Jurnal Teknik Informatika (Jutif)

The Role Of A Decision Support System In Enhancing The Management Of Sexual Violence Cases In Higher Education Using The Saw Method Through An Android-Based Application

Elyza Gustri Wahyuni — 2025-06-23

The rising prevalence of sexual violence in higher education institutions demands urgent attention, highlighting the need for an efficient and responsive reporting system. Android-based applications for reporting sexual violence play a vital role in addressing this issue. This research proposes an application designed to provide ease of access and usability, enabling victims or witnesses to promptly submit reports supported by video, audio, and real-time GPS data. Such empirical evidence increases the likelihood of successful follow-up actions and strengthens legal claims against perpetrators. Timely responses are especially critical for the Sexual Violence Prevention and Handling Task Force (PPKS) to mobilize campus security teams effectively and reduce long-term trauma experienced by victims. An integral component of the application is a Decision Support System (DSS) that utilizes the Simple Additive Weighting (SAW) method to assess the severity of reported cases—categorized into mild, moderate, or severe. This system facilitates faster and more accurate decision-making during the investigation and handling phases. Functional and case testing resulted in 100% success, aligning perfectly with manual calculations and real-world scenarios. The urgency of this research lies in the pressing need for a reporting system that is not only reactive but also proactive in preventing sexual violence. The application demonstrates strong potential to support systemic reform in campus reporting mechanisms, enhance victim trust in reporting processes, and shift the paradigm from reactive intervention to preventive action. Ultimately, this research contributes to building a safer, more responsive, and survivor-centered campus environment.

Marketing Analysis of Shoe Products Using Principal Coordinates Analysis and K-Means Clustering Based on the Marketing Mix at Bintang Sepatu Purwokerto MSME

Samuel Sinaga — 2025-06-23

Bintang Sepatu Purwokerto MSME is a micro, small, and medium enterprise engaged in the production of local shoes. Recently, this MSME faced a significant issue in the marketing aspect, namely the low achievement of sales targets. Consequently, inventory will accumulate in the warehouse. Accordingly, this research aimed to formulate targeted marketing strategies by clustering customers based on demographic and marketing mix influencing purchasing behavior. This study applied principal coordinate analysis (PCoA) and k-means clustering to manage categorical and numerical data types within the dataset comprising 179 customers and 16 attributes.. The PCoA algorithm was utilized to derive object configurations that were subsequently employed in k-means. The clustering result produced three clusters with good clustering quality based on the Silhouette score, namely 0.790, indicating accurate and representative segmentation. Each cluster obtained had a different customer characteristic. The first cluster, comprising 68 customers (38%), was oriented towards fundamental needs and tended to shop traditionally, classified as a segment of conventional rational customers. Additionally, the second cluster, with 70 customers (39%), exhibited planned and stable decision-making, categorized as mature rational customers. Furthermore, the third cluster comprises 41 customers (23%) who are digitally aware and combine conventional shopping approaches with technological utilization, identified as rational consumers. The segmentation results provide a data-driven foundation for designing targeted marketing strategies, thereby potentially increasing sales, supporting the sustainability of MSMEs, and encouraging the application of unsupervised learning techniques in decision-making processes.

Stock Price Prediction and Risk Estimation Using Hybrid CNN-LSTM and VaR-ECF

Alvi Yuana Febriyanti — 2025-06-30

Stock price prediction is a major challenge in the financial domain due to high volatility and complex movement patterns. Traditional methods such as fundamental and technical analysis often fail to capture the non-linear characteristics and fast-changing market dynamics, highlighting the need for more adaptive approaches. This study proposes a hybrid deep learning model, CNN-LSTM, which combines CNN's local feature extraction capabilities with LSTM’s ability to model long-term temporal dependencies. To incorporate risk management, the model is also integrated with the Value at Risk (VaR) approach using the Cornish-Fisher Expansion (ECF) to estimate potential losses under extreme market conditions. The study utilizes daily historical stock price data of PT Unilever Indonesia Tbk retrieved from Yahoo Finance. Model performance is evaluated using Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE), where the model achieves an MAE of 78.13 and a MAPE of 2.72%, indicating relatively low absolute and relative prediction errors. These results confirm that the CNN-LSTM approach effectively models stock price movements in dynamic market environments, and the integration with VaR-ECF provides a more comprehensive risk estimate. Thus, this approach not only enhances predictive accuracy but also offers valuable decision-support tools for investors in planning investment strategies.

Identifying Academic Excellence: Fuzzy Subtractive Clustering of Student Learning Outcomes

Muhammad Bagas Satrio Wibowo — 2025-06-30

Education forms a vital foundation for a nation's future. In this digital era, while the use of Information and Communication Technology (ICT) in education is increasing, it brings increasingly complex challenges in education data management and analysis. The growing number of students each year results in a large volume of data, which would be difficult to manage if still relying on manual methods. Manual approaches are inefficient, time-consuming, prone to inconsistencies and human error, especially when identifying outstanding students in large and complex data. This research aims to implement a clustering system to group outstanding students at XYZ elementary school using the Fuzzy Subtractive Clustering (FSC) method. FSC was chosen for its ability to identify data groups based on the density of data points. FSC involves several important parameters, including radius, squash factor, acceptance ratio, and rejection ratio. Added variabel of social and spiritual values aims to enhance grouping quality by offering a broader perspective on students' character, attitudes, and social interactions. Parameter exploration shows an increase in the silhouette score from 0.20–0.45 to 0.45-0.57 and variable addition spiritual and social values, which indicates clearer cluster separation and provides better insights. The best parameters results were achieved with radius 0.3, accept ratio 0.5, reject ratio 0.04, and squash factor 1.25, resulting in a Silhouette Score of 0.57 and forming 5 student groups. Cluster results can guide special mentoring for students with low academic, spiritual, and social values, and support personalized learning programs based on each cluster’s characteristics.

Classification of Helmet and Vest Usage for Occupational Safety Monitoring using Backpropagation Neural NetworkClassification of Helmet and Vest Usage for Occupational Safety Monitoring using Backpropagation Neural Network

Nurhikma Arifin — 2025-06-10

Occupational Safety and Health (OSH) is a critical aspect in high-risk work environments, where the consistent use of Personal Protective Equipment (PPE) plays a vital role in preventing workplace accidents. However, non-compliance with PPE regulations remains a significant issue, contributing to a high number of work-related injuries in Indonesia. This study proposes an automated detection and classification system for PPE usage, specifically helmets and vests, using the Backpropagation algorithm in artificial neural networks. A total of 100 images were utilized, equally divided between complete and incomplete PPE usage. The dataset was split into 60% training and 40% testing. Image segmentation was performed using HSV color space conversion and thresholding, followed by RGB color feature extraction. The Backpropagation algorithm was then employed for classification. Experimental results show an average accuracy of 90%, with precision, recall, and F-measure all reaching 0.9. Despite some misclassifications due to color similarity between helmets and head coverings, the model demonstrated robust performance with relatively low computational requirements. This study contributes to the field of computer vision and intelligent safety systems by demonstrating the practical effectiveness of lightweight ANN architectures for PPE detection in real-time industrial scenarios, thereby highlighting the potential of backpropagation as an adaptive and practical alternative to more complex deep learning approaches for real-time PPE detection in occupational safety monitoring systems.

