JITK (Jurnal Ilmu Pengetahuan dan Teknologi Komputer)

IMPLEMENTATION OF CNN FOR CLASSIFYING PATCHOULI LEAF IMAGES BASED ON ACCURACY AND EVALUATION

Arif Rahman Hakim — 2025-05-30

Patchouli (Nilam leaves) holds significant potential as a high-value natural material, especially in the perfume and essential oil industries. However, the classification and quality analysis of patchouli leaves remain a challenge that requires an automated solution based on technology. This study aims to develop a Convolutional Neural Network (CNN) model capable of automatically classifying the condition of patchouli leaves. The image data of patchouli leaves were processed through several preprocessing stages and divided into training and testing data. The designed CNN model utilizes several convolutional layers, pooling, dropout, and dense layers for the training process. The evaluation results using the confusion matrix showed that the model had a very low error rate, with only 1 misprediction in the training data. For the testing data, the model achieved an accuracy of 85% with a loss value of 0.6191496. The model also demonstrated an accuracy of 98.75% with a loss of 0.443462 on the training data. However, improvements in model generalization are still needed to achieve more consistent performance on new data

DEEP GATED RECURRENT UNITS PARAMETER TRANSFORMATION FOR OPTIMIZING ELECTRIC VEHICLE POPULATION PREDICTION ACCURACY

Jeni Sugiandi — 2025-05-30

The development of electric vehicles is an important innovation in reducing greenhouse gas emissions while reducing dependence on fossil fuels. The main problem in developing electric vehicles is the lack of adequate infrastructure. Inaccurate predictions regarding the number of electric vehicles hinder adequate infrastructure planning and development. This research proposes the use of the Gated Recurrent Units (GRU) algorithm to improve the accuracy of electric vehicle population predictions by carrying out GRU parameter transformations. This parameter transformation involves searching and adjusting the parameters of the GRU model in more depth to increase its ability to handle uncertainty in electric vehicle population data. After carrying out the training and testing process, the result was that the hyperparameter model using RandomizedSearchCV was the best model because it had the highest accuracy compared to other models tested with a combination of GRU_unit 64 and 128, dropout 0.5 and 0.6, batch size 64 and the number of epochs was 100 which had MAE results: 257.94, MSE: 66655.087, RMSE: 258.176, and Accuracy of 100%.

MODEL OF CYBERBULLYING DETECTION ON SOCIAL MEDIA USING MULTI-LABEL DEEP LEARNING: A COMPARATIVE STUDY

Lemi Iryani — 2025-05-30

Cyberbullying is the deliberate act of using technology to harm others. This study aims to analyze 400 Instagram comments obtained via API from previous research. The data were labeled into three classes: negative (containing cyberbullying), positive (non-bullying, supportive), and neutral (neither positive nor negative). The data for experiment was divided into 70% for training and 30% for testing. The research methodology consists of three main stages. The first stage is text preprocessing, which includes tokenization (splitting comments into tokens), filtering (removing unimportant words or stop-words), and stemming (converting words with affixes into their root forms). The second stage is classification analysis using BiLSTM, LSTM, RNN, and CNN-1D methods. The third stage is evaluation by comparing the model's classification results with manually labeled data using accuracy as the evaluation metric. The results show that the BiLSTM model performed the best, achieving an accuracy of 98.51% on the training data and 81.82% on the testing data. The BiLSTM method used in this study can be further adapted to enhance the effectiveness of cyberbullying detection in various applications.

FOREST FIRE LOCATION AND TIME RECOGNITION IN SOCIAL MEDIA TEXT USING XLM-ROBERTA

Hafidz Sanjaya — 2025-05-30

Forest fires have become a serious global threat, significantly impacting ecosystems, communities, and economies. Although remote sensing technology shows potential, limitations such as time delays, limited sensor coverage, and low resolution reduce its effectiveness for real-time forest fire detection. Additionally, social media can serve as a multimodal sensor, presenting multilingual text data with rapid and global coverage. However, it may encounter challenges in obtaining location and time information on forest fires due to limitations in datasets and model generalization. This study aims to develop a multilingual named entity recognition (NER) model to identify location and time entities of forest fires in social media texts such as tweets. Utilizing a transfer learning approach with the XLM-RoBERTa architecture, fine-tuning was performed using the general-purpose Nergrit corpus dataset containing 19 entities, which were relabeled into 3 main entities to detect location, date, and time entities from tweets. This approach significantly improves the model's ability to generalize to disaster domains across multiple languages and noisy social media texts. With a fine-tuning accuracy of 98.58% and a maximum validation accuracy of 96.50%, the model offers a novel capability for disaster management agencies to detect forest fires in a scalable, globally inclusive manner, enhancing disaster response and mitigation efforts.

