating harvest yields, and volatile Fresh Fruit Bunch (FFB) prices. This study aims to develop a web-based harvest and income recording system integrated with a Geographic Information System (GIS) and forecasting methods to support decision-making. The system is developed using a Research and Development (R&D) approach by comparing Moving Average and a dynamically weighted Moving Average that adapts to price fluctuations for predicting future net income. Model performance is evaluated using Mean Absolute Percentage Error (MAPE) and validated with the Diebold–Mariano test, while system usability is assessed through User Acceptance Testing (UAT). The results show that the dynamically weighted Moving Average achieves a prediction accuracy of 93.08% (MAPE 6.92%), slightly outperforming the standard Moving Average (93.03%), although no statistically significant difference is found based on the Diebold–Mariano test. The system also obtains a “Very Good” usability rating with a UAT score of 95.11%. These findings demonstrate that the proposed approach provides a practical and adaptive forecasting mechanism integrated within a spatial financial management system, contributing to improved decision support and offering methodological value in time-series forecasting for agricultural informatics.

Hybrid Unsupervised-Supervised Learning for Housing Submarket Segmentation and Price Prediction in Surabaya Urban Areas

Rinabi Tanamal — 2026-06-15

Surabaya’s rapid population growth, reaching 3.02 million residents, has intensified housing affordability challenges and increased structural variability in residential markets. This study proposes a hybrid machine learning framework that combines unsupervised clustering with supervised classification to identify submarket segments and predict housing price categories. A dataset of 490 properties containing structural, land, ownership, and contextual features was preprocessed and analyzed using K-Means. Cluster quality assessment through elbow inspection and a silhouette score of 0.45 indicated the presence of five meaningful market segments. These segments served as targets for a supervised classification stage that evaluated seven models, optimized via randomized hyperparameter search within a standardized preprocessing pipeline.

The RBF-SVM achieved the strongest performance, reaching 97 percent accuracy and a macro-F1 score of 0.97, representing an 8 percent improvement over non-hybrid baselines and outperforming boosted ensembles such as XGBoost. Permutation importance analysis identified number of floors, building orientation, position rank, and ownership status as dominant drivers of segment differentiation. The integration of clustering and classification enhances predictive reliability while improving interpretability, offering a transparent analytical toolkit for housing market assessment.

The proposed framework provides actionable insights for developers, appraisers, and policymakers in Surabaya, enabling data-driven identification of submarkets and supporting more equitable housing strategies aligned with SDG 11 on sustainable urban development. The approach is scalable to other Indonesian cities and establishes a foundation for future work incorporating spatial, socioeconomic, or temporal predictors.

Prediction of Indonesian Banking Stock Prices Using a Hybrid LSTM and XGBoost Model with Optuna Based Hyperparameter Optimization

Admaja Admaja — 2026-06-15

Stock price prediction is a critical task in investment decision-making, particularly in highly volatile financial markets such as the Indonesian banking sector. While Long Short-Term Memory (LSTM) networks are effective in modeling temporal dependencies, they often fail to capture nonlinear residual patterns in financial time-series data, and their performance is highly sensitive to hyperparameter selection. To address these limitations, this study proposes a residual learning–based hybrid LSTM–XGBoost framework optimized using Optuna for predicting stock prices of major Indonesian banking stocks, namely BBCA, BBNI, BBRI, and BMRI. LSTM is employed as the base learner to model log-return sequences, while XGBoost is used to learn nonlinear residual structures from LSTM predictions. Latent embeddings extracted from the LSTM are further refined using Principal Component Analysis (PCA) to reduce redundancy and improve generalization. Hyperparameters of the LSTM, PCA, XGBoost, and calibration components are jointly optimized using Optuna with a Tree-structured Parzen Estimator (TPE) strategy. Experimental results demonstrate that the Optuna-optimized hybrid model consistently outperforms the baseline hybrid model across all datasets, achieving lower Mean Absolute Percentage Error (MAPE) values of 1.196% for BBCA, 1.67% for BBNI, 1.53% for BBRI, and 1.70% for BMRI. Additional stability analyses confirm that the proposed framework delivers stable and reliable predictions on unseen data. These findings provide a scalable hybrid forecasting framework that contributes to the development of intelligent financial decision-support systems and demonstrates the effectiveness of adaptive hybrid deep learning optimization techniques in real-world time-series prediction problems within the field of informatics.

