Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi)
Not a member yet
1071 research outputs found
Sort by
HyRoBERTa: Hybrid Robustly Optimized BERT Approach Model for Sentiment and Sarcasm Detection in Post-Flood Social Media Analysis
No full textThe detection problem is a crucial step in sentiment classification because it strengthens the validity and reliability of the model's interpretation of ambiguous text, especially in complex social contexts such as post-disaster public communication. Without this detection, the model is prone to significant classification errors. This study presents a hybrid approach for sentiment analysis with sarcasm detection after a flood disaster by combining the RoBERTa model with sequential deep learning architectures such as GRU, LSTM, and BiLSTM. We used a dataset of 17,520 tweets that were pre-processed using cleaning, normalization, and tokenization. Then, the positive class is further detected to determine whether it is sarcasm. The model was trained using a transformer-based transfer learning method with a combination of hyperparameters: the number of epochs, batch size, dropout rate, and learning rate. The experimental results show that the RoBERTa-GRU model achieved the highest accuracy for sentiment classification at 97. 26%, whereas the RoBERTa-BiLSTM model excels in detecting sarcasm with an accuracy of 98. 74%. RoBERTa-BiLSTM excels in sarcasm detection because it provides a bidirectional sequential mechanism and better long-term memory, effectively leveraging RoBERTa's rich embedding to identify contextual contradictions that are characteristic of sarcasm. Meanwhile, RoBERTa-GRU succeeds in sentiment classification because its architecture is more concise yet effective enough to infer dominant sentiments that have been filtered from the robust representation provided by RoBERTa, making the model more efficient for less complex tasks
Bayesian Hyperparameter Optimization of Lightweight CNNs for Facial Dermatological Classification
No full textConvolutional Neural Networks (CNNs) have been widely applied for skin condition classification. However, fair comparisons across lightweight architectures are often hindered by inconsistent hyperparameter settings. This study investigates the performance of two efficient CNN architectures, EfficientNetB3 and MobileNetV3, for facial dermatological classification across seven skin condition categories. To ensure optimal and comparable performance, Bayesian hyperparameter optimization was employed, alongside data augmentation to improve generalization. Experimental results show that EfficientNetB3 achieved the highest accuracy of 91.91%, outperforming MobileNetV3 at 90.44%. Beyond model comparison, this work highlights the novelty of applying Bayesian optimization to achieve fair benchmarking of lightweight CNNs under limited dataset conditions. The best-performing model was further deployed as a mobile application using TensorFlow Lite and Flutter, demonstrating its potential for real-world dermatological support
The Effect of Gamma Correction on the Accuracy of Vehicle Detection Using the YOLOv8 Algorithm
No full textAccurate vehicle detection under low-light conditions is a significant challenge in traffic surveillance systems and computer vision applications. Although YOLOv8 performs well under normal illumination, its accuracy decreases when processing low-light images due to reduced contrast and limited visual details. This study proposes the integration of gamma correction as a preprocessing method to enhance image brightness and improve YOLOv8 detection performance. The dataset consists of real ATCS traffic camera recordings from Medan City under varying lighting conditions. Gamma correction with three values (0.5, 1.5, and 2.0) was applied to evaluate its effect on detection accuracy. The results show that gamma 1.5 provides the best improvement, increasing [email protected] by 0.14% and [email protected]:0.95 by 0.74%, and achieving the highest confidence score of 0.9678 while also producing more stable training convergence. The novelty of this study lies in applying gamma correction to YOLOv8 using real-world ATCS low-light data, demonstrating that simple preprocessing can enhance detection robustness without modifying the model architecture
Forecasting IHSG Stock Prices Using an Attention-Based CNN-BiGRU Hybrid Deep Learning