Optimization Of Extreme Learning Machine Models Using Metaheuristic Approaches For Diabetes Classification

Gilang Sulaeman — 2025-06-23

Proper classification of diabetes is a significant challenge in contemporary healthcare, especially related to early detection and clinical decision support systems. This study aims to optimize the Extreme Learning Machine (ELM) model with a metaheuristic approach to improve performance in diabetes classification. The data used was an open dataset containing the patient's medical attributes, such as age, gender, smoking status, body mass index, blood glucose level, and HbA1c. The initial process includes data cleansing, one-hot coding for categorical features, MinMax normalization, and unbalanced data handling with SMOTE. The ELM model was tested with four activation functions (Sigmoid, ReLU, Tanh, and RBF) each combined with three metaheuristic optimization strategies, namely Particle Swarm Optimization (PSO), Genetic Algorithm (GA), and Bat Algorithm. The results of the evaluation showed that the combination of the Tanh activation function with GA optimization obtained the highest accuracy of 87.98% and an F1-score of 0.5489. Overall, GA optimization appears to be superior to all other measurement configurations in consistent classification performance. The main contribution of this study is to offer a systematic approach to select the best combination of activation functions and optimization algorithms in ELM, as well as to provide empirical evidence to support the application of metaheuristic strategies to improve the accuracy of disease classification based on health data. This research has direct implications for the development of a more precise and data-based medical diagnostic classification system for diabetes.

Multivariate Forecasting of Paddy Production: A Comparative Study of Machine Learning Models

Feri Yasin — 2025-06-23

Accurate rice production forecasting plays an important role in supporting national food security planning. This study aims to evaluate the performance of four machine learning algorithms, namely Random Forest, XGBoost, Support Vector Regression (SVR), and Linear Regression, in predicting three target variables simultaneously: harvest area, productivity, and production. The dataset used includes annual data per province in Indonesia from 2018 to 2024 obtained from the Central Statistics Agency (BPS). Evaluation was conducted using five metrics: MAE, RMSE, MAPE, R², and training time. The results of the experiment showed that the Random Forest Regressor performed best in the 80:20 scenario, with an MAE of 76,259.52, an RMSE of 154,036.91, a MAPE of 0.61%, and an R² of 0.997. XGBoost showed a competitive performance with an MAE of 79,381.44 and faster training times. In contrast, the SVR showed the worst performance with the MAPE reaching 198.56% and the R² of 0.209. Linear Regression as baseline recorded an MAE of 1,194,355.28 and an R² of 0.503, indicating that the linear model is not effective enough for this data. The 80:20 scenario is considered the best configuration because it is able to balance the accuracy and generalization of the model. These findings show that the use of ensemble algorithms, especially Random Forest and XGBoost, has the potential to be applied practically by agricultural agencies or local governments in designing data-driven policies for more proactive and predictive rice production management. Furthermore, this study contributes to the advancement of applied informatics by demonstrating how machine learning models can be effectively used in multivariate forecasting for complex, real-world problems, thereby supporting the development of intelligent decision-support systems in the agricultural domain.

Performance Optimization of ERD Designs Using Cost-Based Optimization for Large-Scale Query Processing

Juanda Hakim Lubis — 2025-06-23

The rapid growth of stored data, particularly on magnetic disks, is doubling annually for each department within a company, creating a pressing need for efficient database management. While database design is a fundamental step in establishing a high-performance system, it alone is insufficient to ensure optimal efficiency. Query optimization plays a critical role in improving data transaction speed, reducing query execution time, and enhancing overall system responsiveness. This study evaluates various relational database models under different data volumes to analyze their impact on query performance. Using the Cost-Based Optimizer method and access time measurements, we assess query costs and determine the factors influencing performance. The results indicate that among the three database models analyzed, ERD-3 consistently delivers superior performance, especially in handling complex queries. This is attributed to its modular structure, strategic indexing, and reduced full table scans, which collectively minimize query execution costs. Additionally, several key factors significantly affect query performance, including record count, attribute size, query complexity, primary and unique key usage, indexing strategies, order-by clauses, index sequences, and SQL function application. This research contributes to the field of database optimization by demonstrating that ERD structuring and cost-based query analysis significantly improve system efficiency in large-scale environments. These findings emphasize the necessity of a well-structured, scalable database model and efficient query processing techniques to accommodate large-scale data growth. The study’s conclusions provide a foundation for advanced optimization strategies, ensuring that modern database systems remain efficient and adaptable to evolving data demands.

Comparative Analysis of Machine Learning Algorithms with RFE-CV for Student Dropout Prediction

Sekar Gesti Amalia Utami — 2025-06-23

The high dropout rate of students in higher education is a problem faced by educational institutions, impacting quality assessments and accreditation evaluations by BAN-PT. This study aims to develop an early prediction model of potential dropout students using demographic data with a learning analytics approach. Five classification algorithms are used in this research, namely Random Forest (RF), Decision Tree (DT), Logistic Regression (LR), Light Gradient Boosting Machine (LGBM), and Support Vector Machine (SVM). The dataset used consists of undergraduate student data of Sebelas Maret University in 2013 (n=2476) which is processed through preprocessing techniques, resampling with SMOTE, and validation using K-Fold Cross-Validation. The results showed that the RF model gave the best performance with an accuracy of 96.01%, followed by LGBM (95.26%), DT (91.24%), LR (83.68%), and SVM (83.19%). The use of the Recursive Feature Elimination with Cross-Validation (RFE-CV) feature selection method was able to improve the efficiency of the model by reducing the number of features without significantly degrading performance. The best feature selection was obtained when using 75% features, which provided an optimal balance between the number of features and model accuracy. The most contributing features include IPS_range (Semester GPA range), parents' income, students' regional origin, as well as several other demographic factors. This study contributes to the development of early warning systems in higher education by providing accurate predictive models and identifying key risk factors.