Development of Smart Study Web Application for Classifying Student Material Understanding Levels Using Naive Bayes Classifier

Purnomo Hadi Susilo — 2026-06-15

The rapid development of information and communication technology requires adaptive digital learning systems that are able to evaluate students’ learning outcomes objectively. However, the Smart Study application previously functioned only as a quiz delivery platform and lacked analytical capabilities to assess students’ levels of material understanding, particularly in practical courses such as Computer Networks. This study aims to design and develop a web-based Smart Study application integrated with the Naive Bayes classification algorithm to determine students’ understanding levels based on quiz performance data. The research methodology includes data collection from Informatics Engineering students at Universitas Islam Lamongan, followed by data preprocessing through cleaning and categorical conversion of features, including final score, average response time, response time variability, and correct incorrect response time ratio. The dataset was divided into 80% training data and 20% testing data. The Naive Bayes model was trained and evaluated using accuracy, precision, recall, F1-score, and a confusion matrix. The results show that the proposed model achieved an accuracy of 75%, correctly classifying 15 out of 20 testing samples. The model demonstrated strong performance in identifying the Comprehended class with an F1-score of 0.83, while performance for the Not Comprehended class was lower with an F1-score of 0.55 due to class imbalance. This study contributes to the fields of learning analytics and educational data mining by demonstrating the integration of a simple machine learning method into an e-learning application to support early detection of learning difficulties and data-driven evaluation of digital learning processes in higher education.

Improving RoBERTa Performance through Hyperparameter Optimization for Sentiment Analysis of Indonesian Tourism Reviews

Imamah Imamah — 2026-06-15

The performance of transformer models such as RoBERTa in sentiment classification is influenced by hyperparameter settings, especially the epoch and batch sizes. However, no previous study has examined the impact of changes in the number of epochs and batch sizes on the performance of each class in classification tasks, especially in Indonesian-language sentiment analysis of tourism reviews. Therefore, this study aims to fill this gap by analyzing the performance of RoBERTa and the impact of various hyperparameter settings on sentiment for each class. The dataset consists of 3,875 reviews from visitors to Lake Sarangan on Google Maps. The batch sizes used in this study are 8 and 16, and the epoch range is 2 to 4. There are three classes of sentiment: negative, neutral, and positive. The results demonstrate that increasing the batch size from 8 to 16 does not linearly improve model performance. The optimal combination of epoch=4 and batch size=8 achieved 91% accuracy, with significant improvements in recall and F1-score across all classes, especially in positive sentiment classification. This research offers valuable insights into fine-tuning RoBERTa for sentiment analysis in Indonesian contexts, providing recommendations for future sentiment analysis tasks in natural language processing.

Comparative Analysis of the Performance of Random Forest and CatBoost for Air Quality Prediction Based on Meteorological Factor

Nirsal Nirsal — 2026-06-15

Air quality in urban centers such as Tangerang City has become an increasingly urgent issue due to the expansion of industrial activities, rapid population growth, and rising vehicle emissions. As a key city within the Greater Jakarta metropolitan area, Tangerang is highly vulnerable to air pollution caused by human activities and varying meteorological conditions. This study aims to assess the performance of two machine learning algorithms, Random Forest and CatBoost, in predicting air quality in Tangerang under two scenarios: models that incorporate meteorological factors and models that exclude them. The dataset includes concentrations of key air pollutants alongside meteorological variables such as temperature, humidity, and wind speed. Model performance was evaluated using MAE, MSE, RMSE, and R². The findings indicate that both algorithms perform excellently when meteorological variables are included. Random Forest achieved an MAE of 0.0099, MSE of 0.000309, RMSE of 0.0152, and an R² of 0.9931, slightly outperforming CatBoost, which recorded an MAE of 0.0135, MSE of 0.000419, RMSE of 0.0170, and an R² of 0.9907. Excluding meteorological variables decreased accuracy for both models, with Random Forest reaching an R² of 0.9519 and CatBoost 0.9487. These results underscore the importance of temperature, humidity, and wind speed in enhancing predictive accuracy. Notably, this study introduces a comparative evaluation of machine learning models in a unique urban context, providing new insights into how meteorological factors influence air quality predictions. The study contributes to the development of adaptive air quality prediction models, supporting sustainable environmental management planning in Tangerang City.