No full textThis study develops an IHSG stock price forecasting model using a hybrid CNN–BiGRU architecture enhanced by an attention mechanism. The key novelty lies in combining CNN-based local pattern extraction with BiGRU-based bidirectional temporal modeling, while attention selectively emphasizes the most informative time steps, improving representation quality for complex and noisy financial series. Historical IHSG data from public sources were preprocessed through feature engineering and normalization, followed by XGBoost-based feature selection to retain the most predictive variables. Model robustness was assessed in two settings: (i) the full dataset and (ii) a “cleaned” dataset excluding the extreme COVID-19 volatility period. The proposed model achieved strong accuracy, with MAE/RMSE of 0.0125/0.02 on the full dataset and 0.0167/0.03 on the cleaned dataset, while Pearson correlation remained close to 1 in both scenarios, indicating high alignment with actual IHSG movements. A 30-day ahead forecast produced a stable and realistic trend. Overall, the CNN–BiGRU with attention provides an effective and robust approach for capturing multi-scale temporal patterns in IHSG forecasting
Firefly Algorithm Under-sampling for Imbalance Data in Breast Cancer Survival Prediction
No full textBreast cancer remains a major health challenge, affecting approximately 1.7 million individuals annually and often leading to severe complications. Predicting survival outcomes is difficult due to highly imbalanced data, with 3,408 death cases compared to only 616 survival cases. To address this issue, we applied the Firefly Algorithm–based under-sampling (FAUS) to balance the dataset and combined it with three machine learning classifiers: Random Forest (RF), Decision Tree (DT), and K-Nearest Neighbor (KNN). Experimental results show that FAUS substantially improves predictive performance compared to conventional under-sampling. Among the tested models, RF achieved the highest F1-score of 0.79, while DT and KNN reached 0.72 and 0.68, respectively. The results indicate that FAUS is effective in preserving representative samples, thereby enhancing model performance in breast cancer survival prediction.
 
Serendipitous Recommendations for Handicraft Store Discovery in Social Commerce Using a Genetic Algorithm with Adaptive Selection
No full textIn social commerce, particularly among small and medium-sized handicraft enterprises (SMEs), personalized recommender systems (RS) are crucial for enhancing store and product discovery. Conventional content-based filtering (CBF) often overemphasizes accuracy, leading to over-specialization and limiting exposure to novel or diverse items, an issue in the handicraft sector where uniqueness is valued. This study proposes a serendipitous recommendation approach using a Genetic Algorithm (GA) with adaptive selection strategies, Roulette Wheel Selection (RWS), Tournament Selection (TnS), and Rank-Based Selection (RBS), to balance relevance and unexpectedness. Handicraft store attributes, such as product types, materials, and services, are encoded in a 19-bit chromosome and evaluated via a hybrid fitness function. Tested on real data from West Sumatra SMEs, the model is assessed using Precision, Recall, Novelty, and Serendipity metrics. Results show that the GA-based adaptive selection approach outperforms baseline CBF in producing more diverse and surprising recommendations, fostering exploratory shopping experiences and supporting the discovery of unique local products in social commerce ecosystems
A Comparative Evaluation of Federated Learning Algorithms for Privacy-Preserving Academic Prediction on Heterogeneous Data
No full textThe rapid growth of educational data enables predictive analytics for academic performance, yet privacy regulations like GDPR and FERPA severely restrict centralized data sharing. Although Federated Learning (FL) has succeeded in privacy-sensitive fields such as healthcare, its application in education remains underexplored, lacking systematic comparative studies of multiple FL algorithms across diverse educational datasets—especially emphasizing recall and ROC-AUC as critical metrics for early identification of students at academic risk. This study fills this gap by evaluating five FL algorithms—Federated Averaging (FedAvg), Federated Proximal (FedProx), Federated Dynamics (FedDyn), Fair Federated Averaging (q-FedAvg), and SCAFFOLD—for privacy-preserving prediction of academic outcomes. Three public datasets were purposefully selected for their representativeness and heterogeneity: Predict Students Dropout and Academic Success (binary dropout prediction with socioeconomic factors), Student Performance (multi-class grade prediction in secondary education), and xAPI-Edu-Data (multi-class performance based on online learning activities). Local neural networks employed Stratified 5-Fold Cross-Validation, while FL algorithms ran for 50 communication rounds. Global models, particularly q-FedAvg and FedProx, consistently surpassed local models, with q-FedAvg achieving 0.7668 accuracy, 0.6813 recall, and 0.8810 ROC-AUC on Predict Students Dropout; 0.8580 accuracy and 0.9871 recall on Student Performance; and 0.7396 accuracy and 0.8815 ROC-AUC on xAPI-Edu-Data. Paired T-tests confirmed significant recall gains for most global models (p < 0.05). These results highlight FL’s ability to handle data heterogeneity and privacy constraints while improving predictive performance, thereby supporting timely educational interventions and enhanced student retention policies
