Jurnal Politeknik Negeri Batam (PoliBatam)
Not a member yet
3001 research outputs found
Sort by
Clustering of Aquaculture Productivity Villages in East Aceh Using the K-Means Algorithm
No full textThis study aims to classify villages based on the level of pond utilization and to develop a web-based application for categorizing aquaculture areas in East Aceh Regency. In contrast to traditional definitions based on harvest volume, this research defines productivity functionally—whether the pond area is actively managed or abandoned. The dataset consists of 146 villages and includes five primary variables: number of fish farmers, total pond area, number of pond plots, productive pond area, and abandoned pond area. Clustering was conducted using the K-Means algorithm, resulting in two main groups: productive and non-productive villages. Validation through the Silhouette Score revealed that using k = 2 yielded the highest score of 0.7576, indicating the most optimal clustering structure. The analysis showed that 92% of villages were categorized as productive, while 8% fell into the non-productive cluster. These two clusters differ significantly in terms of land utilization ratios and the number of active aquaculture workers. The findings not only offer a more refined spatial insight but also serve as a basis for the Department of Marine Affairs and Fisheries in formulating aquaculture zoning, revitalization programs, and more targeted resource allocation.This study aims to classify villages based on the level of pond utilization and to develop a web-based application for categorizing aquaculture areas in East Aceh Regency. In contrast to traditional definitions based on harvest volume, this research defines productivity functionally—whether the pond area is actively managed or abandoned. The dataset consists of 146 villages and includes five primary variables: number of fish farmers, total pond area, number of pond plots, productive pond area, and abandoned pond area. Clustering was conducted using the K-Means algorithm, resulting in two main groups: productive and non-productive villages. Validation through the Silhouette Score revealed that using k = 2 yielded the highest score of 0.7576, indicating the most optimal clustering structure. The analysis showed that 92% of villages were categorized as productive, while 8% fell into the non-productive cluster. These two clusters differ significantly in terms of land utilization ratios and the number of active aquaculture workers. The findings not only offer a more refined spatial insight but also serve as a basis for the Department of Marine Affairs and Fisheries in formulating aquaculture zoning, revitalization programs, and more targeted resource allocation
Addition of Non-Skin Classes in Skin Type Classification Using EfficientNet-B0 Architecture
No full textSkin type classification is an essential process in dermatology and skincare, aiming to categorize skin conditions such as dry, normal, and oily. However, image-based skin classification models often struggle when confronted with non-skin objects like clothing, background, or hair that are not accounted for in standard datasets. This study proposes a novel approach by integrating a nonskin class into a skin type classification model based on the EfficientNet-B0 architecture. The dataset used consists of images categorized into four classes: dry, normal, oily, and nonskin. The model was trained using transfer learning and optimized through techniques such as data augmentation, learning rate scheduling, and early stopping. The final evaluation achieved an accuracy of 91%, with the nonskin class showing perfect precision and recall. These results demonstrate that incorporating nonskin data can significantly enhance model robustness and accuracy. This research contributes a practical method for improving the reliability of skin classification systems, especially in real-world environments.Skin type classification is an essential process in dermatology and skincare, aiming to categorize skin conditions such as dry, normal, and oily. However, image-based skin classification models often struggle when confronted with non-skin objects like clothing, background, or hair that are not accounted for in standard datasets. This study proposes a novel approach by integrating a nonskin class into a skin type classification model based on the EfficientNet-B0 architecture. The dataset used consists of images categorized into four classes: dry, normal, oily, and nonskin. The model was trained using transfer learning and optimized through techniques such as data augmentation, learning rate scheduling, and early stopping. The final evaluation achieved an accuracy of 91%, with the nonskin class showing perfect precision and recall. These results demonstrate that incorporating nonskin data can significantly enhance model robustness and accuracy. This research contributes a practical method for improving the reliability of skin classification systems, especially in real-world environments
Public Sentiment Analysis on Corruption Issues in Indonesia Using IndoBERT Fine-Tuning, Logistic Regression, and Linear SVM