Banana Leaf Disease Classification Using CNN Feature Extraction and Naive Bayes Algorithm

Moh. Badri Tamam — 2026-06-15

Banana leaf diseases such as Black Sigatoka, Cordana, and Pestalotiopsis significantly reduce productivity and require early, accurate detection to prevent severe yield losses. While Convolutional Neural Networks (CNN) have demonstrated high performance in plant disease classification, most existing approaches rely on computationally intensive end-to-end deep learning models, limiting their deployment on resource-constrained devices. This study proposes a lightweight hybrid classification framework that integrates MobileNetV2-based CNN feature extraction with a Gaussian Naive Bayes classifier. The novelty of this research lies in the systematic transformation of deep 1,280-dimensional feature representations into a probabilistic classification space, enabling competitive accuracy with substantially lower computational complexity. A balanced dataset consisting of 3,200 training images and 1,311 testing images collected from Pamekasan Regency was preprocessed through resizing, normalization, and augmentation. Experimental results show that the end-to-end CNN achieved 98.70% accuracy, while the proposed hybrid CNN–Naive Bayes model attained 95.73% accuracy with F1-scores above 0.90 across all classes. Despite not relying on backpropagation during classification, the hybrid approach maintains strong predictive performance while reducing training time and memory requirements. These findings demonstrate that integrating deep feature extraction with probabilistic learning provides an efficient and deployable solution for edge-based precision agriculture systems.

Image Cryptography Process Using Arnold’s Cat Map And Henon Map Algorithms

Moch Azhar Al Ghifari — 2026-06-15

The security of digital image data is a crucial aspect in various fields, such as communications, medicine, and the military. The inherent characteristics of digital images—namely high pixel correlation and large data size—render conventional encryption methods less optimal. This study aims to evaluate the encryption quality of images using the Arnold’s Cat Map (ACM) and Henon Map algorithms, both individually and in combination (ACM-Henon and Henon-ACM). ACM is utilized to rearrange pixel positions to create a confusion effect, while the Henon Map is employed to randomly alter pixel values (diffusion). The implementation is carried out using the Python programming language within the Visual Studio Code development environment. Encryption quality is assessed using parameters such as Avalanche Effect (AE), Unified Average Changing Intensity (UACI), Number of Pixels Change Rate (NPCR), and correlation coefficient. Experimental results show that the combined chaos-based methods significantly enhance security compared to the individual algorithms, particularly by analyzing the impact of algorithm order on encryption quality. The best performance was achieved by the Henon→ACM combination, producing NPCR ≈ 99.44%, UACI ≈ 19.93%, entropy ≈ 7.9874, and AE ≈ 50.12%, indicating strong randomness and resistance to differential attacks.

This research demonstrates that combining confusion and diffusion mechanisms yields more secure cipher images than using either method alone. The main contribution of this study lies in providing a systematic comparative evaluation of single and combined chaos-based encryption schemes, including order-sensitive analysis across different image characteristics, rather than proposing a new encryption algorithm. However, the encryption performance is influenced by image size, parameter selection, and iteration count, which may limit consistency across different image characteristics. Future work may explore adaptive parameter optimization and improved diffusion mechanisms for higher UACI values.

Global Inflation Forecasting Using Stacking Ensemble with Elastic Net Meta-Learner Integrating Random Forest, XGBoost, and LightGBM

Fauriza Wildhani — 2026-06-15

Inflation dynamics have become increasingly complex due to economic volatility and nonlinear interactions, challenging the reliability of conventional forecasting models; therefore, this study develops a robust global inflation forecasting framework using a hybrid stacking ensemble that integrates Random Forest, XGBoost, and LightGBM as base learners with Elastic Net as a regularized meta-learner, applied to annual inflation data from 2000–2024 across five major economic blocs (G7, Europe, BRICS, ASEAN, and the Americas) after temporal feature engineering and time-series–preserving validation; the results demonstrate strong and consistent predictive performance, with very high accuracy in Europe (R² = 0.9282) and the G7 (R² = 0.9122), and the globally trained stacking model (R² = 0.7866) substantially outperforming the region-specific ASEAN model (R² = 0.5243), confirming the advantage of cross-country learning; this research advances informatics and computer science by providing a scalable and stable ensemble learning framework for macroeconomic time-series forecasting in volatile environments, supporting the development of AI-driven economic and policy analytics systems.

