Jurnal Online Informatika
Not a member yet
276 research outputs found
Sort by
The Impact of Online Reviews to Predict The Number of International Tourists
Full text linkThe tourism sector is a potential resource for advancing the Indonesian economy. The development of the tourism industry is represented by the number of international tourist arrivals. Therefore, this indicator becomes an objective in development programs. To accomplish this goal and assess the demand aspect of the tourism sector, it is a must to have a precise forecast of the number of international visitors. This research attempts to develop precise methods and models for estimating the number of international tourists based on this premise. This study chooses Bali Province as its focus since nearly half, or 47%, of the tourists who visit Indonesia arrive through the entry point in Bali Province. This research uses the LSTM method and big data online reviews in building prediction models. The results of this study show that sentiment analysis of tourist attractions in Bali using the BERT model has an accuracy of 75%. The results also depict that reviews by visitors about tourist attractions in Bali Province during the period 2012-2023 contain more positive sentiments. Furthermore, the best model to predict the number of international tourists, with the smallest RMSE and MAPE values (39,470.64 and 11.25%, respectively), includes inflation, rupiah exchange rates, TPK, monthly sentiment scores, and the number of reviews as dependent variables. The prediction model also show that the review variables (sentiment score and number of reviews) can improve prediction accuracy
Simulation and Empirical Studies of Long Short-Term Memory Performance to Deal with Limited Data
Full text linkThis research is proposed to determine the performance of time series machine learning in the presence of noise, where this approach is intended to forecast time series data. The approach method chosen is long short-term memory (LSTM), a development of recurrent neural network (RNN). Another problem is the availability of data, which is not limited to high-dimensional data but also limited data. Therefore, this study tests the performance of long short-term memory using simulated data, where the simulated data used in this study are data generated from the functional autoregressive (FAR) model and data generated from the functional autoregressive model of order 1 FAR(1) which is given additional noise. Simulation results show that the long short-term memory method in analyzing time series data in the presence of noise outperforms by 1-5% the method without noise and data with limited observations. The best performance of the method is determined by testing the analysis of variance against the mean absolute percentage error. In addition, the empirical data used in this study are the percentage of poverty, unemployment, and economic growth in Java. The method that has the best performance in analyzing each poverty data is used to forecast the data. The comparison result for the empirical data is that the M-LSTM method outperforms the LSTM in analyzing the poverty percentage data. The best method performance is determined based on the average value of the mean absolute percentage error of 1-10%
Comparison Airport Traffic Prediction Performance Using BiGRU and CNN-BiGRU Models
Full text linkCOVID-19 pandemic has significantly disrupted the aviation industry, highlighting the critical need for accurate airport traffic predictions. This study compares the performance of BiGRU and CNN-BiGRU models to enhance airport traffic forecasting accuracy models from March to December 2020. Data preprocessing was performed using Python\u27s Pandas library. This involved filtering, scaling using min-max normalization, and splitting the data into 80:20 training-testing split using Python\u27s Pandas library. Various optimization techniques—RMSProp, Adam, Nadam, Adamax, AdamW, and Lion—were applied, along with ReduceLROnPlateau, to optimize model performance. The models were evaluated using Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), and Mean Squared Error (MSE). The best predictive performance was observed in the United States using the CNN-BiGRU model with the Adam optimizer, achieving the lowest MAE of 0.0580, MSE of 0.0097, and MAPE of 0.0979. The use of a balanced dataset, representing each airport\u27s traffic as a percentage of a baseline period, significantly improved prediction accuracy. This research provides valuable insights for stakeholders seeking effective airport traffic prediction methods during unprecedented times
Generative Adversarial Networks In Object Detection: A Systematic Literature Review