Comparative Analysis of CNN, MobileNetV2 and EffecientNetBO in Smart Farming System for Chili Leaf Disease Detection
No full textChili leaf diseases greatly affect agricultural productivity, making early and accurate detection essential to support smart farming systems. This study presents a comparative analysis of three deep learning architectures—Convolutional Neural Network (CNN), MobileNetV2, and EfficientNetB0—for detecting chili leaf diseases using RGB images. The dataset consists of three main disease classes: Bacterial Spot, Curl Virus, and White Spot. Each model was trained and evaluated using accuracy, precision, recall, F1-score, macro AUC, and training time as performance metrics. Experimental results show that MobileNetV2 achieved the highest performance with 99% accuracy, 0.99 F1-score, and 0.99 macro AUC, although it required the longest training time of 115.12 seconds. CNN demonstrated competitive results with 96% accuracy and the shortest training time of 60 seconds, while EfficientNetB0 performed poorly with only 38% accuracy and an F1-score of 0.18. These findings highlight that model architecture, dataset characteristics, and training configuration significantly influence performance outcomes. This study contributes to the development of intelligent agricultural monitoring systems by identifying the most suitable deep learning architecture for real-time chili leaf disease detection in smart farming applications
Comparative Analysis of Multispectral Image Classification Based on EfficientNetB0, ResNet152, DenseNet161, DenseNet121, and HSV Segmentation
No full textThis study established a classification system based on Convolutional Neural Networks (CNNs) to detect High-High Fluctuation (HHF) patterns in multispectral data derived from pure water (H2O) and a water-sodium hydroxide (NaOH) solution. This study combines HSV color-space-based segmentation to identify areas with the highest signal amplitude, thereby enhancing the feature extraction of the CNN model. Data augmentation techniques, including random flipping, rotation, and color jitter, along with training parameters such as a learning rate of 0.0001 and a batch size of 32, have been shown to effectively improve model generalization and reduce overfitting. Four different CNN architectures were evaluated: ResNet-152, DenseNet-161, DenseNet-121, and EfficientNet-B0. As a result, ResNet152 achieved the highest accuracy of 97.6%, attributed to its network depth and residual connections that effectively address the vanishing gradient problem. DenseNet161 and DenseNet121 also demonstrated competitive performance, achieving accuracies of 96.7% and 96.2%, respectively, which is supported by their dense connectivity that optimizes feature reuse. Conversely, EfficientNetB0, despite showing lower accuracy (90%), provides significant computational efficiency, making it suitable for real-time applications. These results underscore the importance of selecting a CNN architecture that balances accuracy and efficiency for multispectral data classification
A Validated Blockchain Model and Architecture for Public Health System for Data Security
No full textBlockchain has been identified as a technology that can be used in healthcare systems, especially in terms of data security, privacy, and interoperability. This article explores the application of blockchain in various medical processes ranging from patient registration, visit registration, initial examination, patient examination, diagnostic input, and drug preparation. This study was conducted at a healthcare facility that utilizes blockchain as a technology to manage patient data, medical records, and drug prescriptions. This study uses a combined approach between literature studies and qualitative methods using the Focus Group Discussion (FGD) technique for evaluates the impact of model blockchain. Model validation with domain experts in the health sector (doctors, nurses, administrators, IT experts, and pharmacists). The results research is a validated Model blockchain public health sector improves the security and privacy of patient data but also accelerates operational processes in the healthcare chain. This study concludes that blockchain can be a transformational solution for addressing major problems in the healthcare sector, although cross-sector collaboration is needed to ensure successful implementation