Classification of Eyewitness Social Media Messages for Natural Disaster Monitoring using BERT Variants

Muhammad Bashir Hanafi — 2026-06-15

The rapid growth of disaster-related social media data demands effective monitoring. However, its real-time source presents challenges due to large volumes of unstructured and noisy data. This study aims to improve effective monitoring with BERT variants to classify eyewitness reports on Twitter/X. Earlier studies have applied machine-learning and deep-learning models to automate the monitoring of eyewitness messages on social media, but these models still have shortcomings. Traditional machine-learning models rely on handcrafted and frequency-based features, limiting their ability to capture contextual semantics. Deep-learning models offer improved performance but still face challenges in modeling long-range dependencies and handling high-volume social media streams. This issue is pronounced in social media streams. This study employs transformer-based models using several BERT variants (BERT, RoBERTa, DistilBERT, ELECTRA, and ALBERT). Each model is pre-trained with the Masked Language Modeling (MLM) objective, and batch-size optimization is applied to boost performance. Experimental results indicate that a batch size of 16 consistently yields the best performance, with the standard BERT model achieving the highest macro-F1 score of 0.762. By disaster type, macro-F1 scores reach 0.744 for hurricane, 0.793 for flood, 0.756 for earthquake, and 0.750 for wildfire. BERT (16) outperforms the other BERT variants and twelve baseline models from prior research. Unlike previous approaches, this study leverages pre-trained Masked Language Models to optimize classification on disaster-related datasets. The findings contribute to the development of transformer-based architectures for text classification in real-time disaster informatics, leading to more accurate situational awareness and reduced delays in emergency decision-making.

Preference-Driven Medical Image Retrieval using a Dual-Head DenseNet-121 and Multi-Objective Skyline Query for COVID-19 Detection

Slamet Handoko Handoko — 2026-06-15

This study addresses the limitation of single-objective content-based image retrieval in medical imaging, which fails to consider multiple clinical preferences such as image quality. The objective is to develop a preference-driven retrieval system for COVID-19 chest radiography images. A hybrid approach is proposed by integrating a Dual-Head DenseNet-121 model for feature extraction and quality regression with a multi-objective skyline query algorithm for retrieval optimization. The system evaluates multiple image quality dimensions, including sharpness, contrast, exposure, signal-to-noise ratio, and entropy. Experimental results demonstrate that the proposed method achieves 100% Pareto efficiency and improves diversity and hypervolume coverage compared to conventional methods. This approach provides a more flexible and effective multi-objective retrieval mechanism, contributing to the advancement of intelligent medical image retrieval systems in computer science.

Oil Palm Stem Disease Detection Based on Color Moments and GLCM Texture Features Using Artificial Neural Networks

Hamdani Hamdani — 2026-06-15

Oil palm is an essential commodity for the economy; however, basal stem rot caused by Ganoderma boninense poses a significant threat to plantation productivity and long-term vitality. It highlights the importance of early detection of stem disease to facilitate timely intervention and minimize potential economic losses. This study presents an image-based approach to diagnosing oil palm stem maladies, leveraging handcrafted color and texture features within a supervised machine learning framework. The dataset contained 525 images of oil palm stems, of which 205 depicted healthy specimens, and 320 depicted diseased ones. These were captured within their natural environment. Color features were derived by analyzing color moments within the HSV color space, while texture features were extracted from the Grey-Level Co-occurrence Matrix (GLCM). The extracted features were classified employing an Artificial Neural Network (ANN) and were subsequently contrasted with classifiers including Decision Tree, K-Nearest Neighbors, Naive Bayes, and Support Vector Machine. Model performance was evaluated using k-fold cross-validation with k = 5 and k = 10 to ensure the consistency and reliability of the assessment. The experimental results demonstrated that the highest accuracy of 97.52% was achieved when the ANN model was used to classify the integrated color and texture features. The innovative aspect of this research resides in demonstrating that handcrafted features integrated with artificial neural networks can attain high detection accuracy in scenarios with limited data, providing a viable alternative to data-intensive deep learning techniques. This method facilitates a dependable, computer vision-driven early detection system for oil palm stem diseases, thereby promoting sustainable plantation management.

Reliable Intent Detection in Public Service Chatbots Using Hybrid IndoBERT and Bidirectional Long Short-Term Memory with Confidence-Based Decision Strategy

