Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi)
Not a member yet
1071 research outputs found
Sort by
Deep Learning-Based Soybean Leaf Disease Classification Using DenseNet121, Xception, and MobileNetV2
No full textThis study is driven by the challenge of soybean leaf diseases, which significantly reduce agricultural productivity and pose a threat to food security. To address this issue, we developed a deep learning–based classification model for soybean leaf disease detection, employing three prominent architectures: DenseNet121, Xception, and MobileNetV2. The dataset comprised 770 images representing six disease categories and one healthy category, which was expanded to 5,880 images using data augmentation techniques. The dataset was evaluated under three experimental scenarios with splits of 70% training, 10% validation, and 20% testing. Experimental results demonstrated that the DenseNet121 model, optimized with AdamW, achieved the highest accuracy at 90.14%, outperforming MobileNetV2 (85.48%) and Xception (65.37%). Moreover, DenseNet121 exhibited the most consistent performance in classifying the diverse categories of soybean leaf diseases
Prediction of Employee Recruitment Selection in Indonesia Pharmaceutical Company Using Backpropagation Networks
No full textPT K-24 Indonesia is one of the foremost companies in Indonesia, with a primary focus on distributing pharmaceutical products and healthcare services. During the last 2 years, PT K-24 has received more than 110,000 job applicants, with various position vacancies offered. The recruitment process begins with registration, online tests, and interviews. The need for manpower increases annually. More attention is required to select prospective employees who match the selection criteria. However, during the process, PT K-24 found that the recruitment process was less efficient because the applicants did not meet the company’s criteria. To overcome this problem, it is necessary to create a recommendation system for candidate selection. This study developed a recommendation using the multi-layer perceptron method, namely backpropagation. According to the prior literature, this method effectively reduces the error rate of prediction and recommendation results. This study also found that relevant data, the number of input parameters is not big enough, and the minimum network model can make better predictions with a considerable mean square error of 0.029. Our study contributes to the methodological approach by implementing real-world problems and measuring additional parameters that fit the selection requirement
Adaptive Stress Prediction with GSR, SMOTE Balancing, and Random Forest Models
No full textStress is a pervasive condition that affects mental health, productivity, and quality of life across populations. Traditional methods for stress assessment, such as the Perceived Stress Scale (PSS), rely on retrospective self-reporting and are limited by subjectivity and delayed feedback. To address this gap, this study developed an integrated real-time stress monitoring system combining Galvanic Skin Response (GSR) sensors, Internet of Things (IoT) technology, and machine learning algorithms. Primary GSR data were collected from 30 participants under varied conditions, supplemented by secondary data from the WESAD dataset. A Random Forest classifier was employed to categorize stress into four levels: normal, mild, moderate, and severe. To address class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was applied, leading to improved model robustness. The system achieved a cross-validated classification accuracy of 69%, with substantial improvements in the detection of moderate and severe stress cases compared to traditional threshold-based methods. A strong agreement (Cohen’s Kappa κ = 0.82) was observed between system predictions and PSS-based stress assessments. Feature importance analysis identified mean GSR value and Skin Conductance Response (SCR) amplitude as the most influential indicators of stress. The system was evaluated for usability, receiving high user ratings in terms of accessibility, simplicity, and interactivity. A simple Python-based command-line interface (CLI) was also developed for real-time stress prediction based on input features. This research demonstrates the feasibility and effectiveness of combining physiological sensing, predictive analytics, and user-friendly interfaces to enable scalable and adaptive stress monitoring. Future developments will focus on integrating additional physiological modalities and deep learning techniques to enhance predictive performance and personalization in clinical and everyday contexts
Optimization Techniques and Programming for Developing Cost-Effective and Balanced Diet Schedules for Preschoolers