No full textSentiment analysis is a method in Natural Language Processing (NLP) that aims to understand public perceptions based on textual data from social media. Opinions expressed in digital platforms play an important role as they reflect public trust and attitudes toward strategic issues in Indonesia. This study aims to compare the performance of three IndoBERT-based approaches for sentiment classification, namely IndoBERT with full fine-tuning, IndoBERT as a feature extractor combined with Logistic Regression, and IndoBERT as a feature extractor combined with Linear SVM. The dataset was collected through the Twitter API, consisting of 2,012 tweets, which after preprocessing and balancing resulted in 2,252 labeled data for positive and negative sentiments. The preprocessing stage included cleansing, normalization, tokenization, stopword removal, and stemming. The dataset was then split into 80% training data, 10% validation data, and 10% testing data. Experimental results show that IndoBERT with full fine-tuning achieved the best performance, with an accuracy of 82.67%, an F1-score of 82.35%, and an AUC value of 0.87. Logistic Regression and Linear SVM produced lower accuracies of 80.20% and 78.22%, respectively. These findings indicate that fine-tuned IndoBERT is more effective in capturing the semantic nuances of the Indonesian language, while the non fine-tuning approaches offer better computational efficiency at the cost of reduced accuracy. This study contributes to the development of NLP methods for the Indonesian language, particularly in sentiment analysis, and highlights the potential of transformer-based models for analyzing strategic issues in social media.Sentiment analysis is a method in Natural Language Processing (NLP) that aims to understand public perceptions based on textual data from social media. Opinions expressed in digital platforms play an important role as they reflect public trust and attitudes toward strategic issues in Indonesia. This study aims to compare the performance of three IndoBERT-based approaches for sentiment classification, namely IndoBERT with full fine-tuning, IndoBERT as a feature extractor combined with Logistic Regression, and IndoBERT as a feature extractor combined with Linear SVM. The dataset was collected through the Twitter API, consisting of 2,012 tweets, which after preprocessing and balancing resulted in 2,252 labeled data for positive and negative sentiments. The preprocessing stage included cleansing, normalization, tokenization, stopword removal, and stemming. The dataset was then split into 80% training data, 10% validation data, and 10% testing data. Experimental results show that IndoBERT with full fine-tuning achieved the best performance, with an accuracy of 82.67%, an F1-score of 82.35%, and an AUC value of 0.87. Logistic Regression and Linear SVM produced lower accuracies of 80.20% and 78.22%, respectively. These findings indicate that fine-tuned IndoBERT is more effective in capturing the semantic nuances of the Indonesian language, while the non fine-tuning approaches offer better computational efficiency at the cost of reduced accuracy. This study contributes to the development of NLP methods for the Indonesian language, particularly in sentiment analysis, and highlights the potential of transformer-based models for analyzing strategic issues in social media
Enhancing E-Commerce Customer Segmentation with Fuzzy C-Means Soft Clustering Probabilities
No full textCustomer segmentation is of paramount importance in the e-commerce industry, enabling businesses to improve marketing strategies and customer engagement. This study compares the performance of two clustering algorithms, K-Means and Fuzzy C-Means (FCM), using Walmart’s public e-commerce dataset of 550,068 transactions. After preprocessing and normalization, the elbow method was applied to determine the optimal number of clusters, yielding seven clusters for K-Means and eight for FCM. Experimental evaluation based on the silhouette score shows that FCM achieved 0.48, outperforming K-Means which scored 0.36, indicating that FCM generated clusters with stronger cohesion and separation. However, this improvement comes at a computational cost. K-Means consistently required less than 0.02 seconds per run, while FCM averaged 0.3 seconds and peaked at 1.38 seconds when the number of clusters increased, making it approximately 20–30 times slower. Cluster distribution analysis further revealed that K-Means produced an uneven segmentation dominated by a single large cluster, whereas FCM generated a more balanced distribution across its clusters. This demonstrates the advantage of FCM in capturing overlapping and multidimensional customer behaviors through partial memberships, in contrast to the rigid and oversimplify assignments of K-Means. These findings highlight the benefit of adopting FCM for e-commerce segmentation, as it provides more interpretable and actionable insights for personalized marketing. At the same time, the trade-off between clustering quality and computation time suggests that future research should explore optimization techniques such as parallelization, approximate fuzzy clustering, or hybrid models that combine the efficiency of hard clustering with the interpretability of soft clustering
Comparative Analysis Transfer Learning Models for Early Detection of Pneumonia using Chest X-ray Images
No full textPneumonia is a serious respiratory disease that continues to be a major worldwide health issue, especially in nations that are struggling with limited medical resources. Early and accurate detection is essential to improve patient outcomes and reducing the rate of death. This study compares the performance of two deep learning architectures, DenseNet121 and ResNet50, using transfer learning for pneumonia detection from chest X-ray images. The dataset consists 5,856 images with two classes, NORMAL and PNEUMONIA, split into training 60%, validation 20%, and testing 20%. Pretrained ImageNet weights were used as fixed feature extractors, with a custom classification layers. Evaluation metrics included accuracy, precision, recall, F1-score, and confusion matrix. On the internal test set, DenseNet121 achieved 92% accuracy, with precision 0.79, recall 0.94, and F1-score 0.86 for NORMAL class, and precision 0.98, recall 0.91, and F1-score 0.94 for PNEUMONIA class. ResNet50 reached 81% accuracy, with precision 0.61, recall 0.80, and F1-score 0.70 for NORMAL class, and precision 0.92, recall 0.81, and F1-score 0.86 for PNEUMONIA class. External testing on an independent set of 200 images (100 images per class) yielded 98% accuracy for DenseNet121 and 85% for ResNet50. These results show that DenseNet121 provides better overall performance and lower false-negative risk for pneumonia cases, highlight the potential of DenseNet121 as a foundation for AI-assisted diagnostic tools in clinical practice
