Jurnal Politeknik Negeri Batam (PoliBatam)
Not a member yet
3001 research outputs found
Sort by
YOLOv11-Based Detection of Indonesian Traffic Signs: Transfer Learning vs. From-Scratch Training
Get PDFTraffic sign detection is a fundamental component in intelligent transportation systems (ITS), autonomous driving, and advanced driver assistance systems (ADAS), enabling vehicles to interpret road conditions and enhance safety. Developing robust traffic sign detection models for specific regions requires high-quality, well-annotated local datasets, which are often challenging and costly to create. Even when such datasets are available, training deep learning models from scratch demands substantial computational resources and time. This study compares models trained from scratch and those using transfer learning based on the lightweight YOLOv11s architecture on an Indonesian traffic sign dataset. Evaluations using precision, recall, mean Average Precision at IoU 0.5 ([email protected]), and mean Average Precision across IoU thresholds 0.5 to 0.95 ([email protected]:0.95) demonstrate that the transfer learning model consistently outperforms the from-scratch model across all metrics. These findings highlight the effectiveness and efficiency of transfer learning for developing accurate and practical traffic sign detection systems adapted to local contexts
A Forecasting Modeling of Imported Goods Release Waiting Time in Importer Logistics Operations Using Multiple Linear Regression
No full textImport activities play a critical role in international trade, directly affecting logistics efficiency and the competitiveness of importing companies. The process of releasing imported goods at ports often involves complex administrative procedures that can cause delays, leading to increased logistics costs. This study aims to predict the waiting time for the release of imported goods using a machine learning approach. A case study was conducted at PT. Sentra Sarana Logistic, a licensed customs broker responsible for import administration. The primary model applied was Multiple Linear Regression (MLR), and its performance was compared with Neural Network (NN) and Support Vector Machine (SVM) algorithms. Several influencing factors were considered, including tax payment time, inspection duration, and inspection status. Evaluation results indicate that the MLR model achieved the best performance, with an RMSE of 0.00653, MAE of 0.00544, and R-squared of 0.99999, demonstrating high prediction accuracy and a strong linear correlation. The SVM model yielded acceptable results (RMSE 0.74107, R-squared 0.98388) but underperformed compared to MLR. The NN model showed the lowest accuracy with RMSE 2.86599, MAE 2.38831, and R-squared 0.69510. The findings suggest that MLR, despite its simplicity, is highly effective for predicting waiting times in import logistics operations. This research not only offers a practical decision-support tool for importers but also contributes to the existing literature on machine learning applications in logistics operations and customs processing.Import activities play a critical role in international trade, directly affecting logistics efficiency and the competitiveness of importing companies. The process of releasing imported goods at ports often involves complex administrative procedures that can cause delays, leading to increased logistics costs. This study aims to predict the waiting time for the release of imported goods using a machine learning approach. A case study was conducted at PT. Sentra Sarana Logistic, a licensed customs broker responsible for import administration. The primary model applied was Multiple Linear Regression (MLR), and its performance was compared with Neural Network (NN) and Support Vector Machine (SVM) algorithms. Several influencing factors were considered, including tax payment time, inspection duration, and inspection status. Evaluation results indicate that the MLR model achieved the best performance, with an RMSE of 0.00653, MAE of 0.00544, and R-squared of 0.99999, demonstrating high prediction accuracy and a strong linear correlation. The SVM model yielded acceptable results (RMSE 0.74107, R-squared 0.98388) but underperformed compared to MLR. The NN model showed the lowest accuracy with RMSE 2.86599, MAE 2.38831, and R-squared 0.69510. The findings suggest that MLR, despite its simplicity, is highly effective for predicting waiting times in import logistics operations. This research not only offers a practical decision-support tool for importers but also contributes to the existing literature on machine learning applications in logistics operations and customs processing
Exploring AI-Based Music Tool Affordances to Enhance Gen Z Creativity: A Study Using Affordance Theory in Papua
Get PDFArtificial intelligence (AI) plays an important role in the world of music, especially for Generation Z who grew up in the digital age. However, understanding of how AI drives musical creativity in the context of music education is still limited, especially in West Papua and Papua. This study aims to analyze the influence of awareness, perception, and intention to use AI on Generation Z\u27s musical creativity and innovation, using the theory of affordance as a theoretical foundation. The study employs a quantitative approach with the Partial Least Squares Structural Equation Modeling (PLS-SEM) method on 219 respondents from West Papua and Papua provinces. The analysis results indicate that the intention to use AI (p = 0.000) and positive perceptions of AI in general (p = 0.000) have a significant influence on AI-based musical creativity, with moderate