No full textProper nutrition is important for the growth, motor and cognitive development of young children since the foods consumed determine how well-rounded a child's diet is. However, preschool menu planning is complex because it requires balancing multiple constraints such as cost, dietary guidelines, and food variety. This study introduces a computational approach to menu planning for preschools through Linear Programming (LP), Integer Programming (IP), and Binary Programming (BP). This study highlights algorithmic design, constraint modelling, and computational efficiency in solving optimization problems, rather than focusing primarily on dietary outcomes. The models were tested using Malaysian food database to evaluate both feasibility and efficiency. The findings indicate that all models successfully fulfilled the Recommended Nutrient Intakes (RNI 2017) for children aged 4 to 6, ensuring adequate levels of energy, protein, calcium, carbohydrates, and fat. In terms of cost, the LP model was the most economical at RM4.20 per day, but impractical due to fractional servings. The IP model produced a more realistic balance between cost and practicality at RM4.40 per day. The BP model generated the most diverse and implementable menus at RM5.00 per day, though at a higher cost. Overall, these optimization methods provide decision-support tools for enhancing the efficiency of preschool menu planning
Explainable DDoS Detection with a CNN-LSTM Hybrid Model and SHAP Interpretation
No full textThe rising frequency and complexity of Distributed Denial of Service (DDoS) attacks pose a severe threat to network security. This study aims to develop an effective and interpretable DDoS detection framework using a hybrid deep learning approach. The proposed method integrates Convolutional Neural Networks (CNN) to capture local traffic patterns and Long Short-Term Memory (LSTM) networks to model temporal dependencies. The CICIDS 2017 dataset, after preprocessing steps including data cleaning, standardization, and class balancing with SMOTE, was used to train and evaluate the model. Experimental results show that the framework achieved 99.98% accuracy and a 99.83% F1-Score, with minimal false positive and false negative rates. This study integrates SHAP to improve model interpretability, aligning feature importance with network security expertise. Future research will focus on real-time deployment, cross-dataset validation, and exploring alternative explainable AI techniques for improved scalability
Student-Generated User Story Quality: A Study on Practitioner and ChatGPT Evaluation
No full textEvaluating the quality of student-generated user stories is important in software engineering education, but only a limited number of industry practitioners can assist. The integration of generative AI can facilitate this process. To do so, the INVEST quality evaluation framework is widely recognized for assessing user story quality; however, prior research has not explored its use in conjunction with generative AI. This study investigated ChatGPT's ability to evaluate user stories using the INVEST framework. This study compares two ChatGPT-based evaluation approaches with those of experienced practitioners, focusing on student-generated user stories. Discrepancies between ChatGPT and practitioner evaluations were measured using Mean Absolute Deviation (MAD), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE). Statistical significance was tested using the Mann-Whitney U Test. The results indicate that ChatGPT’s 1st approach yielded lower discrepancies than practitioner evaluations. Moreover, significance testing showed no statistically significant differences between the ChatGPT and practitioner results for the two INVEST criteria- Independent and Estimable. These findings suggest that the 1st approach can assist in the evaluation process, although practitioners must ensure comprehensive and accurate evaluations. ChatGPT can provide preliminary evaluations in educational contexts, enabling students to receive formative feedback and allowing educators to streamline evaluation processes. Although practitioner validation is still required, their role may shift toward verifying AI-generated results, thus reducing the overall workload and accelerating quality evaluatio
Boosting YOLO11: Global Attention & Hyperparameter Tuning for High-Fidelity Military Aircraft Detection
No full textMilitary aircraft detection from aerial and satellite imagery is crucial for strategic surveillance and intelligence. This study evaluated the impact of the Global Attention Mechanism (GAM) and hyperparameter optimization on the performance of the YOLO11 model for military aircraft detection. Utilizing a traditional YOLO model as a baseline, we compared precision, recall, and mean Average Precision (mAP) metrics across various configurations. These configurations included the implementation of GAM and variations in n, s variant of YOLO11, optimizers (Adam, NAdam, RAdam, Adamax, AdamW, SGD) and learning rates (0.01, 0.001, 0.0001, 0.00417). Experimental results demonstrate that the integration of GAM significantly enhances the model's detection capabilities within 300 iterations, particularly when combined with the Adamax optimizer and a learning rate of 0.001. This specific configuration achieved the highest mAP performance of 98.5%, outperforming other setups. Further confusion matrix analysis confirmed high accuracy in classifying various aircraft types, while also highlighting some challenges in distinguishing certain classes. The primary contribution of this study is the empirical demonstration of improved military aircraft detection performance by YOLO11 through the utilization of four global attention mechanism modules and effective hyperparameter tuning. These findings offer valuable insights for developing more accurate and robust object detection systems for defense and security applications