Sentiment Analysis of Coretax on Social Media X Using Naive Bayes, SVM, and LSTM for Service Improvement
No full textIn January 2025, Indonesia’s Ministry of Finance launched Coretax to replace DJP Online. However, the launch triggered widespread dissatisfaction among users, reflecting negative public sentiment. This study aims to analyze public perception of Coretax and evaluate the performance of machine learning models in sentiment classification. A total of 6.036 Indonesian language tweets related to Coretax, posted between January and April 2025, were collected using Tweet Harvest. The dataset consists of 0,83% positive, 51,05% negative, and 48,11% neutral sentiments. The research methodology involved several stages: data crawling, manual labeling, preprocessing (cleaning, case folding, stopword removal, tokenization, normalization, stemming, and specifically for LSTM: conversion of tokens into numerical indices, padding, and embedding), feature representation using TF-IDF for classical models and word embedding for deep learning, data balancing with SMOTE, model implementation (Naive Bayes, Support Vector Machine with various kernels, and LSTM), model evaluation and comparison, and visualization through word clouds. The application of SMOTE succeeded in improving the performance of all algorithms. After applying SMOTE, the SVM with the RBF kernel achieved the best performance with 90,70% accuracy, 91% precision, 90,66% recall, and 90,66% F1-score. Keyword analysis revealed that terms such as “data” and “mudah” dominated positive sentiment, “silakan” and “kakak” were prevalent in neutral sentiment, while “sistem” and “error” frequently appeared in negative sentiment. The findings highlight the urgent need for system infrastructure improvements, user-centered features, responsive technical support, taxpayer training, and continuous updates to enhance Coretax and restore public trust.In January 2025, Indonesia’s Ministry of Finance launched Coretax to replace DJP Online. However, the launch triggered widespread dissatisfaction among users, reflecting negative public sentiment. This study aims to analyze public perception of Coretax and evaluate the performance of machine learning models in sentiment classification. A total of 6.036 Indonesian language tweets related to Coretax, posted between January and April 2025, were collected using Tweet Harvest. The dataset consists of 0,83% positive, 51,05% negative, and 48,11% neutral sentiments. The research methodology involved several stages: data crawling, manual labeling, preprocessing (cleaning, case folding, stopword removal, tokenization, normalization, stemming, and specifically for LSTM: conversion of tokens into numerical indices, padding, and embedding), feature representation using TF-IDF for classical models and word embedding for deep learning, data balancing with SMOTE, model implementation (Naive Bayes, Support Vector Machine with various kernels, and LSTM), model evaluation and comparison, and visualization through word clouds. The application of SMOTE succeeded in improving the performance of all algorithms. After applying SMOTE, the SVM with the RBF kernel achieved the best performance with 90,70% accuracy, 91% precision, 90,66% recall, and 90,66% F1-score. Keyword analysis revealed that terms such as “data” and “mudah” dominated positive sentiment, “silakan” and “kakak” were prevalent in neutral sentiment, while “sistem” and “error” frequently appeared in negative sentiment. The findings highlight the urgent need for system infrastructure improvements, user-centered features, responsive technical support, taxpayer training, and continuous updates to enhance Coretax and restore public trust
Sentiment Analysis on Google Play Store Reviews to Measure User Perception of the Gojek Application Using CNN
No full textThis study was conducted to analyze sentiment towards user reviews from the Google Play Store regarding the Gojek application. The analysis aims to measure user perceptions using a Convolutional Neural Network (CNN). This study aims to understand user views on the Gojek application. By understanding user perceptions, the information obtained can be utilized by the company\u27s service team to improve the quality of the application for users. User perceptions are grouped into three labels: positive, neutral, and negative. To produce an effective model, this study uses three data sharing ratios simultaneously with the same parameters: 90:10, 80:20, and 70:30. Due to the large amount of data, random sampling is needed to balance the data and thus increase accuracy in the data processing process. Model evaluation was carried out using a confusion matrix, precision, recall, and F1-Score. The results obtained with the highest accuracy of 84.29%. This study successfully demonstrates that CNN is able to process user review data well.This study was conducted to analyze sentiment towards user reviews from the Google Play Store regarding the Gojek application. The analysis aims to measure user perceptions using a Convolutional Neural Network (CNN). This study aims to understand user views on the Gojek application. By understanding user perceptions, the information obtained can be utilized by the company\u27s service team to improve the quality of the application for users. User perceptions are grouped into three labels: positive, neutral, and negative. To produce an effective model, this study uses three data sharing ratios simultaneously with the same parameters: 90:10, 80:20, and 70:30. Due to the large amount of data, random sampling is needed to balance the data and thus increase accuracy in the data processing process. Model evaluation was carried out using a confusion matrix, precision, recall, and F1-Score. The results obtained with the highest accuracy of 84.29%. This study successfully demonstrates that CNN is able to process user review data well