Impact of Optimizer Selection on MobileNetV1 Performance for Skin Disease Detection Using Digital Images

Khairul Fathan Habie — 2025-07-09

Automatic detection of skin diseases using digital images is a growing field in the application of deep learning in the medical world, especially to help the early diagnosis process. One of the most widely used models is MobileNetV1 because it is lightweight and efficient in image processing. However, the performance of the model is greatly affected by the training configuration, including the type of optimizer used. This study aims to compare the effectiveness of six types of optimizers, namely SGD, RMSprop, Adam, Adadelta, Adagrad, Adamax, and Nadam in training MobileNetV1 models for human skin disease image classification. The model was trained on annotated skin image dataset with predetermined training parameters: batch size 32, learning rate of 0.0001, and 10 epochs. Performance evaluation was performed using accuracy metrics. The results obtained demonstrate that RMSprop performs best, with 99.10% accuracy, 99.14% precision, 99.10% recall, and a 99.10% F1-score. Adadelta showed the lowest performance consistently, with only 22.22% accuracy, 20.34% precision, 22.22% recall, and 18.42% F1-score. This finding confirms that the type of optimizer affects the effectiveness of model training, especially in medical image classification tasks. This research provides empirical insights that are useful in selecting the optimal optimizer for MobileNetV1 model implementation in the healthcare domain.

A Hybrid Recommendation System for Pregnancy-Safe Skincare: Integrating Keyword-Based and Rule-Based Classification with Content-Based Filtering

Jihan Syahira Adnanda Putri — 2025-07-09

Pregnant women struggle to select safe skincare products, often relying on social media and blog searches, and manual ingredient checking. Choosing safe ingredients is essential, as exposure to unsafe substances may lead to teratogenic effects and endocrine disruption, which can result in fetal abnormalities such as retinoic acid embryopathy and neurodevelopmental disorders. Exposure to retinoids, for instance, has been associated with a 20–30% incidence of fetal retinoid syndrome in affected pregnancies. This study develops an integrated recommendation system using three techniques: (1) keyword-based classification with regular expressions to detect 50 unsafe ingredients across 8 categories; (2) rule-based classification using IF-THEN statements matching products with 5 pregnancy-related skin conditions; and (3) content-based filtering utilizing TF-IDF vectorization and cosine similarity for safer alternatives. The system achieved 86.25% accuracy in safety classification, with high recall (97.50%) indicating strong ability to identify safe products. However, moderate precision (79.59%) suggests some unsafe products were misclassified as safe, highlighting need for improvement in safety-critical contexts. Pilot user evaluation using ResQue framework with 10 participants yielded scores of 4.50–4.85 across 8 dimensions, achieving 4.65 overall average. This research demonstrates effective integration of multiple recommendation methods in context-sensitive applications, enabling safer product selection during pregnancy. By providing accessible, personalized, and evidence-based information, the system enables pregnant women to make informed skincare decisions and continue their routines despite limited access to healthcare services.

Optimal Phase Selection Of Single-Phase Appliances In Buildings Using String-Coded Genetic Algorithm

Novalio Daratha — 2025-06-23

Phase imbalance in buildings, primarily caused by single-phase loads and generation, leads to increased neutral current, voltage imbalance, reduced energy efficiency, and potential equipment damage. To address these challenges, an optimal phase selection method is proposed for single-phase loads and generation. This method integrates integer programming with a string-coded genetic algorithm (GA). The GA employs string encoding to represent phase connections. Initially, a Mixed Integer Programming (MIP) solver identifies an initial solution, which is subsequently transformed into a string to initialize the GA’s genes. Subsequently, the GA executes standard operations such as mutation, crossover, evaluation, and selection. Case studies demonstrate the efficacy of this method in achieving substantial load balancing. Notably, the identification of multiple solutions with identical objective function values renders this approach suitable for smart buildings equipped with energy management systems that participate in ancillary services between low-voltage and medium-voltage networks. This research pertains to the domains of computer science, power engineering, and energy informatics.

Enhancing Monkeypox Skin Lesion Classification With Resnet50v2: The Impact Of Pre-Trained Models From Medical And General Domains

Saifulloh Azhar — 2025-06-23

The monkeypox outbreak has emerged as a pressing global health concern, as evidenced by the rising number of cases reported in various countries. This rare zoonotic disease, caused by the Monkeypox virus (MPXV) of the Poxviridae family, is commonly found in Africa. However, since 2022, cases have also spread to various countries, including Indonesia. The dermatological symptoms exhibited by affected individuals vary, with the potential for further transmission through contamination. Early and accurate detection of monkey pox disease is therefore essential for effective treatment. The present study aims to improve the classification of Monkey Pox using the modified Resnet50V2 model, trained using pre-training datasets namely ImageNet and HAM10000, where batch size and learning rate parameters were adjusted. The study achieved high accuracy in distinguishing monkeypox cases, with 98.43% accuracy for Resnet50V2 with pretrained ImageNet and 70.57% accuracy for Resnet50V2 with pretrained HAM10000. Future research will focus on refining these models, exploring hybrid approaches incorporating convolutional neural networks, this advancement contributes to the development of automated early diagnosis tools for monkeypox skin conditions, especially in resource-limited clinical settings where access to dermatology experts is limited.

Optimization Artificial Neural Network (ANN) Models with Adam Optimizer to Improve Customer Satisfaction Business Banking Prediction

Yahya Nur Ifriza — 2025-06-23

Customer satisfaction prediction is critical for business banking to retain clients and optimize services, yet existing models struggle with imbalanced data and suboptimal convergence. Traditional approaches lack adaptive learning mechanisms, limiting accuracy in real-world applications. This study developed an optimized Artificial Neural Network (ANN) model using the Adam algorithm to improve prediction accuracy for banking customer satisfaction. We trained an ANN on the Santander Customer Satisfaction Dataset (76,019 entries, 371 features) with Adam optimization. Preprocessing included normalization, removal of quasi-constant features, and an 80-20 train-test split. Adam’s adaptive learning rates and momentum were leveraged to address gradient instability. The model achieved 95.82% accuracy, 99.99% precision, 95.83% recall, a 97.87% F1-score, and 0.82 AUC, outperforming traditional optimizers like SGD. Training loss reduced by 30% with faster convergence. This work demonstrates Adam’s efficacy in handling imbalanced banking data, providing a scalable framework for customer analytics. The results advance computer science applications in fintech by integrating adaptive optimization with deep learning for high-stakes decision-making. This research contributes to the growing body of knowledge in machine learning applications for business analytics and provides a valuable framework for improving customer satisfaction prediction models in various industries and the advancement of deep learning applications in business intelligence, particularly in banking service quality prediction.