Impact of Adaptive Synthetic on Naïve Bayes Accuracy in Imbalanced Anemia Detection Datasets
No full textThis research aims to analyze the impact of the Adaptive Synthetic (ADASYN) oversampling technique on the performance of the Naïve Bayes classification algorithm on datasets with class imbalance. Class imbalance is a common problem in machine learning that can cause bias in prediction results, especially in minority classes. ADASYN is one of the oversampling methods that focuses on adaptively synthesizing new data for minority classes. In this study, the performance of the Naïve Bayes algorithm was tested on Anemia Diagnosis datasets before and after the application of ADASYN. This dataset contains 104 instances, 5 attributes, and 2 classes, and has an imbalance ratio of 3. The evaluation was carried out by comparing accuracy, confusion matrix, precision, recall, and F1-score to obtain a more comprehensive picture of the effectiveness of ADASYN in improving Naïve Bayes. The results of the study show that the performance of the oversampling method depends on the imbalance ratio so it is important to ensure that the oversampling method does not cause overfitting and this can be overcome by using ADASYN which only selects Selected Neighbors. The results showed that ADASYN significantly increased accuracy from 0.57 to 0.78, precision from 0.17 to 0.74, recall from 0.20 to 0.88, and F1-Score from 0.18 to 0.80. In this study, we also compared the application of ADASYN and SMOTE on the Naïve Bayes algorithm. The results show that ADASYN outperforms SMOTE across all key metrics—accuracy, precision, recall, and F1-Score—while the accuracy improvements were statistically significant (p-value = 0.00903)
Hybrid Gradient Descent Grey Wolf Optimizer for Machine Learning Performance Enhancement
No full textAdvancements in machine learning have enabled the development of more accurate and efficient health prediction models. This study aims to improve diabetes prediction performance using the Support Vector Machine (SVM) model optimized with the Hybrid Gradient Descent Gray Wolf Optimizer (HGD-GWO) method. SVM is a robust machine learning algorithm for classification and regression. Still, its performance depends significantly on selecting appropriate hyperparameters such as regularization (C), kernel coefficient (γ), and polynomial kernel degree (d). The HGD-GWO method synergizes gradient descent for local optimization and the Gray Wolf Optimizer for global solution exploration. Using the Pima Indians Diabetes dataset, the process includes normalization, hyperparameter optimization, data division, and performance evaluation using accuracy, precision, recall, and F1-score metrics. The optimized SVM achieved an accuracy of 81.17%, with precision, recall, and F1-score values of 75.00%, 57.45%, and 65.06%, respectively, at a data ratio of 80%:20%. These findings highlight the potential of HGD-GWO in enhancing predictive models, particularly for early diabetes detection
An In-depth Exploration of Sentiment Analysis on Hasanuddin Airport using Machine Learning Approaches
No full textMachine learning-based sentiment analysis has become essential for understanding public perceptions of public services, including air transportation. Sultan Hasanuddin Airport, one of the main gateways in eastern Indonesia, faces the challenge of improving services amid changing user needs due to the COVID-19 pandemic. This study aims to compare the effectiveness of three machine learning algorithms- Support Vector Machine (SVM), Naive Bayes Multinomial, and K-Nearest Neighbor (KNN)-in analyzing the sentiment of user reviews related to airport services. The research also explores data splitting techniques, text preprocessing, data balancing using SMOTE, model validation, and method parameterization to ensure optimal results. The review data was retrieved from Google Maps (2021-2024) and underwent manual labelling. Text preprocessing includes normalization, stemming using Sastrawi, and stopword removal. The data-balancing technique uses SMOTE, while model evaluation is done with stratified k-fold cross-validation. SVM with a linear kernel showed the best performance, achieving an F1-score of 98.4%. Naive Bayes performed optimally, achieving an F1-score of 93.9%, while KNN recorded the best F1-score of 92.0%. SMOTE was shown to improve Naive Bayes' performance on unbalanced datasets, although it did not significantly impact SVM. The findings of this study provide data-driven recommendations to improve services at Sultan Hasanuddin Airport, such as the management of cleaning facilities, waiting room comfort, and passenger flow efficiency. In addition, this research opens up opportunities for developing real-time sentiment analysis systems that can be applied in other air transportation sectors