predictive power (R² = 0.461). This creativity was then found to encourage the use of AI in musical activities (p = 0.000) and the music creation process (p = 0.000), although both showed weak predictive power (R² = 0.292 and R² = 0.251). Conversely, awareness of AI (p = 0.509) and perceptions of AI in the context of music education (p = 0.135) did not significantly influence creativity. These findings suggest that positive intentions and perspectives toward AI are more decisive in driving creativity than awareness levels. Therefore, a contextual approach to digital music education that encourages active exploration is needed to optimize the potential of AI in enhancing creativity.Artificial intelligence (AI) plays an important role in the world of music, especially for Generation Z who grew up in the digital age. However, understanding of how AI drives musical creativity in the context of music education is still limited, especially in West Papua and Papua. This study aims to analyze the influence of awareness, perception, and intention to use AI on Generation Z\u27s musical creativity and innovation, using the theory of affordance as a theoretical foundation. The study employs a quantitative approach with the Partial Least Squares Structural Equation Modeling (PLS-SEM) method on 219 respondents from West Papua and Papua provinces. The analysis results indicate that the intention to use AI (p = 0.000) and positive perceptions of AI in general (p = 0.000) have a significant influence on AI-based musical creativity, with moderate predictive power (R² = 0.461). This creativity was then found to encourage the use of AI in musical activities (p = 0.000) and the music creation process (p = 0.000), although both showed weak predictive power (R² = 0.292 and R² = 0.251). Conversely, awareness of AI (p = 0.509) and perceptions of AI in the context of music education (p = 0.135) did not significantly influence creativity. These findings suggest that positive intentions and perspectives toward AI are more decisive in driving creativity than awareness levels. Therefore, a contextual approach to digital music education that encourages active exploration is needed to optimize the potential of AI in enhancing creativity
Machine Learning-Based Sentiment Analysis on Twitter (X): A Case Study of the “Kabur Aja Dulu” Issue Using SVM
Get PDFThis study aims to analyze public sentiment toward the phenomenon of “Kabur Aja Dulu” on Twitter (X) using the Support Vector Machine (SVM) method. The data used consists of 4,768 Indonesian-language tweets collected through web scraping. The pre-processing process includes data cleaning, tokenization, stemming, and translation into English for automatic sentiment labeling using TextBlob. The data is then classified into three sentiment categories: positive, negative, and neutral. To address class imbalance, the SMOTE method is applied to the training data, along with TF-IDF techniques for feature extraction. The model was evaluated using the K-Fold Cross Validation method and Grid Search for hyperparameter tuning. The results of the study show that the SVM model with a linear kernel and parameter C=10 provides the best performance with an accuracy value of 85.56%, precision of 845.19%, recall of 85.56%, and F1-score of 85.30%. The main finding of this study is that the linear SVM method is capable of classifying sentiment well, particularly for neutral sentiment data, and has proven effective as an approach to sentiment analysis in the context of social media using the Indonesian language.This study aims to analyze public sentiment toward the phenomenon of “Kabur Aja Dulu” on Twitter (X) using the Support Vector Machine (SVM) method. The data used consists of 4,768 Indonesian-language tweets collected through web scraping. The pre-processing process includes data cleaning, tokenization, stemming, and translation into English for automatic sentiment labeling using TextBlob. The data is then classified into three sentiment categories: positive, negative, and neutral. To address class imbalance, the SMOTE method is applied to the training data, along with TF-IDF techniques for feature extraction. The model was evaluated using the K-Fold Cross Validation method and Grid Search for hyperparameter tuning. The results of the study show that the SVM model with a linear kernel and parameter C=10 provides the best performance with an accuracy value of 85.56%, precision of 845.19%, recall of 85.56%, and F1-score of 85.30%. The main finding of this study is that the linear SVM method is capable of classifying sentiment well, particularly for neutral sentiment data, and has proven effective as an approach to sentiment analysis in the context of social media using the Indonesian language
Few-Shot Learning for Classifying Genuine and Bot Comments on YouTube Using Transformer Models
Get PDFThis study aims to develop a comment classification system on the YouTube platform to distinguish between real accounts and bot accounts, addressing the challenge of limited labeled data through a few-shot learning approach. The issue of bot accounts masquerading as real users in comment sections is becoming increasingly prevalent and has the potential to spread spam, misinformation, and influence public opinion. In this study, a Transformer-based model, DistilBERT, is used, which is known for its efficiency in understanding natural language context. The model is trained in a few-shot scenario (N5 to N50) using a very limited amount of training data. Testing results show that the model maintains high and stable performance even with minimal data (N5), achieving an F1-score above 0.90. In addition, this system is implemented into a web application using Flask to enable direct and interactive comment detection. The main contribution of this research is the proof that the combination of few-shot learning and the DistilBERT model can provide a practical and efficient solution for classifying YouTube bot account comments even with limited data conditions, as well as providing a replicable approach for similar problems on other digital platforms