APPLICATION OF MACHINE LEARNING MODELS FOR FRAUD DETECTION IN SYNTHETIC MOBILE FINANCIAL TRANSACTIONS

Imam Mulyana — 2025-05-30

The financial industry faces challenges in detecting fraud. The 2023 Basel Anti-Money Laundering (AML) Index report shows a worsening money laundering risk trend over the last five years in 107 countries. And according to the Financial Action Task Force (FATF) in 2023, this is exacerbated by financial institutions which have problems with low reporting of suspicious financial transactions (Suspicious Transaction Report). Limited access to confidential financial transaction data is an obstacle in developing machine learning-based fraud detection models. To overcome this challenge, the research uses PaySim synthetic datasets that mimic real financial transaction patterns. The CRISP-DM approach is used, including the Business Understanding, Data Understanding, Data Preparation, Modeling, Evaluation and Deployment stages. The algorithms used are Decision Tree, Random Forest, and XGBoost. Model evaluation is carried out using accuracy, precision, recall, F1-score, specificity, cross-validation and ROC-AUC metrics. The results show that the Random Forest algorithm has the best performance with 99% accuracy, followed by XGBoost (98.9%) and Decision Tree (97%). Data analysis shows that cash-out and transfer transactions have the highest risk of fraud. This model has proven effective in detecting suspicious financial transactions with a high level of accuracy. This research makes a significant contribution to mitigating financial risks, supporting anti-fraud policies, and encouraging innovation in fraud detection using synthetic data.

DEVELOPMENT OF SKIN CANCER PIGMENT IMAGE CLASSIFICATION USING A COMBINATION OF MOBILENETV2 AND CBAM

Juni Ismail — 2025-05-30

Skin cancer is one of the most common types of cancer worldwide, making early detection a crucial factor in improving patient recovery rates. This study compares three classification methods for pigmented skin cancer images using a combination of VGG16 with CBAM, MobileNetV2 with CBAM, and a hybrid VGG16-MobileNetV2 approach with transfer learning. The dataset used in this study is the Skin Cancer ISIC - The International Skin Imaging Collaboration (HAM10000) from Kaggle, which consists of 10,015 images covering seven types of skin cancer. After balancing, the dataset was reduced to 2,400 images with three main classes: Actinic Keratosis (AKIEC), Basal Cell Carcinoma (BCC), and melanoma (MEL), each containing 800 images. This study involves data preprocessing stages such as augmentation, normalization, and image resizing to ensure optimal data quality. The model training process was conducted using the Adam optimizer, a batch size of 16, and an Early Stopping mechanism to prevent overfitting. Evaluation results indicate that the MobileNetV2 with CBAM model achieved the best performance with a validation accuracy of 86%, followed by the VGG16-MobileNetV2 combination at 77%, while VGG16 with CBAM experienced overfitting with an accuracy of 54%. Additionally, the best-performing model demonstrated a precision of 86.53% and a recall of 86.46%, highlighting its superior stability in detecting skin cancer compared to previous single-model approaches. With these results, the developed system can serve as an effective tool for medical professionals in performing early and more accurate skin cancer diagnoses

CLASSIFICATION OF NATURAL DISASTERS IN WEST SEMARANG BASED ON WEATHER DATA USING DEEP LEARNING

Nicholas Martin — 2025-05-30

Natural disasters like floods, landslides, and fires pose serious threats to both life and mental well-being, especially in vulnerable areas like West Semarang, which frequently experiences extreme weather. To mitigate these risks, an accurate classification system is essential for timely prevention and response. This study compares the performance of three neural network models—Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU)—in classifying natural disasters using weather data. LSTM and GRU are particularly effective for handling long-term dependencies and addressing vanishing gradient problems common in time series data. Data for the study comes from the Semarang City Regional Disaster Management Agency (BPBD) and the Meteorology, Climatology, and Geophysics Agency (BMKG), spanning 2019 to 2022. The models achieved a high accuracy of 95.8%, but this is due to an imbalanced dataset—70 records of natural disasters versus 1377 without—resulting in classification favoring "no disaster." Among the models, LSTM performed the best, reaching optimal accuracy in just 20.0671 seconds per epoch. This suggests LSTM is the most effective model for this classification task.

