Jurnal Politeknik Negeri Batam (PoliBatam)
Not a member yet
3001 research outputs found
Sort by
Sentiment Analysis on Rupiah Depreciation Against USD Using XGBoost
No full textThe depreciation of the rupiah against the United States dollar (USD) affects purchasing power and economic stability. Public responses are widely expressed through social media such as X and Instagram. This study aims to analyze public sentiment using the Extreme Gradient Boosting (XGBoost) algorithm. Data were collected through crawling and scraping, consisting of 13,443 X comments and 11,287 Instagram comments between January 2024 until April 2025. Preprocessing included emoji conversion, cleaning, case folding, normalization, tokenization, stopwords removal, and Stemming. Sentiment labeling was performed using the InSet Lexicon, TF-IDF weighting, and data splitting into 70:30, 80:20, and 90:10. The XGBoost model was trained with parameters: 100 estimators, learning rate 0.1, max depth 6, and subsample 0.8. Results showed accuracies of 74–76% on X data and stable 77% on Instagram. Model evaluation using precision, recall, and F1-score confirmed consistency: precision 0.76% – 0.84%, recall 0.86%–0.88%, and F1-score 0.82%–0.86%, reflecting a balance between accuracy and robustness in detecting sentiments. Sentiment distribution revealed that X is dominated by negative opinions (38%), while Instagram is more positive (41%). These findings confirm the effectiveness of XGBoost in sentiment classification and provide valuable insights for policymakers to design adaptive communication and monetary strategies based on digital public opinion.The depreciation of the rupiah against the United States dollar (USD) affects purchasing power and economic stability. Public responses are widely expressed through social media such as X and Instagram. This study aims to analyze public sentiment using the Extreme Gradient Boosting (XGBoost) algorithm. Data were collected through crawling and scraping, consisting of 13,443 X comments and 11,287 Instagram comments between January 2024 until April 2025. Preprocessing included emoji conversion, cleaning, case folding, normalization, tokenization, stopwords removal, and Stemming. Sentiment labeling was performed using the InSet Lexicon, TF-IDF weighting, and data splitting into 70:30, 80:20, and 90:10. The XGBoost model was trained with parameters: 100 estimators, learning rate 0.1, max depth 6, and subsample 0.8. Results showed accuracies of 74–76% on X data and stable 77% on Instagram. Model evaluation using precision, recall, and F1-score confirmed consistency: precision 0.76% – 0.84%, recall 0.86%–0.88%, and F1-score 0.82%–0.86%, reflecting a balance between accuracy and robustness in detecting sentiments. Sentiment distribution revealed that X is dominated by negative opinions (38%), while Instagram is more positive (41%). These findings confirm the effectiveness of XGBoost in sentiment classification and provide valuable insights for policymakers to design adaptive communication and monetary strategies based on digital public opinion
Fuzzy Logic and Neural Network-Based Self-Tuning PID for Vacuum Pressure Stabilization
No full textThe conventional PID controller is widely used for vacuum pressure control; however, it has limitations when faced with nonlinear system characteristics and external disturbances, leading to a decline in performance. Several previous studies have proposed the integration of PID with intelligent methods, such as neural networks or fuzzy logic separately. Nevertheless, these singular approaches still encounter limitations in terms of adaptability and robustness. This study aims to develop a self-tuning PID method based on the combination of Neural Networks (NN) and Fuzzy Inference Systems (FIS) to enhance the stability and accuracy of vacuum pressure control. A nonlinear vacuum system plant model is constructed within the Simulink environment to generate a dataset used for training the NN with the Levenberg-Marquardt algorithm. The NN is employed to predict changes in PID parameters adaptively, while the FIS provides fine corrections to strengthen system stability. Simulation results demonstrate that the proposed approach effectively reduces overshoot from 36.47% to 31.51%, decreases steady-state error from 0.069 to 0.052, and lowers the RMSE value from 0.125 to 0.108 compared to conventional PID. Thus, the integration of NN and FIS within the self-tuning mechanism proves to be more effective in addressing nonlinear dynamics and external disturbances, resulting in a more stable and accurate system response
Multi-Modal Sensor Integration in Smart Rooms to Optimize Internet of Things-Based Monitoring and Security Control of Autistic Child Detection Activities