Implementation of Convolutional Neural Networks (CNN) for Breast Cancer Detection Using ResNet18 Architecture
Get PDFEarly detection of breast cancer is crucial for improving patient survival rates. This study implements a Convolutional Neural Network (CNN) architecture based on ResNet18 using a transfer learning approach to classify breast ultrasound (USG) images into three categories: normal, benign, and malignant. The dataset, comprising 1,578 grayscale images collected from Baheya Hospital in Egypt, underwent preprocessing steps including image conversion, normalization, and augmentation. The ResNet18 model was fine-tuned using selective layer unfreezing to better adapt to the medical imaging domain. Evaluation was conducted using stratified 5-fold cross-validation and assessed with accuracy, precision, recall, F1-score, and AUC metrics. The best results were achieved by fine-tuning layer2, layer3, and the fully connected layer, yielding 95% accuracy, a macro F1-score of 0.93, and an AUC of 0.9906. The findings demonstrate that ResNet18, when properly fine-tuned with transfer learning, delivers high performance in breast cancer detection via ultrasound and holds strong potential as a reliable clinical decision-support tool.Early detection of breast cancer is crucial for improving patient survival rates. This study implements a Convolutional Neural Network (CNN) architecture based on ResNet18 using a transfer learning approach to classify breast ultrasound (USG) images into three categories: normal, benign, and malignant. The dataset, comprising 1,578 grayscale images collected from Baheya Hospital in Egypt, underwent preprocessing steps including image conversion, normalization, and augmentation. The ResNet18 model was fine-tuned using selective layer unfreezing to better adapt to the medical imaging domain. Evaluation was conducted using stratified 5-fold cross-validation and assessed with accuracy, precision, recall, F1-score, and AUC metrics. The best results were achieved by fine-tuning layer2, layer3, and the fully connected layer, yielding 95% accuracy, a macro F1-score of 0.93, and an AUC of 0.9906. The findings demonstrate that ResNet18, when properly fine-tuned with transfer learning, delivers high performance in breast cancer detection via ultrasound and holds strong potential as a reliable clinical decision-support tool
Application of CNN-BiLSTM Algorithm for Ethereum Price Prediction
Get PDFThe volatile and dynamic Ethereum (ETH) market demands an accurate predictive model to support investment decision making. The complexity of ETH time series data and the influence of various external factors make price prediction a challenge in itself. This study aims to develop an ETH price prediction model using a combined architecture of Convolutional Neural Network (CNN) and also Bidirectional Long Short-Term Memory (BiLSTM). CNN is used to extract local features from historical ETH closing price data, while BiLSTM models bidirectional temporal patterns. The dataset used includes ETH daily price from January 2020 to January 2025, which are obtained from Yahoo Finance and have gone through a normalization process and transformation into sequential form. The model is trained for 100 epochs with an early stopping mechanism to prevent overfitting and evaluated using the MAPE and coefficient of determination (R²) metrics. The evaluation results show that the CNN-BiLSTM model is able to predict ETH prices with a MAPE value of 2.8546% and an R² of 0.9415, indicating high performance in capturing actual data trends. This study shows that the hybrid CNN-BiLSTM approach is effective for Ethereum price prediction.The volatile and dynamic Ethereum (ETH) market demands an accurate predictive model to support investment decision making. The complexity of ETH time series data and the influence of various external factors make price prediction a challenge in itself. This study aims to develop an ETH price prediction model using a combined architecture of Convolutional Neural Network (CNN) and also Bidirectional Long Short-Term Memory (BiLSTM). CNN is used to extract local features from historical ETH closing price data, while BiLSTM models bidirectional temporal patterns. The dataset used includes ETH daily price from January 2020 to January 2025, which are obtained from Yahoo Finance and have gone through a normalization process and transformation into sequential form. The model is trained for 100 epochs with an early stopping mechanism to prevent overfitting and evaluated using the MAPE and coefficient of determination (R²) metrics. The evaluation results show that the CNN-BiLSTM model is able to predict ETH prices with a MAPE value of 2.8546% and an R² of 0.9415, indicating high performance in capturing actual data trends. This study shows that the hybrid CNN-BiLSTM approach is effective for Ethereum price prediction
Development of a Classification Model for Underpriced Issuers Using Machine Learning Algorithms