COMPARATIVE OF LSTM AND GRU FOR TRAFFIC PREDICTION AT ADIPURA INTERSECTION, BANDAR LAMPUNG

Ilham Firman Ashari — 2025-05-30

The Tugu Adipura intersection in Bandar Lampung is a vital traffic hub connecting four major roads. Rapid population growth and increasing vehicle numbers challenge traffic flow and urban quality of life. Despite its importance, there is limited research using predictive models to analyze traffic patterns at complex intersections in mid-sized Indonesian cities. This study addresses that gap by examining traffic growth on four connected roads using deep learning models. Traffic data were collected hourly from June 1, 2021, to July 31, 2023. A comparative analysis of Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models was conducted, with SGD and Adam as optimizers. Results show the GRU model with Adam achieved the lowest RMSE (0.23) on road section 1, indicating its superior ability to model short-term fluctuations and non-linear growth in traffic volume. The study offers practical implications for traffic management by highlighting GRU’s capacity to capture seasonal trends and rapid growth, supporting proactive infrastructure planning and congestion mitigation strategies. These findings demonstrate the value of data-driven approaches in enhancing transportation systems in growing urban areas.

MODERN APPLICATION FOR IMPROVING AND REHABILITATING PRISONERS' MENTAL HEALTH

Yonata Laia — 2025-05-30

This study evaluates the Fuzzy Tsukamoto method as an effective rehabilitation solution for young inmates facing mental health challenges, including pre-existing conditions, confinement stress, and educational deficits. Mental health issues in correctional facilities remains a growing concern, affecting not only the well-being of inmates but also their chances of successful reintegration into society. The method employs Electroencephalogram/EEG to monitor tracked brain activity, providing real-time data that refined the treatment protocols and allowed for personalized adjustments. Conducted in a correctional facility in Medan, Indonesia, the study found significant reductions in anxiety and depression among participants, along with improved self-efficacy and emotional resilience. The results highlight the potential of the Fuzzy Tsukamoto method in not only improving inmate mental health but also in reducing recidivism rates and supporting social reintegration. These findings underscore the critical need to adopt more rehabilitative correctional strategies to address the complex mental health challenges within the incarcerated population.

OPTIMIZING TRANSPORTATION SURVEILLANCE WITH YOLOV7: DETECTION AND CLASSIFICATION OF VEHICLE LICENSE PLATE COLORS

Ridho Sholehurrohman — 2025-05-30

Optimizing transportation surveillance requires accurate vehicle license plate color detection and classification; however, existing systems face significant challenges in achieving real-time accuracy and robustness, particularly in crowded traffic scenarios with varying lighting and plate conditions. In Indonesia, vehicle license plates are color-coded based on their usage, including white and black for private vehicles, yellow for public vehicles, red for government vehicles, and green for free-trade areas. Each plate color plays a crucial role in transportation management, enabling proper vehicle identification and regulation. Existing surveillance systems struggle with real-time detection accuracy, especially in distinguishing plate colors in crowded traffic. Traditional methods may not efficiently classify plate colors due to limitations in feature extraction and processing. To address this, this study implements the YOLOv7 model to improve vehicle license plate color detection (black, white, yellow, and red) while distinguishing non-plate vehicles in diverse scenarios. The model's effectiveness is evaluated using precision, recall, and F1-score to ensure robustness for surveillance applications. Results show an average precision of 95.27%, recall of 94.60%, and F1-score of 94.93%, demonstrating strong detection capabilities. Optimizing the Non-Plate category further improves system accuracy, efficiency, and scalability, enhancing transportation monitoring reliability.