Enhancing The Precision Detection and Grading of Diabetic Retinopathy through Digital Retinal Imaging Using 3D Convolutional Neural Networks

Autho Allwine — 2025-06-30

Diabetic retinopathy (DR) is a pressing global health issue that affects the retina and is closely linked to diabetes, leading to vision impairment and blindness, particularly in adults. With the rising incidence of diabetes, the need for efficient and accurate DR screening is critical for early intervention and improved patient outcomes. Automated screening solutions can streamline this process, allowing healthcare professionals to focus more on patient care.In this study, we harnessed advanced deep learning techniques, specifically 3D convolutional neural networks (3D-CNNs), to classify DR into binary categories (presence or absence) and five multiclass categories: mild, moderate, no DR, proliferative DR, and severe DR. Our goal was to enhance diagnostic Precision in ophthalmology. To optimize our models, We embraced two methods transformative data augmentation: random shifting and random weak Gaussian blurring, empowering our model to reach new heights,as well as their combination. Our results showed that, for binary classification, the combined augmentation achieved significant success, The multiclass model was trained without any data augmentation excelled in Precision. These findings highlight the importance of large, high-quality research datas in deep learning algorithms. By leveraging advanced methodologies and robust data, we can transform diabetic retinopathy screening, promoting earlier detection and better treatment outcomes for those affected.

Comparative Analysis of Supervised Learning Algorithms for Delivery Status Prediction in Big Data Supply Chain Management

Riri Damayanti Apnena — 2025-06-23

This study addresses the problem of predicting delivery status in supply chain data, a critical task for optimizing logistics and operations. The dataset, which includes multiple features like order details, product specifications, and customer information, was pre-processed using oversampling to address class imbalance, ensuring that the model could handle rare cases of late or canceled deliveries. The data cleaning process involved handling missing values, removing irrelevant columns, and transforming categorical variables into numerical formats. After pre-processing and cleaning, five machine learning models were applied: Logistic Regression, Random Forest, SVM, K-Nearest Neighbors (KNN), and XGBoost. Each model was evaluated using metrics such as accuracy, precision, recall, and F1-score. The results showed that XGBoost outperformed the other models, achieving the highest accuracy and providing the most reliable predictions for the delivery status. This makes XGBoost the best choice for supply chain data analysis in this context. This study contributes to the growing application of machine learning in supply chain optimization by identifying XGBoost as a robust model for delivery status prediction in large datasets. For future research, exploring hybrid models and advanced feature engineering techniques could further improve prediction accuracy and address additional challenges in supply chain optimization, especially in the context of real-time data processing and dynamic supply chain environments.

Implementation of Enhanced Confix Stripping Stemming and Chi-Squared Feature Selection on Classification UIN Walisongo Website with Naïve Bayes Classifier

Muhammad Naufal Muhadzib Al-Faruq — 2025-06-23

Academic news classification on university websites remains a challenge due to the growing volume of content and lack of efficient categorization systems. At UIN Walisongo Semarang, this problem hinders students, faculty, and the public from easily accessing relevant information. This study aims to develop an automated academic news classification system to address this issue. We applied a Naïve Bayes Classifier model, enhanced with Term Frequency weighting, the Enhanced Confix Stripping Stemmer for Indonesian language preprocessing, and Chi-Squared feature selection to identify the most informative terms. The dataset consisted of 880 academic news articles from UIN Walisongo’s website, split into 704 training and 176 testing documents. The system achieved 95% accuracy on the test set. To evaluate generalizability, we used a separate evaluation set of 12 new articles, obtaining 83.3% accuracy. The preprocessing stage played a vital role in reducing morphological complexity, while Chi-Squared scoring improved the relevance of selected features. This research highlights the importance of robust text classification techniques in academic information systems, particularly in Indonesian language contexts where language morphology poses unique challenges. The proposed model demonstrates strong performance, scalability, and potential for integration into academic portals to improve information retrieval. This study contributes significantly to the field of Natural Language Processing and applied machine learning in academic settings, especially for Indonesian-language content. It provides an effective solution for automated academic content management in institutional information systems.

Improving Term Deposit Customer Prediction Using Support Vector Machine with SMOTE and Hyperparameter Tuning in Bank Marketing Campaigns

Dodo Zaenal Abidin — 2025-06-23

Identifying potential customers for term deposit products remains a challenge in the banking industry due to class imbalance in marketing datasets. This study proposes an integrated approach that combines Support Vector Machine (SVM) with the Synthetic Minority Oversampling Technique (SMOTE) and hyperparameter tuning via GridSearchCV to enhance prediction performance. The dataset comprises 45,211 records containing demographic and campaign-related features. Preprocessing steps include categorical encoding, feature scaling, and SMOTE-based resampling. The optimized SVM model achieves an accuracy of 91% and an AUC of 0.96, outperforming the baseline model and demonstrating strong discriminatory ability, particularly for the minority class. This method improves the balance between precision and recall while reducing bias toward the majority class. The findings confirm the effectiveness of combining SMOTE and SVM for imbalanced classification tasks in the financial domain. These results contribute to the advancement of applied machine learning in informatics, particularly in developing robust decision support systems for data-driven banking strategies. Future work may extend this approach to diverse datasets and explore advanced resampling or ensemble techniques to improve model generalization.

Leveraging Convolutional Block Attention Module (Cbam) For Enhanced Performance In Mobilenetv3-Based Skin Cancer Classification

Anas Rachmadi Priambodo — 2025-06-23

As the incidence of skin cancer continues to rise globally, effective automated classification methods become crucial for early detection and timely intervention. Lightweight neural networks such as MobileNetV3 offer promising solutions due to their minimal parameters, making them suitable for environment with low resource. This study aims to develop an automated multiclass skin cancer classification system by enhancing MobileNetV3 with the Convolutional Block Attention Module (CBAM). The primary goal is to achieve high classification accuracy without significantly increasing computational demands. We employed Bayesian optimization to automatically fine-tune model parameters and applied targeted data augmentation techniques to address class imbalance. CBAM was integrated to highlight diagnostically relevant regions within images. The proposed method was evaluated using the ISIC 2024 SLICE-3D dataset, which includes over 400,000 dermatoscopic images categorized into benign, basal cell carcinoma, melanoma, and squamous cell carcinoma classes. Preprocessing involved standardized resizing, normalization, and extensive geometric and photometric augmentations. Results demonstrated that our method achieved an accuracy of 98.97%, precision of 98.99%, recall of 98.97%, and an F1-score of 98.98%, surpassing previous state-of-the-art models by 1.86–6.52%. Remarkably, this improvement was achieved with minimal additional parameters due to the effective integration of CBAM. These results represent an advancement in automated medical image analysis, particularly for low resource settings, by combining lightweight CNNs with attention mechanisms and systematic hyperparameter exploration.

