Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi)
Not a member yet
1071 research outputs found
Sort by
Harnessing BERT for Semantic Understanding in Tourism Recommendation Engines
No full textIt will be necessary for attraction managers within hotels to track guests' lifestyles to keep the business running. Such an understanding may be achieved, for example by analyzing reviews on attractions to capture the attitudes of the visitors towards the services and business within the tourism industry. The approach utilizes web scraping to gather user-generated reviews, using text preprocessing, data pre-processing, and further improvement of the model using labelled sentiment data divided into three sentiment classes: positive, negative, or neutral. The dataset consisting of 908 reviews were divided in 70:15:15 ratio for training, validation and testing. Model performance was measured in terms of accuracy, precision, recall and F1-score. In this study, the BERT deep learning model is used to classify sentiments of Indonesian tourist. Using the SmallBERT variant fine-tuned on 515k reviews for 5 epochs, the model achieved 91.40% accuracy, 90.51% precision, recall, and F1 score. The results indicate a dominance of positive sentiments, visualized using tableau. This research provides a robust foundation for developing intelligent sentiment-based recommendation systems in the tourism sector and suggests future exploration using other transformer-based models such as GPT, T5, or BART for comparative analysis
Early Detection of Grasserie Disease in Silkworms Using Computer Vision and Machine Learning
No full textSilkworm diseases are a significant threat to the sericulture industry, with early detection remaining a major challenge due to limited resources. Timely identification of infected silkworms is essential to curb the spread of disease and reduce economic damage. This study focuses on diagnosing Grasserie disease, a highly contagious condition that can devastate silkworm populations, leading to substantial financial losses for farmers. To address the shortcomings of expert manual inspections, this study employed camera-captured images of silkworms for automated disease detection. A newly compiled dataset, consisting of 668 healthy silkworms and 574 infected with Grasserie disease, forms the basis of the investigation. The study applies machine learning techniques for image analysis, combining Histogram of Oriented Gradients (HOG) for feature extraction, Kernel Principal Component Analysis (KPCA) for dimensionality reduction, and supervised classification models. The results highlight the effectiveness of this approach in differentiating healthy silkworms from diseased ones. The machine learning model HOG integrated with KPCA and Decision Trees (DT) achieved strong performance, with accuracy, recall, and precision scores of 94.28%, 94.56%, and 92.48%, respectively. While these outcomes are encouraging, further research is needed to develop a practical IoT-based tool that enables sericulture farmers to quickly detect infections and take preventive measures, minimizing unexpected losses. This study marks a crucial advancement in silkworm disease detection, offering a pathway toward greater sustainability and economic stability in the sericulture sector
Computer Vision-Based Information System for Early Detection of Elderly Patient Falls using YOLOv12
No full textFalls in elderly patients are a significant public health problem due to their high frequency and potential to cause serious injury or even death. Traditional fall detection systems often rely on wearable sensors, which can be intrusive and uncomfortable for long-term monitoring. This study proposes a non-intrusive computer vision-based information system for early fall detection using the YOLOv12 (You Only Look Once version 12) object detection model. The system integrates real-time video processing with a lightweight convolutional neural network architecture to detect falls in indoor care settings. A dataset of 10,793 annotated images, including simulated fall scenarios and daily activities, was used to train and validate the proposed model. The proposed system achieved a Mean Average Precision (mAP) of 90.60%, demonstrating robust performance under various lighting conditions and camera angles when compared with the YOLOv8, YOLOv11, and YOLO-NAS models. This study contributes to the development of intelligent healthcare systems that improve the safety and quality of life of elderly patients through proactive monitoring and rapid response capabilities
Early Stroke Disease Prediction Based on Lifestyle Factors Applied with Machine Learning