THE ROLE OF L1 REGULARIZATION IN ENHANCING LOGISTIC REGRESSION FOR EGG PRODUCTION PREDICTION

Nur Alamsyah — 2025-05-30

Poultry egg productivity is strongly influenced by various environmental factors, such as air and water quality. However, accurately predicting productivity remains a challenge due to the complex interplay of multiple environmental variables and the risk of overfitting in predictive models. This study improves egg productivity prediction using Logistic Regression with L1 regularization, which enhances model generalization by performing automatic feature selection. The research methodology includes data collection of key environmental indicators—Air Quality Index (AQI), Water Quality Index (WQI), and Humidex—followed by data preprocessing, exploratory data analysis (EDA), and model training using L1-regularized Logistic Regression. Model evaluation was performed using classification metrics and learning curve analysis to assess stability and effectiveness. Experimental results indicate that Logistic Regression without regularization achieved an accuracy of 90.7%, with misclassification occurring in the lower production categories. By applying L1 regularization, accuracy increased significantly to 97%, demonstrating its ability to reduce overfitting while improving classification performance. This study also compares its findings with previous research, such as De Col et al. (wheat epidemic prediction, 80–85% accuracy) and Jia Q1 et al. (heart disease prediction, 92.39% accuracy), confirming that our approach outperforms prior Logistic Regression models in similar predictive tasks. These findings suggest that L1 regularization is an effective solution for improving egg productivity prediction, particularly in scenarios with complex environmental influences. Future work will explore advanced ensemble learning techniques and real-time IoT-based monitoring to further enhance prediction accuracy and practical applicability.

COMPARATIVE STUDY OF YOLO VERSIONS FOR DETECTING VACANT CAR PARKING SPACES

Muhammad Fathurrahman — 2025-05-30

The increasing vehicle density in urban areas has made parking space availability a significant challenge. With technological advancements, efficient smart parking systems based on object detection have become essential. This study evaluates the performance of YOLO versions 3 to 11 in detecting vacant parking spaces in urban environments, focusing on real-time processing, high accuracy with limited datasets, and adaptability to varying conditions. Using 4,215 annotated images and two test videos, YOLOv7 achieved the highest overall accuracy of 99.57% with an average FPS of 30.79, making it the most effective model for smart parking applications. YOLOv8 and YOLOv11 followed closely, with accuracies of 98.51% and 98.72%, respectively, and average FPS rates of 32.31 and 31.99, balancing precision and speed, which are ideal for real-time applications. Meanwhile, YOLOv5 stood out for its exceptional processing speed of 33.92 FPS. These results highlight YOLO's potential to revolutionize smart parking systems by significantly enhancing both detection precision and operational efficiency.

OPTIMIZATION OF THE INCEPTIONV3 ARCHITECTURE FOR POTATO LEAF DISEASE CLASSIFICATION

Khairun Nisa Arifin Nur — 2025-05-30

Potato leaf diseases can cause significant yield losses, making early detection crucial to prevent major damages. This study aims to optimize the Inception V3 architecture in a Convolutional Neural Network (CNN) for potato leaf disease classification by applying Fine Tuning Pre-Trained. This method leverages weights from a pre-trained model on a large-scale dataset, enhancing accuracy while reducing the risk of overfitting. The training process involves adjusting several final layers of Inception V3 to better adapt to specific features of potato leaf diseases. The results show that this approach improves classification performance, achieving an accuracy of 97.78%, precision of 98%, recall of 98%, and an F1-score of 98%. With better computational efficiency compared to previous architectures, this model is expected to be widely applicable in plant disease detection systems, particularly for farmers or institutions with limited resources.

ENHANCING HERBAL PLANT LEAF IMAGE DETECTION ACCURACY THROUGH MOBILENET ARCHITECTURE OPTIMIZATION IN CNN

Anan Wibowo — 2025-05-30

Herbal plants have various health benefits, but their type identification remains challenging for the general public. This study aims to improve the accuracy of herbal plant leaf classification using Convolutional Neural Network (CNN) based on MobileNetV2 architecture. To enhance model performance, various optimization techniques including fine-tuning, batch normalization, dropout, and learning rate scheduling were implemented. The experimental results showed that the proposed optimized model achieved an accuracy of 100%, significantly outperforming previous studies that used standard MobileNet with an accuracy of 86.7%. While these perfect results warrant additional validation with more diverse datasets to confirm generalizability, this study contributes to the development of a more accurate herbal plant classification system that is readily accessible to the general public. Future work should explore model performance under varying environmental conditions and with expanded plant species datasets.