Real-Time Rice Leaf Disease Diagnosis: A Mobile CNN Application with Firebase Integration

Abdul Azis — 2025-06-23

Rice, the staple food for the majority of Indonesia's population, faces significant production threats from leaf diseases, which can decrease yields and jeopardize national food security. Traditional manual identification of these diseases is a major challenge for farmers, as it is often subjective, prone to misdiagnosis leading to incorrect treatments, time-consuming, demands specialized expertise, and is difficult to implement widely for effective real-time early prevention, allowing diseases to spread and significantly impact crop yields. This research addresses these challenges by developing an automated and easily accessible rice leaf disease diagnosis system. The system is manifested as a mobile application that integrates a Convolutional Neural Network (CNN) model, specifically utilizing the EfficientNetB0 architecture, for the classification of rice leaf images and leverages key Firebase services such as its Realtime Database for data synchronization and Cloud Storage for image management to ensure a scalable and responsive backend. The methodology involved several key stages. Firstly, the CNN model was developed by employing a transfer learning approach on the pre-trained EfficientNetB0 architecture. Secondly, the model underwent comprehensive testing using a dataset of 1,000 new rice leaf images, which were independently validated by agricultural experts. The results demonstrated that the developed CNN model achieved a global accuracy of 85.9%, with an average precision of 86.1% and recall of 85.9% (macro-average) in the expert validation testing phase with the 1,000 new images. However, the study also identified variations in the model's performance across different disease classes, highlighting areas that require further optimization to enhance detection effectiveness for specific types of rice leaf diseases. The primary benefit of this research is the provision of a practical rice leaf disease diagnosis tool that is readily accessible to farmers via a mobile application, empowering them with timely and accurate information for effective crop management. This can lead to reduced crop losses, improved yield quality, and contribute significantly to national food security. Furthermore, this research contributes to the field of applied machine learning and mobile computing in resource-constrained agricultural environments, offering valuable insights for the development of impactful informatics solutions.

Depression Detection using Convolutional Neural Networks and Bidirectional Long Short-Term Memory with BERT variations and FastText Methods

Leonardus Adi Widjayanto — 2025-06-30

Depression has become a significant public health concern in Indonesia, with many individuals expressing mental distress through social media platforms like Twitter. As mental health issues like depression are increasingly prevalent in the digital age, social media provides a valuable avenue for automated detection via text, though obstacles such as informal language, vagueness, and contextual complexity in social media complicate precise identification. This study aims to develop an effective depression detection model using Indonesian tweets by combining Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM). The dataset consisted of 58,115 tweets, labeled into depressed and non-depressed categories. The data were preprocessed, followed by feature extraction using BERT and feature expansion using FastText. The FastText model was trained on three corpora: Tweet, IndoNews, and combined Tweet+IndoNews corpus; the total corpus will be 169,564 entries. The best result was achieved by BiLSTM model with 84.67% accuracy, a 1.94% increase from the baseline, and the second best was the BiLSTM-CNN hybrid model achieved 84.61 with an accuracy increase of 1.7% from the baseline. These result indicate that combining semantic feature expansion with deep learning architecture effectively improves the accuracy of depression detection on social media platforms. These insights highlight the importance of integrating semantic enrichment and contextual modeling to advance automated mental health diagnostics in Indonesian digital ecosystems.

Comparative Analysis Of Ant Lion Optimization And Jaya Algorithm For Feature Selection In K-Nearest Neighbor (Knn) Based Electricity Consumption Prediction

Retno Wahyusari — 2025-06-23

The increase in demand for electrical energy is in line with increasing population, urbanization, industrial deployment, and technology. Accurate prediction of electrical energy consumption plays an important role in planning, analyzing, and managing electricity systems to ensure sustainable, safe, and economical electricity supply. K-Nearest Neighbors (KNN) is a simple and fast prediction algorithm based on the quality and relevance of the features used. This research proposes to improve the accuracy of energy consumption prediction through feature selection based on metaheuristic algorithms, namely Genetic Algorithm (GA), Ant Lion Optimization (ALO), Teaching Learning Based Optimization (TLBO), and Jaya Algorithm (JA). The dataset used is Tetouan City Power Consumption, with a preprocessing process of time feature extraction, min-max scaling normalization, and feature selection. The ALO+KNN and JA+KNN combinations delivered the best and most stable prediction performance, while TLBO+KNN performed poorly. GA+KNN showed the worst overall results among all combinations. The evaluation of model performance was based on RMSE, MAPE, and R² metrics. These findings highlight the importance of selecting a feature selection algorithm that aligns well with the characteristics of the model and dataset to enhance prediction accuracy.

Multi-Class Mangrove Classification Using Transfer Learning with MobileNet-V3 on Multi-Organ Images

Ari Sudrajat — 2025-07-09

Mangrove ecosystems are important for coastal protection, biodiversity conservation, and climate change mitigation. However, the accurate identification of mangrove species is very challenging due to the morphological similarities between different species, especially when the species are analyzed based on limited plant organs like leaves or stems. Manual identification methods have traditionally been time-consuming, error-prone, and require expert knowledge. Addressing these issues, this research suggests an automatic classification system based on Deep Learning techniques by leveraging the MobileNet-V3 architecture. The system is based on images of three different plant organs—leaves, stems, and seeds—of five mangrove species: Avicennia marina, Avicennia officinalis, Avicennia rumphiana, Rhizophora mucronata, and Sonneratia alba. Data augmentation techniques such as rotation, shifting, and flipping, as well as sharpness enhancement, were applied in the preprocessing step to enhance data variability and ease model generalization. The model was trained with a carefully selected set of hyperparameters and extensively validated through training and testing steps. The experiment results demonstrated outstanding performance with a training accuracy of 99.88% and perfect precision, recall, and F1-score values of 100%. Furthermore, testing with unseen data confirmed the robustness of the model since all test samples were correctly identified. This research concludes that the MobileNet-V3 architecture offers an effective approach to mangrove species classification and suggests that future work should involve larger and more varied datasets, real-world field environments, and the investigation of ensemble models to further extend the adaptability and scalability of mangrove monitoring systems.

Detecting Avocado Freshness In Real-Time: A Yolo-Based Deep Learning Approach

Atika Dwi Febriani — 2025-06-23

The increasing consumption of avocados in Indonesia highlights the need for an effective method to ensure fruit freshness. The main problem lies in the absence of an objective and standardized system for assessing avocado freshness, which may lead to consumer dissatisfaction and food waste. This study aims to address the challenge of identifying avocado freshness to ensure suitability for consumption. Conducted from May 23 to June 5, 2024, the research used butter avocado samples sourced from supermarkets. The method employed is the You Only Look Once version 8 (YOLOv8) deep learning algorithm, known for its real-time object detection capabilities. YOLOv8 offers enhanced performance compared to earlier versions through anchor-free detection, improved speed, and accuracy, making it suitable for fast and reliable freshness detection tasks. Avocados were classified based on estimated spoilage time under room and refrigerator temperatures, ranging from "up to 5 days at room temperature and 14 days in refrigeration" to "not fit for consumption." The model was validated using 120 images categorized into six freshness levels. Evaluation results demonstrated high performance, with 98% accuracy, an F1-Score of 0.978, mAP50 of 0.994, and mAP50-95 of 0.972 after 50 training epochs, confirming the model’s robustness. Real-time tests yielded confidence levels of 96% and 94%, further validating its effectiveness in detecting avocado freshness. To facilitate daily use, a mobile application named Avo Freshify was developed. The app accurately identifies the freshness of avocados and provides valuable information for consumers and sellers. This research contributes to the advancement of artificial intelligence and object detection in food quality control and agricultural technology.