No full textStroke prediction has many supporting features and variables. Some forecasts focus more on health or elements that are already present. Predicting stroke risk by identifying habitual factors provides more advantages for preventive action. In addition, the complexity of features or variables is a concern in predicting stroke risk. In this study, we used a public dataset from Kaggle with 10 features or variables. In this study, we propose to collaborate algorithms and preprocessing in feature selection using Pearson Correlation and Principal Component Analysis (PCA) dimension reduction to unravel the complexity of variables and data processing computing. This aims to predict stroke risk more simply. The results of the experiment show that feature selection using Pearson Correlation between features and labels produces maximum results using 5 features out of 10 provided features. This approach produces the best performance on the Naïve Bayes, Iterative Dichotomiser Tree (ID3), Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and Logistic Regression with 100% accuracy and reduces features by 50% to support the reduction of the complexity of prediction variables and data processing computing
Application of Reinforcement Learning to Solve Rubrik’s Cube with Markov Decision Process
No full textThe Rubik's Cube is a complex puzzle with an enormous number of possible configurations, making it a challenging problem for both humans and computational methods to solve. While traditional solving algorithms rely on predefined strategies, this study explores the application of reinforcement learning (RL) to develop an adaptive and efficient solution model. This study aims to create an RL_based solver using the Markov Decision Process (MDP) framework, optimizing for speed, move efficiency, and solving steps. The proposed model employs Q-learning and Monte Carlo Tree Search (MCTS) to determine the optimal actions at each game state, which are trained through extensive Rubik's Cube simulations. The key novelty of this study lies in the integration of MCTS with Q-learning to enhance decision-making efficiency by reducing the number of moves compared with conventional methods. The experimental results demonstrate that the model achieves near-optimal solutions with fewer moves, outperforming basic rule-based approaches. Additionally, a web-based application was developed to provide real-time solving strategies based on user-input cube configurations. This study contributes to the advancement of RL applications in combinatorial puzzles and offers a practical tool for Rubik's Cube enthusiasts seeking to improve their solving techniques
Deep Learning-Based Visualization of Network Threat Patterns Using GAN-Generated Infographic
No full textDespite the growing sophistication of cyberattacks, current network traffic analysis tools often lack intuitive visual support, limiting human analysts’ ability to interpret complex threat behaviors. To address this gap, this study proposes a novel deep learning-based visualization framework using a Deep Convolutional Generative Adversarial Network (DCGAN) to synthesize threat-specific infographics from structured numerical features in the CICIDS 2017 dataset. Unlike conventional methods, such as PCA or static dashboards, which often result in abstract or non-adaptive visuals, our approach generates class-distinct grayscale images that preserve the behavioral patterns of various attacks, including denial-of-service, brute force, and port scanning. The preprocessing pipeline reshapes the selected flow-based features into 28×28 matrices to train the generative model. Evaluation using the Frechet Inception Distance (FID) yielded a score of 28.4, whereas a CNN classifier trained on the generated images achieved 91.2% accuracy, confirming visual fidelity and semantic integrity. Additionally, a panel of human experts rated the interpretability of the generated images at 4.3 out of 5.0. These findings demonstrate that generative visualization can enhance human-centered threat analysis by bridging raw data with interpretable imagery, thereby offering a scalable and explainable approach for integrating AI into real-time security workflows
Analyzing Public Sentiment on Electric Vehicles Through BERTopic and Emotion-Based Data Clustering
No full textThe escalating impact of technological advancements on worldwide society prompts a closer examination of their profound consequences. Enhanced communication methods and the significant influence of social media platforms stand out as critical factors, with the automotive industry responding to environmental concerns through the emergence of electric vehicles (EVs). In this work the relationship between the trends of EV evolving and social media was utilized using X (aka, Twitter) data. Specifically, this work studies the increasing market demand for EVs due to the impact of social media. Consequently, the study is crucial for both clients and EV manufacturers. To identify the primary discussion themes on Twitter, this article utilizes a topic modelling technique (BERTopic) a data mining method and analyses the production and sales of EV manufacturers. We utilized The National Research Council Canada's Emotion Lexicon (NRCLex) for emotion analysis. Trust, surprise, anger, anticipation, positive, negative, disgust, fear, sadness, and joy are the eight emotions of NRCLex that can provide awareness of the present dynamics. We compared current media coverage of EVs and topic-modeled data. The results showed that BERTopic and NRCLex provided a depth of analysis via the emotional analysis. Consequently, this study contributes to improving the understanding of public sentiment's influence on EV trends