ZTSCAN: ENHANCING ZERO TRUST RESOURCE DISCOVERY WITH MASSCAN AND NMAP INTEGRATION

Reikal Taupaani — 2025-05-30

Implementing Zero Trust Architecture (ZTA) requires a comprehensive understanding of network assets as a fundamental step in implementing security policies. This study proposes ZTscan, an automated tool to increase the efficiency of network asset resource discovery. This proposed tool is then made open source in Github for anyone to evaluate and extend. The research constructs a GNS3-based testing scenario to evaluate the performance of the proposed tool against other scanning tools, including standalone Nmap, Masscan, RustScan, and ZMap. The evaluation focuses on three key metrics: accuracy, scanning speed, and generated data throughput. Experimental results demonstrate that ZTscan achieves 100% accuracy, matching Nmap_Pingsyn while outperforming faster tools such as Masscan, ZMap, and RustScan in precision. ZTscan completes scans 10.64%, faster than Nmap TCP SYN scan while maintaining comparable high accuracy. In terms of throughput, ZTscan reaches a stable peak throughput that is 13.8% lower than Nmap TCP SYN scan without causing disruptive traffic spikes. The findings of this study serve as a reference for resource discovery strategies in ZTA implementation, particularly in scenarios that require fast and accurate network scanning while minimizing potential disruptions or network interference.

DEVELOPMENT OF INFORMATION SYSTEM FOR EMPLOYEE PERFORMANCE ASSESSMENT AT HASNUR CENTRE USING 360° ASSESSMENT

Arif Muhammad Iqbal — 2025-06-04

Hasnur Centre is the CSR institution of Hasnur Group dedicated to the development of human resources in South Kalimantan. The performance assessment for Hasnur Centre employees currently relies on a conventional and unidirectional brief fill-in-the-blank method, reflecting the viewpoint of superior, indicating a need for an adjustment in the employed method. Furthermore, the employee evaluation process at Hasnur Centre still relies on a simple Google Form. Therefore, there is a need for the development of information system integration that can automate employee performance assessment at Hasnur Centre. The system is developed gradually according to the needs of the HR Admin, utilizing the Spiral development method. The tools include Use Case Diagrams, PHP as the programming language, CodeIgniter as the system development framework, and MySQL. This research has resulted in the Employee Performance Assessment Information System for Hasnur Centre employees, introducing a novelty by integrating the 360° Assessment method based on predetermined perspectives and sub-perspectives using a Likert Scale combined with a brief qualitative input method in which evaluators provide written feedback on the assessed employees. The combination of these two methods results in a more measurable, objective, and unbiased performance evaluation, making it a reliable tool for the Executive Director of Hasnur Centre in making decisions related to employee performance

SENTIMENT ANALYSIS OF PLAYER FEEDBACK IN ALGORUN: A STUDY OF DEEP LEARNING MODELS FOR GAME-BASED LEARNING

Rio Andriyat Krisdiawan — 2025-06-04

AlgoRun: Coding Game is a game-based learning application aimed at teaching computational thinking (CT) concepts such as variables, conditions, loops, and functions. Evaluating user feedback in such educational games is challenging, as traditional sentiment analysis techniques often overlook nuanced responses. Despite its potential to inform content improvements, sentiment analysis in game-based learning remains underexplored. This study compares the performance of deep learning models—DNN, CNN, RNN with LSTM, and Bidirectional LSTM—for sentiment classification of AlgoRun user reviews, using TF-IDF and word embeddings as feature extraction methods. A total of 1,440 reviews were scraped from the Google Play Store, translated, and preprocessed using data preparation techniques (dropna, fillna), text preprocessing (case folding, cleaning, tokenization, stopword removal, stemming), and feature extraction (TF-IDF and word embeddings). The dataset was labeled into negative, neutral, and positive classes, and split 80% for training and 20% for testing. Among the tested models, the DNN with TF-IDF achieved the highest accuracy of 98.86%, followed by CNN with Word Embeddings (96.97%), Bidirectional LSTM (96.59%), and RNN with LSTM (92.42%). The DNN also showed stable performance and convergence at the 10th epoch, outperforming other models in precision, recall, and F1-score. These results suggest that DNN with TF-IDF is highly effective for sentiment classification in the context of game-based learning. The findings offer useful guidance for developers to adapt content and enhance game quality based on user feedback. This research also contributes to the growing body of literature on leveraging sentiment analysis to optimize educational applications.