Barka Satya — 2026-06-15

The rapid digitalization of public services has increased the demand for intelligent information systems capable of providing accurate and responsive assistance to citizens on a 24/7 basis. However, many existing public service chatbots still rely on rule-based mechanisms or single-model natural language processing (NLP) approaches, which often fail to handle linguistic variations, informal expressions, and ambiguous user queries. This study proposes a Hybrid Natural Language Understanding (NLU) architecture that integrates a fine-tuned IndoBERT model with a Bidirectional Long Short-Term Memory (BiLSTM) network to improve intent detection performance in public service chatbots. To enhance system reliability, a confidence-based decision-making mechanism is introduced, enabling the system to dynamically select the most reliable prediction or activate a fallback pattern-matching module when confidence thresholds are not met. The proposed approach was evaluated on a custom dataset comprising 53 public service intents, spanning formal and informal Indonesian language use. Experimental results demonstrate that the hybrid architecture achieves an intent classification accuracy of 86.8%, outperforming single-model approaches while maintaining an acceptable response time for practical deployment, particularly in public service scenarios where accuracy and reliability are prioritized over response speed. Furthermore, integrating a continuous learning mechanism enables the system to adapt to low-confidence queries over time, thereby improving robustness in real-world applications. These findings indicate that hybrid NLP architectures with confidence-aware decision mechanisms offer a practical and scalable solution for intelligent public service chatbots.

Systematic Review of Adaptive User Interfaces in E-Commerce for MSMEs: Gaps and User-Centric Indicators

Solehatin Solehatin — 2026-06-15

Objective – Observations of research results related to adaptive user interfaces in e-commerce have been widely conducted; however, there is a need for evaluation and assessment of indicators based on user requirements. This study aims to conduct a systematic literature review and bibliometric analysis on adaptive user interfaces for MSMEs, based on existing empirical research.

Methodology – The methodology applied is a Systematic Literature Review, using the term “adaptive user interface for MSMEs” in “Article Titles, Abstracts, and Keywords” within the Sciencedirect database, resulting in 5,622 publications from 1998 to 2025. The evaluation was carried out on November 21, 2025. The collected articles were analyzed using bibliometric analysis with VOSviewer software, based on fields of study including computer science, decision science, engineering, social sciences, business, management, accounting, and materials science.

Findings – Research on adaptive user interfaces for MSMEs has been extensively conducted in line with the digitalization of the e-commerce sector. The observations sought gaps and indicators in each article. Gaps were identified; however, further research is still needed to provide more specific, comprehensive, and well-founded recommendations. Indicators focus on how to provide ease and comfort for users, as well as offering recommendations to them.

Research Limitations – This study used the Sciencedirect database for articles related to adaptive user interfaces for MSMEs. Future research could enhance generalizability by integrating other databases such as the Web of Science.

Regional Segmentation of School Dropouts Based on Economic and Accessibility Factors Using K-Means Clustering

Juna Eska — 2026-06-15

The high dropout rate in Asahan Regency has become a serious problem affecting the quality of human resources and equitable access to education across various regions. This study aims to identify patterns and characteristics of dropout-prone areas using the K-Means clustering technique. The research method involves collecting dropout data from the Asahan Regency Education Office for the period 2022–2025, followed by data pre-processing for cleaning and normalization, and then clustering analysis to generate three regional clusters based on dropout vulnerability levels. The results indicate that clusters with high dropout rates are largely influenced by economic factors, followed by limited access to education and social conditions in the community. The resulting regional segmentation provides a spatial overview of dropout vulnerability levels in Asahan Regency. These findings offer data-driven insights that can support the formulation of more targeted education policies and programs to encourage inclusive education development in the region. Scientifically, this study contributes to strengthening the validity and effectiveness of the K-Means algorithm as a quantitative approach in mapping and identifying complex patterns in socio-educational data, thereby expanding its application in data-driven analytical studies in the field of education.

Implementation Of Cnn Mobile Netv2 For Classification And Detection Of Diseases In Banana Plants Through Leaf Images

Maria Yunita — 2026-06-15

Banana plants are vulnerable to disease attacks, especially in remote areas with limited access. Banana farmers struggle to identify and classify types of diseases on banana plants early on due to limited information about the types of diseases and the characteristics of diseases that attack bananas.The purpose of this study is he development of a CNN model with a MobileNet architecture for the classification and detection of diseases through banana leaf images, which can be implemented in an Android application. The method used applies a Convolutional Neural Network (CNN) using the MobileNetV2 architecture that can help classify banana plant diseases. The banana leaf image dataset was obtained independently and additionally from the Kaggle platform up to 4135 images. The images were then divided into 6 classes consisting of healthy leaves, panama disease, moko disease, leaf pests, yellow sigatoka and black sigatoka. The image dataset was then divided again into 3 parts: training data, validation data and test data with a data division of 80:10:10. The results showed that CNN with MobileNetV2 architecture can be used for disease classification and detection with an accuracy rate of 87.26% for the test data, 89,59 for validasi and 92.71% for the training data. This model was successfully implemented on the Android platform using Android Studio to detect banana plant diseases in real time without special tools.