Evaluation of K-Means, DBSCAN, and Hierarchical Clustering for Strategic Segmentation of Tourism SMEs in Rembang, Indonesia

Ardiansyah Ramadhan — 2025-07-09

Small and Medium Enterprises (SMEs) play a crucial role in job creation, regional competitiveness, and economic equity. In the tourism sector, particularly in ecotourism and cultural tourism, clustering SMEs presents challenges due to complex and interrelated data variables. This study aims to evaluate the effectiveness of three clustering algorithms—K-Means, DBSCAN, and Hierarchical Clustering—in segmenting SMEs based on real-world tourism datasets. A purposive sampling method was applied to 203 valid respondents from SMEs in Rembang Regency, Central Java. Clustering performance was assessed using the Silhouette Coefficient and Davies-Bouldin Index, while computational efficiency and scalability were analyzed through execution time and memory usage. The results show that DBSCAN achieved the best clustering quality (Silhouette Coefficient: 0.5496, Davies-Bouldin Index: 0.3298), effectively managing noise and irregular cluster shapes. Hierarchical clustering offered moderate quality and helped reveal relationships between SMEs. In contrast, K-Means demonstrated the lowest quality (Silhouette Coefficient: 0.2321) due to its limitation in handling non-spherical clusters. For computational efficiency, Hierarchical Clustering required the least memory (0.14 MB) and shortest execution time (5.73 seconds), while K-Means took the longest time (26.00 seconds). DBSCAN consumed more memory due to density-based processing. K-Means was the most stable in scalability testing with increasing dataset sizes, whereas Hierarchical Clustering showed inefficiency. The findings support selecting appropriate clustering methods based on data complexity and size. This study enhances data-driven tourism development strategies and advances clustering methodology for applied informatics. Future work may explore hybrid clustering and predictive models for deeper insights.

Improving Semantic Segmentation of Flood Areas Using Rotation and Flipping-Based Feature Augmentation

Naili Suri Intizhami — 2025-07-09

Semantic segmentation is one of the powerful methods for analyzing flood video or picture data captured by smartphones. However, achieving accurate semantic segmentation requires the application of several methods. In this work, we address the task of feature augmentation approach using rotation (90°, 180°, 270°) and flipping (horizontal, vertical) to improve semantic segmentation of flood areas in Parepare city using a Fully Convolutional Network (FCN). The experimental results demonstrate that the best augmentation scenario 270° rotation achieved an accuracy of 88% and 90° rotation achieved an mean Intersection over Union (mIoU) of 43%, significantly outperforming the baseline FCN model without augmentation, which achieved 86% accuracy and 35% mIoU.

Prediction Of Clay Mining Production Value Using Linear Regression Model With Multi-Swarm Particle Swarm Optimization

Gusti Eka Yuliastuti — 2025-06-10

The progress of a nation or a country can be recognized from its income through various industries inside. Mining refers to one of the most advanced industries in Indonesia. The majority of mining in Indonesia is open-pit mining which is exposed directly to the sky. This study focuses on modeling data from rainfall, working hours, and production yields. It employed the Multi-Swarm Particle Swarm Optimization (MSPSO) algorithm to find multiple linear regression modeling by minimizing the Mean Squared Error (MSE) value. The value for the production results was then predicted using the existing multiple linear regression model. In terms of testing, the best model having an MSE of 288.0656 occurred at the parameters of Npop 180, acceleration coefficient 1 by 0.7, acceleration coefficient 2 by 0.7, acceleration coefficient 3 by 0.7, wmin 4, wmax 9 within 100 iterations.

Comparative Analysis Retrofit and Ktor Client Performance in Various Internet Speeds Internet on MSMEs Cashier Application

Muhamad Akbar Abdul Kholik — 2025-06-10

MSMEs (Micro, Small, and Medium Enterprises) in Indonesia face uneven network infrastructure, with more than 20% of smartphone users having download speeds below 10 Mbps. This condition hampers the efficiency of data processing between client and server, while MSMEs need innovations such as digitization of bookkeeping to increase competitiveness. The selection of HTTP networking libraries such as Retrofit and Ktor Client is very important, because both play a role in the process of sending and receiving data from the server. This research aims to analyze the performance of both libraries in the Lulu POS application to determine the most optimal library in supporting MSME operations in various network conditions. The test is conducted in two scenarios: the first scenario uses text data and the second scenario uses text and image data. Each scenario has several test cases that will be tested at six different internet speeds. The results show that Retrofit excels in response time for text data with a performance improvement of 18.85% and network usage of 21.33%. Ktor Client is superior in scenarios involving text and image data, with a response time advantage of 7.20% and network usage of 0.08%. On the other hand, Retrofit is more efficient in memory usage in both scenarios, with an advantage of 16.49% in text data and 4.70% in text and image data. In conclusion, Retrofit is more stable for applications focusing on text data such as Lulu POS, while Ktor Client is more suitable for applications that manage images. These results make MSMEs get cashier applications with optimal libraries for various network conditions, so that operations are smoother and data management efficiency increases.

Optimising Bitcoin Price Forecasting Using Lstm, Gru, Prophet, Var, And Es Multi-Model Approaches

Anggito Karta Wijaya — 2025-06-10

This study aims to optimize Bitcoin price forecasting by integrating several multi-model approaches, including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Prophet, as well as risk analysis using Value at Risk (VaR) and Expected Shortfall (ES). The daily Bitcoin price data from the period of July 17, 2010, to June 28, 2024, obtained from Kaggle, were analyzed using accuracy metrics such as Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE), as they provide a more objective and reliable evaluation of prediction effectiveness. The results show that the LSTM model performed the best, with an MSE of 535,419.12, RMSE of 731.72, MAE of 310.72, and MAPE of 159.01. The GRU model produced similar evaluation values with an MSE of 558,868.06 and RMSE of 747.57. In contrast, Prophet demonstrated lower performance, with an MSE of 59,309,927.76 and RMSE of 7,701.29. The risk analysis indicated that at a 95% confidence level, VaR reached 61,676.43, while ES reached 61,737.58, reflecting additional risk in extreme conditions. This study provides valuable insights into the advantages of the LSTM and GRU models for Bitcoin price forecasting, while also emphasizing the importance of risk analysis in supporting cryptocurrency investment decisions.