SENTIMENT ANALYSIS OF GOVERNMENT ON TIKTOK AND X PLATFORMS WITH SVM AND SMOTE APPROACH

Dimar Pateman — 2025-06-16

This study aims to analyze public sentiment toward the government on TikTok and X (formerly Twitter) using the Support Vector Machine (SVM) algorithm optimized with the Synthetic Minority Over-sampling Technique (SMOTE). Data were collected through keyword-based scraping of posts containing the word “pemerintah” (government) and processed using standard NLP pre-processing techniques. Results show that SVM combined with SMOTE significantly improves classification accuracy from 61% to 76% on TikTok, and from 74% to 86% on X. Word cloud analysis confirms these findings: TikTok content tends to reflect neutral and positive sentiments, while X contains predominantly negative expressions. These differences highlight platform-specific public discourse characteristics. The findings suggest that public communication strategies should be tailored accordingly: TikTok for positive narrative and outreach, X for monitoring feedback and criticism. This approach demonstrates the effectiveness of machine learning-based sentiment analysis in supporting data-driven public policy communication.

PWA AND NON-PWA PERFORMANCE ANALYSIS: CHROME EXTENSION TESTING ON E-COMMERCE PLATFRORM

Panji Revolusioner Saputro — 2025-06-04

This study compares Progressive Web Apps (PWA) and traditional web applications performance using a custom Chrome extension and Google Lighthouse, focusing on Tokopedia's e-commerce platform. The research employs a quantitative approach with controlled testing environments across three viewports for the custom extension (desktop, tablet, mobile) and two viewports for Google Lighthouse (desktop, mobile). The custom extension measures eleven metrics, including Core Web Vitals, PWA features, and resource usage, while Google Lighthouse provides five core metrics. Results show PWA implementation improves performance with 9.9% better First Contentful Paint on desktop and significant memory efficiency (29-33MB vs 59-62MB). The comparison between testing tools reveals methodology differences, with custom extension showing optimistic results in real-world conditions and Lighthouse providing more conservative measurements under throttled conditions. This research contributes to PWA performance measurement methodology by combining real-world and standardized testing approaches.

OPTIMIZATION CVRP WITH MACHINE LEARNING FOR IMPROVED CLASSIFICATION OF IMBALANCED DATA FOOD DISTRIBUTION

Muhammad Syahputra Novelan — 2025-06-09

The classification of imbalanced data in food delivery distribution is an important issue that needs to be considered to ensure fairness and efficiency in the food distribution system. This research answers these problems by improving the accuracy of the classification of delivery locations that have imbalanced demand data, so that high priority areas are not neglected. Generating more efficient and cost-effective distribution routes, taking into account vehicle capacity and delivery urgency. Reducing delivery time and potential food waste due to delays or non-optimal route allocation. This study addresses the problem of improving the accuracy of delivery location classification that has imbalanced demand data, so that high priority areas are not neglected. Generate more efficient and cost-effective distribution routes, taking into account vehicle capacity and delivery urgency. Reduce delivery time and potential food wastage due to delays or non-optimal route allocation. This study uses the research stages of data collecting, data preprocessing, and implementation of K-Means and K-NN methods. The results of CVRP testing with K-Means show the value of cluster 7 acc=80, precc=85, recall=84. cluster 9 acc=85, precc=90, recall=91. cluster 11 acc=88, precc=93, recall=94. While the results of CVRP testing with K-NN show the value of K 7 acc=89, precc=88, recall=85. value of K 9 acc=87, precc=90, recall=91. value of K 11 acc=95, precc=97, recall=94. The optimization results show that this approach not only improves operational efficiency but also increases the accuracy of food delivery, which will affect the availability of traditional markets.

PERFORM COMPARATION OF DEEP LEARNING METHODS IN GENDER CLASSIFICATION FROM FACIAL IMAGES

Yosefina Finsensia Riti — 2025-06-09

Identifying gender through facial images is a crucial aspect in various life contexts. Biometric technology, such as facial recognition, has become an integral part of various applications, including fraud detection, cybersecurity protection, and consumer behavior analysis. With the advancement of technology and the progress in artificial intelligence, especially through the use of Convolutional Neural Networks (CNNs), computers can now identify gender from facial images with a high level of accuracy. Although there are still some challenges, such as variations in pose, facial expressions, and different lighting conditions, CNNs can overcome these obstacles. This study uses the CelebA dataset, which consists of 122,000 facial images of both men and women. The dataset has been processed to maintain a balanced number of samples for each gender class, resulting in a total of 101,568 samples. The data is divided into training, validation, and test sets, with 80% used for training, and the remaining 20% split between validation and testing. Eight different CNN architectures are applied, including VGG16, VGG19, MobileNetV2, ResNet-50, ResNet-50 V2, Inception V3, Inception ResNet V2, and AlexNet. Although previous research has shown the potential of CNN architectures for various classification tasks, these studies often encounter issues of overfitting on large datasets, which can reduce model accuracy. This study applies dropout techniques and hyperparameter tuning to address overfitting issues and optimize model performance. The training results indicate that ResNet-50, ResNet-50 V2, and Inception V3 achieved the highest accuracy of 98%, while VGG16, VGG19, MobileNetV2, and AlexNet achieved accuracies of 95% and 97%, respectively. Performance evaluation using confusion matrices, precision, recall, and F1-score demonstrates excellent performance.