Face Gender Classification for Public Facility Access Control using EfficientNet with Penalized-Entropy Loss

Sabrina Adinda Sari — 2026-06-15

Access to public facilities that are restricted based on gender, such as toilets and changing rooms, requires a strict security system because there are still many cases of abuse by irresponsible parties if only gender signs are relied upon. CCTV integrated with facial recognition is becoming more sophisticated every day, but it is limited if the face is covered by attributes such as masks. This is because the less visible the area is, the more difficult it is for the model to determine the label. To overcome this, this study proposes a gender classification approach for faces that may be covered by accessories such as masks, by adding Penalized Entropy loss as a loss function to the EfficientNet-B0 model. This loss function adds a penalty for incorrect predictions even if they are fairly accurate. The evaluation results show that the proposed model, with a penalty weight of 0.5, improved the accuracy by 3% from 90% to 93%. The experimental results show that the determination of the penalty weight has a significant impact on model performance, where a weight of 0.5 produces optimal performance because it provides a balance between penalizing overconfident predictions and the model's ability to maintain relevant feature discrimination; too small a weight does not sufficiently suppress overconfidence, while too large a weight actually reduces classification ability. The proposed method has demonstrated improvements in generalization and reduced overconfidence in gender classification systems. This method contributes to the development of reliable biometric systems suitable for uncontrolled real-world environments.

Classification of Watermelon Flavor Using Artificial Neural Network with Color, Texture, and Shape Features

Muh Faqih S Musgamy — 2026-06-15

Watermelon (Citrullus vulgaris Schard) is a widely produced fruit due to its high nutritional value and health benefits. However, consumers often experience difficulty in distinguishing sweet and bland watermelons because quality assessment is generally conducted manually and subjectively. To address this issue, this study proposes a watermelon flavor classification system based on visual features, including color, texture, and shape, using an Artificial Neural Network approach with digital image processing. The dataset used in this study consists of 214 images collected from 55 watermelon samples, categorized into sweet and bland classes. The proposed method involves several stages, namely image acquisition, preprocessing, grayscale conversion, segmentation, morphological operations, feature extraction, and classification using a feedforward backpropagation learning algorithm. Various combinations of visual features were evaluated to determine the most effective configuration. Experimental results show that the proposed system achieves an accuracy of 93.67% on training data and 92.85% on testing data, with an average computation time of 0.319 seconds per image. The findings indicate that the integration of Hue Saturation Value color features, texture features derived from the Gray-Level Co-occurrence Matrix, and shape features significantly enhances the accuracy of watermelon flavor classification. This study contributes to the development of an objective, efficient, and non-destructive fruit quality assessment system and demonstrates potential applicability to other types of fruits using a similar approach.

Sentiment Analysis Of E-Commerce Reviews Using Fine-Tuned Indobert With Class Weights Strategy

Abdan Syakura — 2026-06-15

MSMEs in the e-commerce sector face difficulties in converting large volumes of unstructured customer review data into actionable business insights. This challenge is exacerbated by the ambiguity of star ratings, which often do not align with the content of the reviews, making automated sentiment analysis of the text essential. This study implements a systematic sentiment analysis workflow on a case study of 15,278 customer reviews of Toko Pasar Stan Jogja. The method used is fine-tuning a pre-trained Transformer model, namely IndoBERT, which is optimized with class weighting techniques to handle unbalanced datasets. The model's performance was comprehensively evaluated using Accuracy, Precision, Recall, F1-Score, Confusion Matrix, and word cloud visualization metrics. The test results showed that the developed model had very high performance, achieving an overall accuracy of 96.99% and an average F1-Score of 0.97 on the test data. Qualitative analysis also successfully identified that product quality (“fresh”) and logistics efficiency (‘fast’) were the main drivers of satisfaction, while the main complaints centered on the condition of the product upon arrival (“damaged,” “rotten”). This research proves that the optimized Transformer model is not only effective for sentiment classification, but also serves as a strategic tool for extracting concrete business insights.