Enhancing Cyberbullying Detection with a CNN-GRU Hybrid Model, Word2Vec, and Attention Mechanism

Kaysa Azzahra Adriana — 2025-06-10

Cyberbullying is an act of violence commonly committed on online platforms such as social media X, often causing psychological effects for victims. Despite prevention efforts, traditional methods for detecting cyberbullying show limited effectiveness due to the complexity of language and diversity of expressions, leading to suboptimal performance. This study aims to enhance detection accuracy by applying Convolutional Neural Networks (CNN) and Gated Recurrent Unit (GRU) with an attention mechanism to analyze textual data from tweets. The model uses Term Frequency-Inverse Document Frequency (TF-IDF) for extracting important words and Word2Vec for expanding text representation. A total of 30,084 labeled datasets from tweets on social media X were utilized. Results indicate the hybrid CNN-GRU model with attention achieved the highest accuracy of 80.96%, outperforming stand-alone CNN and GRU models. Additionally, TF-IDF and Word2Vec significantly improved model performance, with the CNN-GRU combination proving most effective for detecting cyberbullying. This study contributes to computer science by proposing a novel approach that integrates CNN, GRU, and attention mechanisms with advanced feature extraction techniques, providing a more reliable detection system for online platforms. It also highlights the potential for integrating multimodal data to further enhance future performance.

Comparative Analysis of Augmentation and Filtering Methods in VGG19 and DenseNet121 for Breast Cancer Classification

I Kadek Seneng — 2025-06-10

Breast cancer is one of the most prevalent malignancies and a leading cause of mortality among women worldwide. Mammography plays a crucial role in early detection, yet challenges in manual interpretation have led to the adoption of Convolutional Neural Networks (CNNs) to improve classification accuracy. This study evaluates the performance of Visual Geometry Group (VGG19) and Densely Connected Convolutional Networks (DenseNet121) in mammogram classification. It examines the impact of data augmentation and image enhancement techniques, including Contrast-Limited Adaptive Histogram Equalization (CLAHE), Median Filtering, and Discrete Wavelet Transform (DWT), as well as the influence of varying epochs and learning rates. A novel approach is introduced by assessing data augmentation effectiveness and exploring model adaptations, such as layer incorporation and freezing during training. Classification performance is enhanced through fine-tuning strategies combined with image enhancement techniques, reducing reliance on data augmentation. These findings contribute to medical imaging and computer science by demonstrating how CNN modifications and enhancement methods improve mammogram classification, providing insights for developing robust deep learning-based diagnostic models. The highest performance was achieved using VGG19 with DWT, a learning rate of 0.0001, and 20 epochs, yielding 98.04% accuracy, 98.11% precision, 98% recall, and a 97.99% F1-score. Data augmentation did not consistently enhance results, particularly in clean datasets. Increasing epochs from 10 to 20 improved accuracy, but performance declined at 30 epochs. The confusion matrix showed high accuracy for Benign (100%) and Cancer (99.5%), with more misclassifications in the Normal class (94.5%).

An Efficient Model for Waste Image Classification Using EfficientNet-B0

Teofilus Kurniawan — 2025-06-10

Waste management remains a significant challenge, particularly in developing countries. To address this issue, artificial intelligence can be leveraged to develop a waste image classifier that facilitates automatic waste sorting. Previous studies have explored the use of Convolutional Neural Networks (CNNs) for waste image classification. However, CNNs typically require a large number of parameters, leading to increased computational time. For practical applications, a waste image classifier must not only achieve high accuracy but also operate efficiently. Therefore, this study aims to develop an accurate and computationally efficient waste image classification model using EfficientNet-B0. EfficientNet-B0 is a CNN architecture designed to achieve high accuracy while maintaining an efficient number of parameters. This study utilized the publicly available TrashNet dataset and investigated the impact of image augmentation in addressing imbalance data issues. The highest performance was achieved by the model trained on the unbalanced dataset with the addition of a Dense(32) layer, a dropout rate of 0.3, and a learning rate of 1e-4. This configuration achieved an accuracy of 0.885 and an F1-score of 0.87. These results indicate that the inclusion of a Dense(32) layer prior to the output layer consistently improves model performance, whereas image augmentation does not yield a significant enhancement. Furthermore, our proposed model achieved the highest accuracy while maintaining a significantly lower number of parameters compared to other CNN architectures with comparable accuracy, such as ResNet-50 and Xception. The resulting waste classification model can then be further implemented to build an automatic waste sorter.

Machine Learning Models for Metabolic Syndrome Identification with Explainable AI

Egga Asoka — 2025-06-10

Metabolic syndrome (MetS) is a cluster of interrelated risk factors, including hypertension, dyslipidemia, central obesity, and insulin resistance, significantly increasing the likelihood of cardiovascular diseases and type 2 diabetes. Early identification of hypertension, a key component of MetS, is essential for timely intervention and effective disease management. This research aims to develop a hybrid machine learning model that integrates XGBoost classification with K-Means clustering to enhance or strengthening of hypertension prediction and identify distinct patient subgroups based on metabolic risk factors. The dataset consists of 1,878 patient records with metabolic parameters such as systolic and diastolic blood pressure, fasting glucose, cholesterol levels, and anthropometric measurements. Model performance was assessed using accuracy, precision, recall, F1-score, and ROC-AUC. The proposed XGBoost model achieved an outstanding classification performance with 98% accuracy, 98% precision, 98% recall, 98% F1-score, and an ROC-AUC of 1.00. K-Means clustering further identified five distinct patient subgroups with varying metabolic risk profiles. The findings underscore the potential of machine learning-driven decision support systems in improving hypertension diagnosis and MetS management.

Optimizing Indonesian Banking Stock Predictions with DBSCAN and LSTM

Septiannisa Alya Shinta Purwandhani — 2025-06-10

Investing in the stock market is challenged by high volatility, which often leads to inaccurate price predictions. Prediction models often struggle to handle the fluctuation phenomenon and produce unstable forecasts. This study aims to predict stock prices in three banks, namely PT Bank Central Asia Tbk (BBCA), PT Bank Rakyat Indonesia (Persero) Tbk (BBRI), and PT Bank Mandiri (Persero) Tbk (BMRI) using Long Short-Term Memory (LSTM) with the integration of Density-Based Spatial Clustering of Applications with Noise (DBSCAN) for anomaly detection. DBSCAN is applied with an epsilon (ε) of 0.5 and a minimum of 5 samples using Euclidean distance. The LSTM model consists of two hidden layers with 50 units, optimized using Adam, and applying the Mean Squared Error (MSE) loss function. The results show that DBSCAN improves prediction accuracy under several conditions. For BBCA stock, the lowest MSE was 0.003 at the 2nd fold with DBSCAN compared to 0.006 without DBSCAN. For BMRI stock achieved an MSE of 0.003 at the 4th fold with DBSCAN, while the 5th fold without DBSCAN obtained 0.000. For BBRI stock showed the best MSE of 0.003 at the 2nd fold with DBSCAN and the 5th fold without DBSCAN. These results show that the integration of DBSCAN can improve prediction especially when extreme price fluctuations occur. This research contributes to the development of stock price prediction methods that can be one of the benchmarks for investors before making decisions so that they do not experience losses.