IMPLEMENTATION MEAN IMPUTATION AND OUTLIER DETECTION FOR LOAN PREDICTION USING THE RANDOM FOREST ALGORITHM

Nimatul Mamuriyah — 2025-06-10

Loans and credit are among the most in-demand banking products, making accurate loan prediction systems essential for minimizing bank credit risks and boosting profitability. This study proposed a loan prediction model using the Random Forest algorithm, with mean imputation and 3 outlier detection (Boxplot, Z-score, and Interquartile Range (IQR)) as data pre-processing methods. Using Lending Club loan data from 2014-2021 (466,285 records, split 70/30 for training/testing), model performance was assessed using accuracy, recall, and F1 Score. The proposed approach achieved a 95% prediction accuracy, outperforming previous models at 83%. The best results were obtained using mean imputation with IQR-based outlier detection. However, the determination of the mean imputation mean can be a limitation of this study. This highlights the importance of thorough pre-processing in enhancing prediction accuracy. The study underscores the role of machine learning and financial technology (fintech) in informing credit decisions and support incorporating imputation and outlier handling as standard steps in financial modeling pipeline

THE ROLE OF THE INTERNET OF THINGS (IOT) IN ELECTRIC VEHICLE MANAGEMENT AND MAINTENANCE

Callista Fabiola Candraningtyas — 2025-06-10

The growing adoption of electric vehicles (EVs) as an eco-friendly alternative to fossil fuel-based vehicles necessitates more advanced management and maintenance systems. The Internet of Things (IoT) presents significant potential to enhance EV management by enabling real-time monitoring and data analysis through interconnected sensors and technologies. This research investigates the integration of IoT in electric vehicle systems, focusing on real-time battery health monitoring, early detection of technical issues, and route optimization for improved energy efficiency. The study employs a system design and testing approach, supported by descriptive-analytical analysis using data from case studies, literature reviews, surveys, and interviews. Findings indicate that IoT implementation in EVs yields notable advantages. Real-time battery health tracking provides accurate performance insights, achieving a 92% accuracy rate in predicting battery degradation. Technical problem detection through sensor analysis enables timely maintenance, leading to a 30% reduction in vehicle downtime. Furthermore, IoT-based route optimization improves energy efficiency, reducing energy consumption by 15% and extending battery lifespan by 20% compared to traditional systems. These results underscore the practical benefits of IoT in enhancing EV performance and operational efficiency. The system enables users and service providers to make informed decisions regarding vehicle maintenance and usage, promoting better understanding of battery conditions. Ultimately, the application of IoT technology contributes to extending battery life, minimizing vehicle downtime, and supporting broader efforts in energy efficiency and carbon emission reduction

MEASURING PERCEIVED USABILITY OF ARTIFICIAL INTELLIGENCE-BASED QUIZZES IN A VIRTUAL MUSEUM

Shinta Puspasari — 2025-06-16

The transformation of modern museums through digital technology offers added value to visitors, especially in the context of education. Virtual museums, in particular, complement physical museums by providing accessibility and enhancing the learning experience. The SMBII virtual museum includes an AI-based quizzes feature designed to assess the knowledge level of visitors regarding the museum's history and collections as an educational feature. In addition to physical museums, virtual museums offer convenience and enrich the learning process for visitors. The quizzes adapts its questions based on the visitor's profile, leveraging AI to tailor content and maximize learning outcomes. This study aims to compare the effectiveness of two widely used usability metrics—System Usability Scale (SUS) and Usability Metric for User Experience (UMUX)—in evaluating the usability of the AI-driven quiz feature within the SMBII virtual museum. The study specifically seeks to determine whether there are significant differences between SUS and UMUX in measuring user perceptions of the quiz’s usability. The primary respondents of this study were students, who represent the museum's target audience for educational purposes. Hypothesis testing results show no significant difference between the SUS and UMUX scores (P > 0.05), indicating that both metrics offer similar evaluations of usability. Based on these findings, the study recommends the use of UMUX over SUS for future usability assessments in virtual museum systems, as UMUX is more time-efficient without compromising accuracy. This research contributes to advancing the understanding of usability testing methods for AI-based educational features in virtual museum environments