Get PDFThis study develops a classification model to identify underpriced Initial Public Offering (IPO) issuers so that it can help investor decision-making. Using the CRISP-DM methodology, this research uses a sample of 209 non-banking IPO issuers of the OJK E-IPO platform (since establishment until December 31, 2024). In addressing the problem of class imbalance (161 underpriced and 48 not underpriced), SMOTE was used. The model utilizes nine features: Year-on-Year, IHSG, IPO price, ratio of shares issued, age of the firm, size of the firm, sales growth, Return on Assets (ROA), Debt to Equity Ratio (DER), and Asset Turnover Ratio (ATO). Seven classifier algorithms were compared based on accuracy, precision, sensitivity, recall, F1-score, and AUC. Random Forest had the best performance with 89.2% accuracy, 88.9% Macro Average F1-score, and an AUC of 0.946. The findings suggest that the Random Forest model accurately identifies underpriced IPO issuers as a good investment decision-making tool. This research demonstrates that machine learning concepts can be implemented to classify underpriced issuers in Indonesia, continuing previous studies that contributed to understanding the correlation and significance of certain variables to underpricing.This study develops a classification model to identify underpriced Initial Public Offering (IPO) issuers so that it can help investor decision-making. Using the CRISP-DM methodology, this research uses a sample of 209 non-banking IPO issuers of the OJK E-IPO platform (since establishment until December 31, 2024). In addressing the problem of class imbalance (161 underpriced and 48 not underpriced), SMOTE was used. The model utilizes nine features: Year-on-Year, IHSG, IPO price, ratio of shares issued, age of the firm, size of the firm, sales growth, Return on Assets (ROA), Debt to Equity Ratio (DER), and Asset Turnover Ratio (ATO). Seven classifier algorithms were compared based on accuracy, precision, sensitivity, recall, F1-score, and AUC. Random Forest had the best performance with 89.2% accuracy, 88.9% Macro Average F1-score, and an AUC of 0.946. The findings suggest that the Random Forest model accurately identifies underpriced IPO issuers as a good investment decision-making tool. This research demonstrates that machine learning concepts can be implemented to classify underpriced issuers in Indonesia, continuing previous studies that contributed to understanding the correlation and significance of certain variables to underpricing
Implementation of Convolutional Neural Network in Image-Based Waste Classification
Get PDFThe increasingly complex issue of waste management, particularly in the sorting process, demands efficient and accurate technology-based solution. This study aims to implement the Convolutional Neural Network (CNN) method for image-based waste classification, focusing on two classes paper and plastic. The dataset used consists of 2000 images, with an 80% proportion for training and 20% for testing. This study tested four scenarios combining image augmentation and classification methods, namely threshold and one-hot encoding, and evaluated model performance using accuracy, precision, recall, and F1-score metrics. The best results were obtained in the scenario using image augmentation with the one-hot encoding classification method, with an accuracy of 89%, precision of 88.5%, recall of 89%, and F1-score of 88.5%. These findings indicate that implementation of CNN can enhance the effectiveness of image-based waste classification and support recycling efforts through a smarter and more automated sorting system
Classification of Nutritional Status Using the Fuzzy Mamdani Method : Case Study at Banjar City Hospital
Get PDFThe problem of nutritional status in adults requires accurate and adaptive classification methods. This study aims to develop a decision support system using the Fuzzy Mamdani method to classify nutritional status based on Body Mass Index (BMI). A dataset consisting of 237 anthropometric records from Banjar City Regional General Hospital was utilized. The system applies five fuzzy rules to map BMI values into nutritional categories: malnutrition, underweight, normal, overweight, and obesity. The classification process involves fuzzification, inference, and defuzzification using the centroid method. System performance evaluation shows an overall accuracy of 91.13%, with the highest classification precision achieved in the normal category (98.54%) and the lowest in the malnutrition category (30.77%). The results demonstrate that the Fuzzy Mamdani method is effective for nutritional classification, although refinement is needed for underrepresented categories. This system can serve as a useful tool for supporting clinical decision-making in public health services.The problem of nutritional status in adults requires accurate and adaptive classification methods. This study aims to develop a decision support system using the Fuzzy Mamdani method to classify nutritional status based on Body Mass Index (BMI). A dataset consisting of 237 anthropometric records from Banjar City Regional General Hospital was utilized. The system applies five fuzzy rules to map BMI values into nutritional categories: malnutrition, underweight, normal, overweight, and obesity. The classification process involves fuzzification, inference, and defuzzification using the centroid method. System performance evaluation shows an overall accuracy of 91.13%, with the highest classification precision achieved in the normal category (98.54%) and the lowest in the malnutrition category (30.77%). The results demonstrate that the Fuzzy Mamdani method is effective for nutritional classification, although refinement is needed for underrepresented categories. This system can serve as a useful tool for supporting clinical decision-making in public health services