No full textThe advancement of Internet of Things (IoT) technology has opened new opportunities for automated monitoring systems, especially for children with Autism Spectrum Disorder (ASD). These children require intensive supervision due to communication limitations and unpredictable behavior. This study aims to design and implement a smart room system integrated with multi-modal sensors to monitor autistic children\u27s activities in real time.Using a Research and Development (R&D) approach with the ADDIE model, the system was developed with an ESP32 microcontroller and sensors including PIR (motion), DHT22 (temperature), microphone (sound), and LDR (light). The Mamdani fuzzy logic algorithm processes sensor data to classify safety levels. Data is visualized and notified via the Blynk platform.Test results show the system effectively detects "safe," "needs attention," and "critical" conditions with high accuracy, providing timely alerts for parents. This solution enhances home-based supervision and offers a practical, IoT-based approach to child safety and care
Predicting Cryptocurrency Prices Using Machine Learning: A Case Study on Bitcoin
No full textThe rapid growth of cryptocurrencies, particularly Bitcoin, has drawn significant attention from investors and researchers due to its extreme price volatility. However, predicting the price of Bitcoin against the Indonesian Rupiah (BTC/IDR) remains a major challenge, especially in emerging markets such as Indonesia. This study aims to conduct an empirical comparison among three deep learning models Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and one-dimensional Convolutional Neural Network (CNN-1D) in forecasting Bitcoin prices based on historical data obtained from the Indodax platform for the period 2018–2025. The dataset consists of five main variables: opening price, highest price, lowest price, closing price, and trading volume. Prior to model training, preprocessing steps were conducted, including handling missing values using the forward fill method, normalization with MinMaxScaler, and constructing time series data with a 60-day look-back window. The models were trained using an 80% training and 20% testing data split, the Adam optimizer, Mean Squared Error (MSE) as the loss function, for 50 epochs with a batch size of 32. Evaluation was performed using five quantitative metrics: MSE, RMSE, MAE, MAPE, and R², along with validation techniques to prevent data leakage. The results indicate that the GRU model achieved the best performance, with a MAPE of 1.77% and an R² of 0.9916, outperforming LSTM (MAPE 3.90%) and CNN-1D (MAPE 6.17%). These findings suggest that GRU is computationally more efficient and better adapted to nonlinear temporal dependencies in highly volatile markets. This research contributes to the academic discourse on the application of deep learning for digital asset price forecasting and provides practical implications for investors and developers of financial predictive systems in Indonesia. Future studies are expected to explore hybrid models or multi-step forecasting approaches to enhance real-time predictive performance.The rapid growth of cryptocurrencies, particularly Bitcoin, has drawn significant attention from investors and researchers due to its extreme price volatility. However, predicting the price of Bitcoin against the Indonesian Rupiah (BTC/IDR) remains a major challenge, especially in emerging markets such as Indonesia. This study aims to conduct an empirical comparison among three deep learning models Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and one-dimensional Convolutional Neural Network (CNN-1D) in forecasting Bitcoin prices based on historical data obtained from the Indodax platform for the period 2018–2025. The dataset consists of five main variables: opening price, highest price, lowest price, closing price, and trading volume. Prior to model training, preprocessing steps were conducted, including handling missing values using the forward fill method, normalization with MinMaxScaler, and constructing time series data with a 60-day look-back window. The models were trained using an 80% training and 20% testing data split, the Adam optimizer, Mean Squared Error (MSE) as the loss function, for 50 epochs with a batch size of 32. Evaluation was performed using five quantitative metrics: MSE, RMSE, MAE, MAPE, and R², along with validation techniques to prevent data leakage. The results indicate that the GRU model achieved the best performance, with a MAPE of 1.77% and an R² of 0.9916, outperforming LSTM (MAPE 3.90%) and CNN-1D (MAPE 6.17%). These findings suggest that GRU is computationally more efficient and better adapted to nonlinear temporal dependencies in highly volatile markets. This research contributes to the academic discourse on the application of deep learning for digital asset price forecasting and provides practical implications for investors and developers of financial predictive systems in Indonesia. Future studies are expected to explore hybrid models or multi-step forecasting approaches to enhance real-time predictive performance