DETECTION OF FRAUDULENT ATM TRANSACTIONS USING RULE-BASED CLASSIFICATION TECHNIQUES

Deni Ekel Ramanda Sembiring Pelawi — 2025-06-16

The significant rise in ATM fraud—reflected in 130,472 suspicious transactions reported in Indonesia in 2022—highlights the urgent need for accurate and efficient real-time fraud detection systems. This study evaluates two complementary detection approaches using a dataset of 20,000 anonymized ATM transactions collected from XYZ Bank between January and December 2022, each labeled by internal fraud analysts as fraud or non-fraud. The models compared are a Rule-Based Classifier and a Decision Tree classifier. The Decision Tree demonstrates strong overall performance, achieving 98% accuracy, 75% precision, 79% recall, and a 77% F1-score, indicating a reliable ability to detect diverse fraud patterns. In contrast, the Rule-Based Classifier yields 60% accuracy, 97% precision, 60% recall, and a 74% F1-score, showing high precision with fewer false alarms but a limited ability to detect varied fraud cases. These results emphasize the trade-off between specificity and sensitivity in static versus adaptive models. To address this, a hybrid detection framework is proposed—combining rule-based screening to filter obvious non-fraud cases, followed by Decision Tree analysis to handle more complex patterns. This approach aims to reduce unnecessary transaction holds and improve detection reliability. This study contributes to the limited comparative research on fraud detection methods using real ATM transaction data within the Indonesian banking context. Future research will focus on adaptive learning models to maintain performance against evolving fraud behaviors in dynamic financial systems.

INTEGRATING AUGMENTED REALITY WITH C4.5 ALGORITHM TO ENHANCE TOURISM EXPERIENCE IN PEKALONGAN

Muhamad Rizaludin — 2025-06-16

The tourism industry demands interactive and personalized solutions to enhance the traveler experience. However, providing relevant and customized travel recommendations based on individual preferences remains a challenge. This study integrates Augmented Reality (AR) technology with the C4.5 algorithm to address this issue and improve the tourism experience in Pekalongan Regency. The research method involved collecting data from 500 respondents through an online questionnaire. The collected data underwent preprocessing, including handling missing data, data transformation, and class balancing. The C4.5 algorithm was applied to build a tourism recommendation model, while AR technology presented 3D visualizations of tourist destinations through an interactive application. The research results show that the recommendation model achieved an accuracy rate of 76.92%. The integration of AR provided an interactive experience that enhanced tourist engagement and satisfaction, although limitations were found in AR visualization quality and the completeness of destination information. Further improvements are needed to enhance AR realism, provide more detailed content, and optimize user satisfaction. This study contributes to the development of AR-based tourism technology integrated with the C4.5 algorithm. The findings encourage local tourism innovation and have the potential to enhance the traveler experience in Pekalongan Regency. This model can also be applied to other tourist destinations across Indonesia.

DEVELOPMENT OF CNN-LSTM-BASED IMAGE CAPTIONING DATASET TO ENHANCE VISUAL ACCESSIBILITY FOR DISABILITIES

Muhammad Rifki — 2025-06-18

Visual accessibility in public spaces remains limited for individuals with visual impairments in Indonesia, despite technological advancements such as image captioning. This study aims to develop a custom dataset and a baseline CNN-LSTM image captioning model capable of describing sidewalk accessibility conditions in Indonesian language. The methodology includes collecting 748 annotated images from various Indonesian cities, with captions manually crafted to reflect accessibility features. The model employs DenseNet201 as the CNN encoder and LSTM as the decoder, with 70% of the data used for training and 30% for validation. Evaluation was conducted using BLEU and CIDEr metrics. Results show a BLEU-4 score of 0.27 and a CIDEr score of 0.56, indicating moderate alignment between model-generated and reference captions. While the absence of an attention mechanism and the limited dataset size constrain overall performance, the model demonstrates the ability to identify key elements such as tactile paving, signage, and pedestrian barriers. This study contributes to assistive technology development in a low-resource language context, providing foundational work for future research. Enhancements through data expansion, incorporation of attention mechanisms, and transformer-based models are recommended to improve descriptive richness and accuracy