Deep Reinforcement Learning for Autonomous System Optimization in Indonesia: A Systematic Literature Review

Dedi Yusuf — 2025-06-10

Background: The development of artificial intelligence (AI) technology, including Deep Reinforcement Learning (DRL), has brought significant changes in various industrial sectors, especially in autonomous systems. DRL combines the capabilities of Deep Learning (DL) in processing complex data with those of Reinforcement Learning (RL) in making adaptive decisions through interaction with the environment. However, the application of DRL in autonomous systems still faces several challenges, such as training stability, model generalization, and high data and computing resource requirements. Methods: This study uses the Systematic Literature Review (SLR) method to identify, evaluate, and analyze the latest developments in DRL for autonomous system optimization. The SLR was conducted by following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework, which consists of four main stages: identification, screening, eligibility, and inclusion of research articles. Data were collected through literature searches in leading scientific journal databases such as IEEE Xplore, MDPI, ACM Digital Library, ScienceDirect (Elsevier), SpringerLink, arXiv, Scopus, and Web of Science. Results: This study found that DRL has been widely adopted in various industrial sectors, including transportation, industrial robotics, and traffic management. The integration of DRL with other technologies such as Computer Vision, IoT, and Edge Computing further enhances its capability to handle uncertain and dynamic environments. Therefore, this study is crucial in providing a comprehensive understanding of the potential, challenges, and future directions of DRL development in autonomous systems, in order to foster more adaptive, efficient, and reliable technological innovations.

Digital Forensic Chatbot Using DeepSeek LLM and NER for Automated Electronic Evidence Investigation

Nuurun Najmi Qonita — 2025-06-10

The growing complexity of cybercrime necessitates efficient and accurate digital forensic tools for analyzing electronic evidence. This research presents an intelligent digital forensic chatbot powered by DeepSeek Large Language Model (LLM) and Named Entity Recognition (NER), designed to automate the analysis of various digital evidence, including system logs, emails, and image metadata. The chatbot is deployed on the Telegram platform, providing real-time interaction with investigators. The metric results show that the chatbot achieves a precision of 83.52%, a recall of 88.03%, and an F1-score of 85.71%. These results demonstrate the chatbot's effectiveness in accurately detecting forensic entities, significantly improving investigation efficiency. This study contributes to digital forensics by integrating LLM and NER for enhanced evidence analysis, offering a scalable and adaptive solution for automated cybercrime investigations. Future research may explore integrating anomaly detection and blockchain-based evidence integrity.

Hybrid Neural Network-Based Road Damage Detection Using CNN-RNN and CNN-MLP Models

Ani Dijah Rahajoe — 2025-06-10

Currently, there are many applications of image processing in various fields. One of them is the recognition of paved road images. Detection through images helps in handling infrastructure development roads. With the advancement of technology, especially in the field of deep learning, the process of detecting road damage can be done automatically and more efficiently. The road damage detection system can be integrated into the smart city system to monitor infrastructure conditions in real time. This study will use a combined deep learning algorithm between Convolutional Neural Network- Recurrent Neural Network (CNN-RNN) and as a comparison using Convolutional Neural Network- MultiLayer Perceptrons (CNN-MLP). The study aims to analyze the accuracy of using the CNN-RNN and CNN-MLP algorithms for detecting paved roads that have categories of undamaged roads, damaged roads, and damaged roads with holes. The detection of paved roads has complex details so an algorithm that has good performance with high accuracy is needed. The results of the study showed that the CNN-RNN hybrid had a better accuracy of 96.59 percent than the CNN-MLP hybrid model of 95.9 percent.

Comparative Analysis of DBSCAN, OPTICS, and Agglomerative Clustering Methods for Identifying Disease Distribution Patterns in Banjarnegara Community Health Centers

Dillyana Tugas Setiyawan — 2025-06-10

The variation in disease distribution patterns across community health centers in Banjarnegara Regency necessitates a precise segmentation analysis to support effective allocation of healthcare resources. This study aims to compare the effectiveness of three clustering methods DBSCAN, OPTICS, and Agglomerative Clustering in grouping Puskesmas based on the type and number of diseases they manage. The evaluation methods used include the Silhouette Score and the Davies-Bouldin Index, which assess the quality of the clustering results. The analysis indicates that Agglomerative Clustering produces the most stable cluster structures, reflected in its highest Silhouette Score, compared to DBSCAN and OPTICS, which tend to yield more noise and less optimal clustering quality. These findings suggest that hierarchical clustering approaches are more effective in the context of healthcare service distribution data at the primary care level. The results of this study are expected to serve as a foundation for the formulation of data-driven and region-based health policies, particularly in designing more targeted interventions and optimizing the distribution of healthcare services.

Comparative Analysis of Hybrid Intelligent Algorithms for Microsleep Detection and Prevention

Arvina Rizqi Nurul'aini — 2025-06-10

Microsleep is a critical factor contributing to traffic accidents, posing significant risks to road safety. Research by the AAA Foundation for Traffic Safety found that 328,000 sleep-related driving accidents happen annually in the United States, underscoring the widespread and dangerous nature of drowsy driving. These incidents often occur without warning, making them especially hazardous and difficult to prevent through conventional means alone. This research aims to improve the accuracy of microsleep detection by developing a hybrid intelligent algorithms. It compares three intelligent algorithms: Fuzzy Logic (FL), representing scheme A; Fuzzy Logic combined with Artificial Neural Networks (FL-ANN), representing scheme B; and a combination of Fuzzy Logic, ANN, and Decision Trees (FL-ANN-DT), representing scheme C. These methods were evaluated using performance metrics such as MSE, MAE, RMSE, R², and response time. The results indicate that Scheme C (FL-ANN-DT) significantly outperforms the other approaches, achieving an MSE of 5.3617e-32, MAE of 4.3823e-17, R² of 1.0, and an RMSE close to zero, demonstrating near-perfect accuracy. Compared to previous models, this hybrid approach enhances prediction precision while maintaining real-time feasibility. The findings highlight the potential of FL-ANN-DT as an advanced microsleep detection system, contributing to improved road safety and real-time monitoring applications. This system can serve as a proactive safety layer in driver assistance technologies, reducing the risk of fatigue-related accidents and potentially saving lives.