Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi)

Enhancing Areca Nut Detection and Classification Using Faster R-CNN: Addressing Dataset Limitations with Haar-like Features, Integral Image, and Anchor Box Optimization

Yovi Pratama, Errissya Rasywir, Suyanti, Agus Siswanto, Fachruddin — Wed, 25 Jun 2025 00:00:00 +0000

The classification and detection of areca nuts are essential for agriculture and food processing to ensure product quality and efficiency. The manual classification of areca nuts is time-consuming and prone to human error. For a more accurate and efficient automated approach, a deep learning-based framework was proposed to address these challenges. This study optimizes the Faster R-CNN by integrating Haar-like features and integral images to enhance object detection. However, dataset limitations, including low image quality, inconsistent lighting, cluttered backgrounds, and annotation inaccuracies, affect the model performance. In addition, the small dataset size and class imbalance hindered generalization. The Faster R-CNN model was trained with and without Haar-like Features and Integral Image enhancement. Performance was evaluated based on training loss, accuracy, precision, recall, F1-score, and mean average precision (mAP). The effects of the dataset limitations on detection performance were also analyzed. The optimized model achieved better stability, with a final training loss of 0.2201, compared to 0.1101 in the baseline model. Accuracy improved from 62.60% to 73.60%, precision from 0.6161 to 0.7261, recall from 0.3094 to 0.4194, F1-score from 0.2307 to 0.3407, and mAP from 0.1168 to 0.2268. Despite these improvements, dataset constraints remain a limiting factor. While the integration of Haar-like features and integral images into faster R-CNN contributes to detection accuracy, the study also reveals that high-resolution images, precise annotations, and dataset scale significantly amplify model performance.

Automatic Classification of Multilanguage Scientific Papers to the Sustainable Development Goals Using Transfer Learning

Mon, 23 Jun 2025 00:00:00 +0000

The classification of scientific papers according to their relevance to Sustainable Development Goals (SDGs) is a critical task in identifying the research development status of goals. However, with the growing volume of scientific literature published worldwide in multiple languages, manual categorization of these papers has become increasingly complex and time-consuming. Furthermore, the need for a comprehensive multilingual dataset to train effective models complicates the task, as obtaining such datasets for various languages is resource intensive. This study proposes a solution to this problem by leveraging transfer learning techniques to automatically classify scientific papers into SDG labels. By fine-tuning pretrained multilingual models mBERT on SDG publication datasets in a multilabel approach, we demonstrate that transfer learning can significantly improve classification performance, even with limited labelled data, compared to SVM. Our approach enables the effective processing of scientific papers in different languages and facilitates the seamless mapping of research to the relevance of SDGs, the four pillars of SDGs, and the 17 goals of SDGs. The proposed method addresses the scalability issue in SDG classification and lays the groundwork for more efficient systems that can handle the multilingual nature of modern scientific publications.

A New Framework for Dynamic Educational Marketing Segmentation in Student Recruitment: Optimizing Fuzzy C-Means with Metaheuristic Techniques

Sun, 22 Jun 2025 00:00:00 +0000

An effective educational marketing strategy requires accurate school segmentation to enhance new student recruitment. Traditional segmentation methods such as K-means are often used, but they have limitations in capturing the flexibility of school characteristics. Fuzzy C-Means (FCM) offers a more adaptive approach by allowing each school to simultaneously have a degree of membership in several clusters. However, the performance of FCM highly depends on determining parameters such as the number of clusters (k) and the level of fuzziness (m), which are not always optimal when determined manually. This study develops a new framework for dynamic educational marketing segmentation in student recruitment by optimizing FCM using three metaheuristic techniques: Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Differential Evolution (DE). Performance was evaluated using the Fuzzy Silhouette Index (FSI). The experimental results showed that DE yielded the best results with the highest FSI value (0.8023), producing eight main clusters based on the Recency, Frequency, and Monetary (RFM) model. Based on the clustering results, a personalized and adaptive marketing strategy was designed to enhance the effectiveness of student recruitment. The proposed framework enhances segmentation accuracy and supports the implementation of dynamic data-driven marketing in the context of higher education. This study also opens new directions for educational data mining research and machine-learning-based marketing strategies.

Enhancing Stroke Prediction with Logistic Regression and Support Vector Machine Using Oversampling Techniques

Sun, 22 Jun 2025 00:00:00 +0000

Stroke is a significant health concern that can result in both death and disability, making the early identification of risk factors crucial. Previous studies on stroke prediction have been limited by inadequate handling of class imbalance, lack of comprehensive feature selection, and parameter optimization, with accuracy rates usually below 80%. This study compares the performance of Logistic Regression (LR) and Support Vector Machine (SVM) algorithms combined with different oversampling methods—SMOTE, Borderline-SMOTE, ADASYN, Random Over Sampling (ROS), and Random Under Sampling (RUS)—on a stroke prediction dataset. Correlation-based feature selection identified age, hypertension, and heart disease as significant predictors. GridSearchCV with 10-fold cross-validation was used for hyperparameter optimization, and performance was evaluated using precision, recall, accuracy, and ROC curves. The results showed that SVM significantly outperformed Logistic Regression across all sampling methods. SVM+ROS achieved the highest performance with perfect recall (100%), precision of 97.18%, and accuracy of 98.56% (AUC: 0.9857), whereas SVM + Borderline-SMOTE offered balanced performance with a recall of 94.99%, precision of 95.06%, and accuracy of 95.17% (AUC: 0.9512). LR + Borderline-SMOTE performed the best with an accuracy of 84.98% (AUC: 0.8503), significantly better than previous studies. This improved accuracy shows significant clinical benefits, potentially reducing missed stroke diagnoses by identifying thousands of additional at-risk patients in large-scale screening programs. Healthcare providers should consider implementing SVM with ROS in critical care settings, where potentially missed stroke cases have severe consequences. Simultaneously, SVM with Borderline-SMOTE may be more appropriate for resource-constrained environments.

Stunting Prediction Modeling in Toddlers Using a Machine Learning Approach and Model Implementation for Mobile Application

Eko Abdul Goffar, Rosa Eliviani, Lili Ayu Wulandhari — Sun, 22 Jun 2025 00:00:00 +0000

Children’s health and development are critical for maintaining national productivity and independence, with stunting being a major concern. Stunting, a form of malnutrition, impairs growth and development, affecting millions of people globally, including a significant number in Indonesia. This study addresses the challenge of stunting by developing a predictive model using machine learning techniques to forecast stunting risks based on public health data. The literature review section discusses the factors that influence stunting, and these factors are used as features to build a stunting prediction model. Then the features were used to build a model with three machine learning algorithms Extreme Gradient Boosting (XGBoost), Random Forest, and K-Nearest Neighbor (KNN) to build and evaluate models that predict stunting. The models were trained and assessed using public datasets and the most effective algorithm was integrated into a mobile application for practical use. The results indicate that the XGBoost model outperforms the other models with an accuracy of 85%, making it the optimal choice for implementation in a mobile application. The next-best model is selected to be implemented through a mobile application so that users can directly use the model that has been built. This application aims to enhance early detection and intervention efforts for stunting, potentially improving child health outcomes and contributing to long-term productivity by building predictive models and implementing the models into a mobile application. This study contributes to the implementation of models built using public data for application in mobile applications.