Full text linkThe intersection of Generative Adversarial Networks (GANs) and object detection represents one of the most promising developments in modern computer vision, offering innovative solutions to longstanding challenges in visual recognition systems. This review presents a systematic analysis of how GANs are transforming these challenges, examining their applications from 2020 to 2025. The paper investigates three primary domains where GANs have demonstrated remarkable potential: data augmentation for addressing data scarcity, occlusion handling techniques designed to manage visually obstructed objects, and enhancement methods specifically focused on improving small object detection performance. Analysis reveals significant performance improvements resulting from these GAN applications: data augmentation methods consistently boost detection metrics such as mAP and F1-score on scarce datasets, occlusion handling techniques successfully reconstruct hidden features with high PSNR and SSIM values, and small object detection techniques increase detection accuracy by up to 10% Average Precision in some studies. Collectively, these findings demonstrate how GANs, integrated with modern detectors, are greatly advancing object detection capabilities. Despite this progress, persistent challenges including computational cost and training stability remain. By critically analyzing these advancements and limitations, this paper provides crucial insights into the current state and potential future developments of GAN-based object detection systems
A Comparison of YOLOv8 Series Performance in Student Facial Expressions Detection on Online Learning
Full text linkStudent engagement in online learning is an important factor that can affect learning outcomes. One indicator of engagement is facial expression. However, research on facial expression detection in online learning environments is still limited, especially in the use of the YOLOv8 algorithm. This study aims to compare the performance of several YOLOv8 variants, namely YOLOv8x, YOLOv8m, YOLOv8s, YOLOv8n, and YOLOv8l in recognizing six facial expressions: happy, sad, angry, surprised, afraid, and neutral. Student facial expression data was collected through the Moodle platform every 15 seconds during the learning process. All models were trained using 640x640 pixel images for 100 epochs to improve facial expression detection capabilities. The main contribution of this study is to provide a comprehensive analysis of the effectiveness of YOLOv8 in detecting student facial expressions, which can be used to improve the online learning experience. The evaluation results show that the YOLOv8s model has the best performance with the highest mAP of 0.840 and the fastest inference speed of 2.4 ms per image. YOLOv8m and YOLOv8x also performed well with mAP of 0.816 and 0.815, respectively. Although YOLOv8x had the slowest inference speed, it was superior in detecting fear, happiness, and sadness expressions with mAP above 0.9. YOLOv8n had mAP of 0.636, while YOLOv8l achieved mAP of 0.813 with an inference speed of 9.1 ms per image. This study shows that the YOLOv8 algorithm, especially YOLOv8s, can be an effective solution to analyze student engagement based on their facial expressions during online learning
LLM-Based Information Retrieval for Disease Detection Using Semantic Similarity
Full text linkInformation retrieval systems are vital for disease prediction, but traditional methods like TF-IDF struggle with word meanings and produce long, complex vectors. This research uses Large Language Models (LLMs) and follows the CRISP-DM methodology to improve accuracy. Using health forum discussions labeled with specific diseases, we split the data into queries and a corpus. Semantic similarity is used to retrieve the most relevant text from the corpus. After preprocessing, we compare LLMs and TF-IDF, with LLMs achieving an accuracy of 0.911 (Top-K=30), outperforming TF-IDF. LLMs excel by creating shorter, meaningful vectors that preserve context, enabling precise semantic matching. These results demonstrate LLMs\u27 potential to enhance healthcare information retrieval, offering more accurate and context-aware solutions. This research highlights how advanced AI can overcome traditional methods\u27 limitations, opening new possibilities for medical informatics
Cyberbullying Detection in the Libyan Dialect Using Convolutional Neural Networks
Full text linkecently, the widespread use of social media has increased, leading to increased concerns about cyberbullying. It has become imperative to intensify efforts and methods to detect and manage cyberbullying through social media. Arabic has recently received increasing attention to improve the classification of Arabic texts. Given the multitude of Arabic dialects used on social media platforms by Arabic speakers to express their opinions and communicate with each other, applying this approach to Arabic becomes extremely challenging due to its structural and morphological complexity. Analyzing Arabic dialects using Natural Language Processing (NLP) tools can be more challenging than Standard Arabic. In this paper, the impact of using stopword removal and derivation techniques on detecting cyberbullying in the Libyan dialect was presented. The efficiency of text classification was compared when using a Libyan dialect word list alongside pre-generated Modern Standard Arabic (MSA) lists. The texts were classified using Convolutional Neural Network (CNN) classifiers, and the experiments showed that when using Libyan dialect words, the accuracy results were 92% and 83%, and when using only Standard Arabic stop words, the accuracy results were dropped to 91% and 77%. Based on these results, the higher accuracy was obtained when using the presented stop words list which it is specific to the Libyan dialect, and they had a positive impact on the results, better than Standard Arabic stop words