Comparison of Online Gambling Promotion Detection Performance Using DistilBERT and DeBERTa Models
No full textOnline gambling promotions on social media have become a serious concern in Indonesia, where perpetrators use ambiguous and disguised language to evade detection. This study compares two transformer-based models, DistilBERT and DeBERTa, in detecting such content within Indonesian YouTube comments. Using a balanced dataset of 6,350 comments, both models were fine-tuned with optimized hyperparameters (learning rate 1e-5, batch size 32, 5 epochs) and evaluated through five-fold cross-validation. Results show that DeBERTa achieves superior performance with 99.84% accuracy and perfect recall, while DistilBERT achieves 99.29% accuracy. Error and linguistic analyses indicate that DeBERTa’s disentangled attention and Byte-Pair Encoding provide better understanding of non-standard and ambiguous language. Despite requiring higher computational cost, DeBERTa is ideal for high-accuracy applications, whereas DistilBERT remains suitable for real-time and resource-limited environments.Online gambling promotions on social media have become a serious concern in Indonesia, where perpetrators use ambiguous and disguised language to evade detection. This study compares two transformer-based models, DistilBERT and DeBERTa, in detecting such content within Indonesian YouTube comments. Using a balanced dataset of 6,350 comments, both models were fine-tuned with optimized hyperparameters (learning rate 1e-5, batch size 32, 5 epochs) and evaluated through five-fold cross-validation. Results show that DeBERTa achieves superior performance with 99.84% accuracy and perfect recall, while DistilBERT achieves 99.29% accuracy. Error and linguistic analyses indicate that DeBERTa’s disentangled attention and Byte-Pair Encoding provide better understanding of non-standard and ambiguous language. Despite requiring higher computational cost, DeBERTa is ideal for high-accuracy applications, whereas DistilBERT remains suitable for real-time and resource-limited environments
Boosting CNN Accuracy for Sundanese Script Recognition through Feature Extraction Techniques
No full textSundanese script is included in the cultural heritage in Indonesia, especially the culture in West Java. As a society that appreciates and preserves Indonesian culture and art, active participation can be realized through efforts to strengthen and preserve this script, one of which is by utilizing digital media. One of the technology-based digital media that can be used to preserve culture is image detection to make it easier to recognize Sundanese script. One of the models that can be used is the Convolutional Neural Network (CNN) with the MobileNetV2 architecture, with limited resources this architecture is able to produce good detection. This study applies the Convolutional Neural Network (CNN) algorithm with the MobileNetV2 architecture which will be tested with two main test scenarios, namely by applying feature extraction and without using feature extraction. The focus of this study will explore the influence and significance of the influence of feature extraction on the final results of image detection using the Convolutional Neural Network (CNN). The two feature extraction models used are Local Binary Pattern and Gray-Level Co-occurrence Matrix. These two feature extraction models will be tested with Sundanese script image data with data of 2,300 Sundanese script images. The results of this study show that the best results were obtained in the Convolutional Neural Network (CNN) with Gray-Level Co-occurrence Matrix (GLCM) with the best accuracy results at 93.8%. This is because the addition of the Gray-Level Co-occurrence Matrix (GLCM) is able to capture spatial texture statistics such as contrast, homogeneity, entropy, and correlation between pixel pairs. With these results, it can be concluded that in this study feature extraction has an effect and is able to increase the detection accuracy of the Convolutional Neural Network (CNN) model with the MobileNetV2 architecture in Sundanese script image data
MRI Classification of Brain Tumors Using EfficientNetB0 Feature Extraction and Machine Learning Methods