Constructing a Part-of-Speech Tagging based on Lexicon and Rule-based for Sundanese Corpus

Ade Sutedi — 2026-06-15

Part-of-Speech (POS) Tagging is the process of annotating word classes (nouns, verbs, adjectives, etc.) in a sentence, which is used as a basis for natural language processing and artificial intelligence. In this study, a corpus of word classes and word class annotating rules for the Sundanese language, which has limited resources, was developed. The experiments were conducted on an annotated corpus consisting of 104,696 tokens collected from Sundanese dictionaries, Sundanese Literature (Carita Pondok, Guguritan, Mantra, Pupujian, Sisindiran, Sajak, and Wawacan), Babasan and Paribasa, and social media X (Twitter). The annotation process is carried out in several stages that combine manual annotation based on cross-lingual transfer from Indonesian POS to Sundanese POS, then adjusted based on the word class rules in Sundanese. The results of this study are a POS annotation corpus containing Sundanese word-tag pairs and a basic rule-based model compared to the HMM and CRF models. The rule-based model achieves an F1-score of 0.867, the CRF model achieves an F1-score of 0.889, while the HMM model attains the highest score with an F1-score of 1.000. Analysis of POS distributions reveals that nouns (KB) consistently dominate across all models, reflecting the noun-rich nature of Sundanese literary texts. It also highlights the challenges of handling unknown words and the need for richer annotated resources, which are related to tag interoperability with Universal POS standards. This research contributes to the development of NLP resources for low-resource languages and provides a methodological foundation for future Sundanese NLP applications.

Comparative Evaluation of Linear Regression and Ensemble Learning Models for Daily Calorie Prediction Using a Public Lifestyle Dataset with Structured Preprocessing and Recursive Feature Elimination

Yunandra Wahyu Utama — 2026-06-15

Accurate daily calorie estimates are essential for personalized nutrition and prevention of diet-related conditions, yet lifestyle variability can reduce the effectiveness of one-size-fits-all recommendations. This study aims to develop an accurate lifestyle-based calorie estimation model by comparing an interpretable linear approach with ensemble machine learning methods. A publicly available lifestyle dataset from Kaggle was used, containing demographic variables, anthropometric measurements, food intake, dietary patterns, and physical activity attributes. A preprocessing pipeline was applied, including outlier handling using interquartile range capping, categorical encoding, normalization, and feature selection via Recursive Feature Elimination to identify the most relevant predictors. Four models (Linear Regression, Random Forest, XGBoost, and LightGBM) were trained and evaluated, followed by hyperparameter tuning of ensemble models using GridSearchCV. Performance was assessed using R², Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE) and training time. Linear Regression achieved the best overall performance (R² = 0.9650, MAE = 80.95, RMSE = 101.71, training time = 8.95 seconds). Among ensembles, the tuned XGBoost performed best (R² = 0.9646, MAE = 81.34, RMSE = 102.35, training time = 10.55 seconds). Compared with tuned XGBoost, Linear Regression was superior with MAE by 0.39 and RMSE by 0.64, while R² increased by 0.0004 and required less computational time, indicating that added complexity did not yield meaningful gains on this structured dataset. These findings suggest that, for structured lifestyle data, interpretable linear models can match or outperform complex ensembles while remaining computationally efficient for real-time or edge-deployed health applications.

Implementation of IndoBERT for Sustainability Impact Assessment in University Collaboration Information Systems

Ryan Hamonangan — 2026-06-15

University collaboration plays a critical role in enhancing institutional quality and supporting global sustainability agendas. However, many higher education institutions face challenges in managing Memorandum of Understanding (MoU), Memorandum of Agreement (MoA), and Implementation Agreement (IA) documents, particularly in monitoring implementation and assessing their alignment with sustainability goals. This study introduces a University Collaboration Information System enhanced with IndoBERT-based Natural Language Processing (NLP) to automate sustainability impact assessment. A synthetic corpus of 30 annotated collaboration documents was developed, covering multi-label Sustainable Development Goals (SDG) classification and span-level Named Entity Recognition (NER). Two approaches were evaluated: (1) baseline TF-IDF + Support Vector Machine (SVM) for SDG classification and rule-based NER, and (2) fine-tuned IndoBERT for both tasks. Experimental results show that IndoBERT significantly outperforms the baselines, achieving an average F1-score of 0.93 for SDG classification (+16.3%) and 0.96 for NER (+18.5%). The system integrates these models to generate automated entity extraction, sustainability dashboards, and document monitoring features. This work contributes to the advancement of informatics by demonstrating the effectiveness of Transformer-based NLP in processing institutional documents and by providing an integrated information-system framework that strengthens the role of NLP within the field of computer science.