Robust Aggregation Strategies in Federated Learning for Credit Risk Assessment
No full textFinancial institutions face challenges in credit risk assessment due to fragmented data and strict privacy regulations, which hinder predictive modeling and increase financial risks. Federated Learning (FL) enables privacy-preserving collaborative modeling without sharing raw data. This study evaluates five FL aggregation methods—Federated Averaging (FedAvg), Weighted Average, Median Aggregation, Federated Proximal (FedProx), and Stochastic Controlled Averaging (SCAFFOLD)—using logistic regression on the Credit Approval dataset (690 records, five clients) with non-IID label and feature distributions. Local models were trained and aggregated over 50 rounds. Median Aggregation outperformed the other methods, achieving an F1-score of 97.85% and a recall of 80.6% (vs. 72.3% for others), demonstrating robustness against data skewness. However, global model performance (85.22% for FedAvg, Weighted Average, FedProx, SCAFFOLD; 85.80% for Median) remained static across rounds, indicating limited convergence due to rapid local model convergence and non-IID challenges. The high communication cost of 50 rounds highlights a trade-off between accuracy and efficiency, necessitating optimized strategies like adaptive regularization or client sampling. This study advances theoretical understanding of FL under heterogeneity and provides practical guidance for secure, regulation-compliant credit risk modeling in financial institutions. Future work should explore larger datasets, multi-round convergence, and privacy mechanisms like differential privacy to mitigate risks such as model inversion attacks while ensuring complianc
Classification of Retinoblastoma Eye Disease on Digital Fundus Images Using Geometric Features and Machine Learning
Get PDFMedical image analysis is essential for detecting retinoblastoma tumors due to the ability of this method to assist doctors in examining the morphology, density, and distribution of blood vessels. The classification of normal and retinoblastoma-affected retinas is a preliminary step in treating retinoblastoma tumors. Therefore, this study aimed to propose a new method for classifying normal and retinoblastoma-affected retinas using geometric feature extraction and machine learning. The workflow consisted of (1) fundus image data collection for retinoblastomas, (2) image segmentation, (3) feature extraction process, (4) building a classification model using machine learning, (5) splitting testing and training data, (6) classification process using machine learning methods, and (7) evaluation of classification results using a confusion matrix. The results showed that the segmentation method could detect retinoblastoma areas and extract their geometric features. The SVM method achieved an accuracy of 0.96 while the RF and DT had 0.55 and 0.63, respectively. Moreover, a comparison with previous research showed that the proposed method achieved a 4% improvement in the classification performance. This led to the conclusion that classification using geometric features combined with the SVM on digital fundus images of retinoblastoma eye disease produced the best results
Deep learning with Bayesian Hyperparameter Optimization for Precise Electrocardiogram Signals Delineation
No full textElectrocardiography (ECG) serves as an essential risk-stratification tool to observe further treatment for cardiac abnormalities. The cardiac abnormalities are indicated by the intervals and amplitude locations in the ECG waveform. ECG delineation plays a crucial role in identifying the critical points necessary for observing cardiac abnormalities based on the characteristics and features of the waveform. In this study, we propose a deep learning approach combined with Bayesian Hyperparameter Optimization (BHO) for hyperparameter tuning to delineate the ECG signal. BHO is an optimization method utilized to determine the optimal values of an objective function. BHO allows for efficient and faster parameter search compared to conventional tuning methods, such as grid search. This method focuses on the most promising search areas in the parameter space, iteratively builds a probability model of the objective function, and then uses that model to select new points to test. The used hyperparameters of BHO contain learning rate, batch size, epoch, and total of long short-term memory layers. The study resulted in the development of 40 models, with the best model achieving a 99.285 accuracy, 94.5% sensitivity, 99.6% specificity, and 94.05% precision. The ECG delineation-based deep learning with BHO shows its excellence for localization and position of the onset, peak, and offset of ECG waveforms. The proposed model can be applied in medical applications for ECG delineation