No full textBrain tumor classification using MRI images plays a crucial role in modern medical diagnostics, offering fast and accurate support for disease detection. This study proposes a classification approach that combines feature extraction using EfficientNet B0 with conventional machine learning algorithms. MRI brain images are preprocessed and resized to match EfficientNet B0 input dimensions. Feature vectors are extracted and subsequently processed using PCA for dimensionality reduction and SMOTE for class balancing. The resulting data are classified using various machine learning algorithms including Support Vector Machine, XGBoost, LightGBM, and others. Experimental results show that Support Vector Machine achieved the highest accuracy of 96%, followed by XGBoost and LightGBM at 94%. The combination of EfficientNet B0 feature extraction and lightweight classifiers proved to be effective, matching the performance of more complex deep learning models. This study does not focus on measuring computational cost directly, but rather demonstrates that combining EfficientNetB0 feature extraction with machine learning algorithms can achieve performance comparable to deep learning approaches. This highlights that lightweight models remain competitive in terms of accuracy without requiring highly complex architectures. Future work can explore this method on other medical imaging datasets and enhance model interpretability for clinical adoption.Brain tumor classification using MRI images plays a crucial role in modern medical diagnostics, offering fast and accurate support for disease detection. This study proposes a classification approach that combines feature extraction using EfficientNet B0 with conventional machine learning algorithms. MRI brain images are preprocessed and resized to match EfficientNet B0 input dimensions. Feature vectors are extracted and subsequently processed using PCA for dimensionality reduction and SMOTE for class balancing. The resulting data are classified using various machine learning algorithms including Support Vector Machine, XGBoost, LightGBM, and others. Experimental results show that Support Vector Machine achieved the highest accuracy of 96%, followed by XGBoost and LightGBM at 94%. The combination of EfficientNet B0 feature extraction and lightweight classifiers proved to be effective, matching the performance of more complex deep learning models. This study does not focus on measuring computational cost directly, but rather demonstrates that combining EfficientNetB0 feature extraction with machine learning algorithms can achieve performance comparable to deep learning approaches. This highlights that lightweight models remain competitive in terms of accuracy without requiring highly complex architectures. Future work can explore this method on other medical imaging datasets and enhance model interpretability for clinical adoption
Aircraft Image Classification on a Small-Scale Dataset using MobileNetV2 with Grad-CAM as Explainable AI
No full textThis study explores aircraft image classification using MobileNetV2 combined with Gradient-weighted Class Activation Mapping (Grad-CAM) for model interpretability. A dataset of 1,500 balanced images—helicopters, propeller aircraft, and jets—was split into training, validation, and testing sets with data augmentation to reduce overfitting. Transfer learning with pre-trained MobileNetV2 achieved an accuracy of 87.56%, with macro-average precision and recall of 85.76% and 87.69%. Grad-CAM visualizations confirmed that correct predictions relied on distinctive features such as rotor blades, propellers, and engines, while misclassifications often stemmed from background distractions or less discriminative areas. These findings demonstrate the potential of lightweight architectures for small-scale datasets and highlight the value of Explainable AI in validating deep learning models. The study provides a practical reference for educational contexts and offers directions for future work with larger datasets
Evaluating Fasttext and Glove Embeddings for Sentiment Analysis of AI-Generated Ghibli-Style Images
No full textThe development of text-to-image generation technology based on artificial intelligence has triggered mixed public reactions, especially when applied to iconic visual styles such as Studio Ghibli. This research aims to evaluate public sentiment towards the phenomenon of Ghibli-style AI images by comparing two static word embedding methods, namely FastText and GloVe, on three classification algorithms: Logistic Regression, Random Forest, and Convolutional Neural Network (CNN). Data in the form of Indonesian tweets were collected from Twitter using hashtags such as #ghibli, #ghiblistyle, and #hayaomiyazaki during the period 25 March to 25 April 2025. Each tweet was manually labelled with positive or negative sentiment, then preprocessed and represented using pre-trained FastText and GloVe embeddings. Evaluation was conducted using accuracy, precision, recall, and F1-score metrics, both macro and weighted. Results showed that FastText consistently performed the best on most