Bank Customer Churn Prediction Using CTGAN-Augmented Data and Boosting-Based Ensemble Learning with SHAP Explainable AI

Mohamad Syazimmi Hersyaputra — 2026-06-15

Customer churn prediction remains a fundamental concern in the banking domain due to its direct impact on revenue stability and long-term customer value. A key challenge in churn modeling lies in severe class imbalance, which often limits model sensitivity toward minority churn cases. This study aims to develop an integrated and explainable churn prediction framework that effectively addresses class imbalance while maintaining robust predictive performance and interpretability. The proposed approach employs Conditional Tabular Generative Adversarial Networks (CTGAN), comparison of five boosting-based ensemble learning, and SHapley Additive exPlanations (SHAP) to preserve model interpretability. CTGAN is leveraged to synthesize high-fidelity instances for the churn class, yielding a class-balanced dataset that retains intricate tabular feature distributions. Five boosting-based ensemble models, XGBoost, CatBoost, Gradient Boosting Machine (GBM), Stochastic Gradient Boosting (SGB), and LightGBM, are systematically tuned using randomized hyperparameter optimization and evaluated under consistent experimental settings. Model performance is assessed using accuracy, precision, recall, and F1-score to capture classification performance under class imbalance. To ensure transparency, SHAP is applied to analyze global feature importance influencing churn predictions. Experimental results indicate CTGAN enhances model learning stability and detection capability. Among the evaluated models, CatBoost achieves the best results, with an accuracy of 0.9748 and an F1-score of 0.9178. The explainability analysis reveals that transactional features play a dominant role in churn. The novelty of this study lies in a unified and explainable churn prediction framework that integrates CTGAN-data augmentation, boosting ensembles, and interpretability for robust decision support in banking analytics.

Optimizing Naive Bayes for Sentiment Analysis of M-Passport Reviews Using N-Gram and Synthetic Minority Over-sampling Technique

Devia Kartika — 2026-06-15

The diverse user perceptions and increasing number of negative reviews of the M-Passport application indicate the need for sentiment analysis-based evaluation to more accurately measure the quality of digital immigration services. This study aims to analyze user sentiment towards the M-Passport application using an optimized Naïve Bayes classification model. Review data was obtained through web scraping from various digital platforms and processed using text preprocessing, TF-IDF feature extraction, N-Gram representation, and the Synthetic Minority Over-sampling Technique (SMOTE) technique to address data representativeness. The proposed model classifies user reviews into positive, neutral, and negative sentiment categories. Test results show that optimization using N-Gram and SMOTE successfully improved model performance, with accuracy increasing from 61% to 77.51%, precision from 0.75 to 0.78, recall from 0.53 to 0.78, and F1-score from 0.50 to 0.77. These results demonstrate that the combination of feature engineering and data balancing can improve text context representation and sentiment classification stability across multiple classes. Furthermore, sentiment analysis successfully identified key factors contributing to user dissatisfaction, such as technical constraints, feature limitations, and application difficulty. These results demonstrate that the proposed approach is effective in supporting data-driven evaluation to improve the quality of digital immigration services.

Systematic Review of TinyML at the Edge: Optimization, Applications, and Hardware Ecosystem

Very Kurnia Bakti — 2026-06-15

The Internet of Things (IoT) is growing rapidly, making it even more crucial to deploy Machine Learning (ML) models directly on edge devices with limited resources. TinyML fixes this matter by giving microcontroller-class hardware the ability to think for itself. This makes it less reliant on the cloud and better for latency, energy efficiency, and data privacy. This study offers a comprehensive Systematic Literature Review (SLR) of TinyML research published between 2021 and 2025, in accordance with PRISMA principles. We identified 429 records, removed 326 duplicates, and added 83 studies to the final synthesis. The evaluation examines five research inquiries concerning optimization techniques, streamlined architectures, sophisticated learning frameworks, application sectors, and hardware ecosystems. The findings underscore four key themes: enhancing models, utilizing specialized tools and technology, and adapting strategies. Some of the challenges that keep recurring are broken ecosystems, different benchmarking approaches, and on-device learning that isn't compelling when ideas shift. This research presents an open-access taxonomy that categorizes optimization techniques, application trends, and hardware constraints, thereby laying the foundation for a TinyML research agenda within the informatics community. Future directions highlight the importance of adaptive TinyMLOps pipelines, federated learning, LLM-assisted model design, and NVM‑based computing to support scalable and sustainable edge intelligence. The results underscore the relevance of TinyML for advancing informatics and computer science, particularly in enabling secure, efficient, and environmentally aligned IoT systems that support SDG 9 and SDG 12.