Anatomy of Sentiment Analysis in Ontological, Epistemological, and Axiological Perspectives
Full text linkThe aim of this article was to examine sentiment analysis methods from the perspective of the philosophy of science with three approaches, ontological, epistemological and axiological. This research used a qualitative research method (descriptive-analysis) with an ontological, epistemological and axiological approach that uses library research and document studies of previous research results. Data collection was carried out through books and reputable scientific journals on Scopus, ScienceDirect, IEEEXplore and Springer Link. The results of this research showed that sentiment analysis from an ontological perspective describes the definition, development and relationship of sentiment with social reality. Meanwhile, from an epistemological perspective, sentiment analysis is viewed from how the source of knowledge is obtained, explaining the production of sentiment analysis knowledge, and several ways of working that can be applied in studies. Axiologically, sentiment analysis can see the function and value resulting from sentiment analysis, as well as discussing the results of interpretation from sentiment analysis studies. These findings showed the development of sentiment analysis in answering various problems to improve the quality of sustainable services in various fields
Modality-based Modeling with Data Balancing and Dimensionality Reduction for Early Stunting Detection
Full text linkIn Indonesia, the stunting rate has reached 36%, significantly higher than the World Health Organization\u27s (WHO) standard of 20%. This high prevalence underscores the urgent need for effective early detection methods. Traditional data mining approaches for stunting detection have primarily focused on unimodal data, either tabular or image data alone, limiting the comprehensiveness and accuracy of the detection models. Modality-based modeling, which integrates image and tabular data, can provide a more holistic view and improve detection accuracy. This research aims to analyze modality-based modeling for the early detection of stunting. Two modalities, unimodal and multimodal, are used in this study. The main contributions of this research are the development of a comprehensive framework for modality-based analysis, the application of advanced data preprocessing techniques, and the comparison of various machine learning algorithms to identify the best model for stunting detection. The dataset, comprising images and tabular data, is sourced from Posyandu in Sidoarjo, Indonesia. Image data undergoes preprocessing, including background segmentation and feature extraction using the Gray Level Co-occurrence Matrix (GLCM), while tabular data is processed through categorical encoding. The Synthetic Minority Oversampling Technique (SMOTE) addresses class imbalance, and Principal Component Analysis (PCA) is used for dimensionality reduction. Unimodal modeling uses tabular or image data alone, while multimodal modeling combines both before classification. The study achieves the best F1 scores of 0.96, 0.91, and 0.90 for tabular-only, image-only, and image-tabular modalities, respectively, demonstrating the effectiveness of data balancing and dimensionality reduction techniques
A Comparison Analysis Between ResNET50 and XCeption for Handwritten Hangeul Character using Transfer Learning
Full text linkThe enthusiasm for Korean pop culture in Indonesia has contributed to a growing interest in learning the Korean language, including its writing system, Hangeul, which currently ranks as the 6th most studied language. Hangeul has a unique structure, where each character is arranged in syllabic blocks of consonants and vowel combinations. The main challenge in Korean character classification lies in the similarity between characters and the complex structure, making it more difficult for models to recognize.
This study aims to compare two deep convolutional neural networks are ResNet50 and Xception, using transfer learning for handwritten Hangeul character classification. While previous studies have examined CNN-based character recognition, this study highlights the effectiveness of deeper architectures with limited yet augmented data. Unlike earlier works, it incorporates Grad-CAM visualizations, transfer learning with partial fine-tuning, and multiple train-test ratios to analyze model behavior. A total of 1,920 images across 24 classes were evaluated using 5-fold cross-validation, with extensive augmentation and preprocessing to simulate variation. The Machine Learning Life Cycle (MLLC) framework assessed model performance through accuracy, precision, recall, F1-score, and AUC. Both models achieved high performance, with ResNet50 consistently outperforming Xception in most folds, especially in precision and F1-score. ResNet50 achieved perfect scores (100%) across all metrics, while Xception also performed strongly with up to 99.74% accuracy. These results indicate that ResNet50 is more effective in classifying Korean letters on the dataset used in this study. For future research, a robustness evaluation can be applied using data that was not included in previous training or testing