models, especially in terms of precision and overall accuracy, thanks to its ability to handle sub-word information and spelling variations in social media texts. The combination of CNN with FastText yielded the highest performance with a macro F1-score of 76.56% and accuracy of 84.69%. However, GloVe still showed competitive performance in recall on the Logistic Regression model, making it relevant for contexts that prioritise sentiment detection coverage. This study emphasizes the importance of selecting embeddings and models that are appropriate to the characteristics of the data and the purpose of the analysis in informal social media-based sentiment classification.The development of text-to-image generation technology based on artificial intelligence has triggered mixed public reactions, especially when applied to iconic visual styles such as Studio Ghibli. This research aims to evaluate public sentiment towards the phenomenon of Ghibli-style AI images by comparing two static word embedding methods, namely FastText and GloVe, on three classification algorithms: Logistic Regression, Random Forest, and Convolutional Neural Network (CNN). Data in the form of Indonesian tweets were collected from Twitter using hashtags such as #ghibli, #ghiblistyle, and #hayaomiyazaki during the period 25 March to 25 April 2025. Each tweet was manually labelled with positive or negative sentiment, then preprocessed and represented using pre-trained FastText and GloVe embeddings. Evaluation was conducted using accuracy, precision, recall, and F1-score metrics, both macro and weighted. Results showed that FastText consistently performed the best on most models, especially in terms of precision and overall accuracy, thanks to its ability to handle sub-word information and spelling variations in social media texts. The combination of CNN with FastText yielded the highest performance with a macro F1-score of 76.56% and accuracy of 84.69%. However, GloVe still showed competitive performance in recall on the Logistic Regression model, making it relevant for contexts that prioritise sentiment detection coverage. This study emphasizes the importance of selecting embeddings and models that are appropriate to the characteristics of the data and the purpose of the analysis in informal social media-based sentiment classification
Transfer Learning Analysis on Tuberculosis Classification Using MobileNetV2 Architecture Based on Chest X-Ray Images
No full textTuberculosis(TBC) remains a major global health issue, with millions of new cases reported annually. Early and accurate diagnosis is essential, but manual interpretation of chest X-ray(CXR) images is limited by subjectivity and resource constrains. This study applies the MobileNetV2 architecture using transfer learning to classify tuberculosis from CXR images. The publicly available Tuberculosis Chest X-ray dataset containing 4200 images was divided into training (70%), validation (15%), and testing (15%). The pretrained MobileNetV2 model on ImageNet was used as the base network, with additional classification layers and training through the Adam optimizer and early stopping. The model achieved a validation accuracy above 99.84% after the second epoch maintained stable performance. Once the test set, model reached 99.84% accuracy, with precision 99.53% and recall 99.90% for the tuberculosis class. The result demonstrate that the transfer learning with MobileNetV2 provides a fast, efficient, and highly accurate method for tuberculosis detection. This model show potential for integration into Computer-Aided Diagnosis (CAD) system in low resource clinical settings.Tuberculosis(TBC) remains a major global health issue, with millions of new cases reported annually. Early and accurate diagnosis is essential, but manual interpretation of chest X-ray(CXR) images is limited by subjectivity and resource constrains. This study applies the MobileNetV2 architecture using transfer learning to classify tuberculosis from CXR images. The publicly available Tuberculosis Chest X-ray dataset containing 4200 images was divided into training (70%), validation (15%), and testing (15%). The pretrained MobileNetV2 model on ImageNet was used as the base network, with additional classification layers and training through the Adam optimizer and early stopping. The model achieved a validation accuracy above 99.84% after the second epoch maintained stable performance. Once the test set, model reached 99.84% accuracy, with precision 99.53% and recall 99.90% for the tuberculosis class. The result demonstrate that the transfer learning with MobileNetV2 provides a fast, efficient, and highly accurate method for tuberculosis detection. This model show potential for integration into Computer-Aided Diagnosis (CAD) system in low resource clinical settings