Improving YOLO Performance with Advanced Data Augmentation for Soccer Object Detection
No full textThis study developed an object detection system for soccer games using the YOLOv8m algorithm with four main classes: player, goalkeeper, referee, and ball. The dataset, consisting of 372 annotated images, exhibited class imbalance, with significantly fewer ball instances compared to players. The basic YOLOv8m architecture was used without internal modifications, but adjustments were made to the output layer and fine-tuning of the pre-trained weights to adapt to the new dataset. Two models were compared: one without and one with advanced augmentation techniques (mosaic, mixup, cutmix). The experimental results showed an increase in mAP@50 from 74.9% to 81.4% in the augmented model, with a statistically significant difference (p < 0.01). However, model performance still decreased under extreme conditions such as high occlusion, rapid movement, and uneven lighting. The combination of data augmentation, output layer adaptation, and fine-tuning proved effective in improving object detection accuracy and provided the basis for the development of a real-time artificial intelligence-based soccer match analysis system.This study developed an object detection system for soccer games using the YOLOv8m algorithm with four main classes: player, goalkeeper, referee, and ball. The dataset, consisting of 372 annotated images, exhibited class imbalance, with significantly fewer ball instances compared to players. The basic YOLOv8m architecture was used without internal modifications, but adjustments were made to the output layer and fine-tuning of the pre-trained weights to adapt to the new dataset. Two models were compared: one without and one with advanced augmentation techniques (mosaic, mixup, cutmix). The experimental results showed an increase in mAP@50 from 74.9% to 81.4% in the augmented model, with a statistically significant difference (p < 0.01). However, model performance still decreased under extreme conditions such as high occlusion, rapid movement, and uneven lighting. The combination of data augmentation, output layer adaptation, and fine-tuning proved effective in improving object detection accuracy and provided the basis for the development of a real-time artificial intelligence-based soccer match analysis system
Detection of Misoriented Polarized Electronic Components on PCBs Using HOG Features and Neural Networks
No full textMounting misorientation on polar electronic components in printed circuit boards (PCBs) can cause malfunctions in electronic devices. This study proposes an automatic detection system that utilizes the Histogram of Oriented Gradients (HOG) feature and employs classification using an artificial neural network. The research was conducted by collecting data from PCB images featuring polar components, such as diodes, electrolytic capacitors, and transistors. Once the components are identified, the HOG features are extracted to generate feature vectors used in artificial neural network training. The experiment results show that this system can detect component orientation errors with a high degree of accuracy, achieving accuracy values of 99.5% for transistor components, 97% for electrolyte capacitors, and 93.6% for diodes. Additionally, F1 values and high precision are achieved for all three types of components. The ReLU activation function has been shown to perform best among other activation functions. While the results are promising, further research is necessary to automate the identification of component locations without relying on manual cropping processes
Benchmarking Deepseek-LLM-7B-Chat and Qwen1.5-7B-Chat for Indonesian Product Review Emotion Classification
No full textUpon completing their shopping experience on an e-commerce platform, users have the opportunity to leave a review. By analyzing reviews, businesses can gain insight into customer emotions, while researchers and policymakers can monitor social dynamics. Large Language Models (LLMs) utilization is identified as a promising methodology for emotion analysis. LLMs have revolutionized natural language processing capabilities, yet their performance in non-English languages, such as Indonesian, necessitates a comprehensive evaluation. This research objective is to perform a comprehensive analysis and comparison of Deepseek-LLM-7B-Chat and Qwen1.5-7B-Chat, two prominent open-source Large Language Models, for the emotion classification of Indonesian product reviews. By leveraging the PRDECT-ID dataset, this study evaluates the performance of both models in a few-shot learning scenario through prompt engineering. The methodology outlines the data preprocessing pipeline, a detailed few-shot prompt engineering strategy tailored to each model\u27s characteristics, model inference execution, and performance assessment using the accuracy, precision, recall, and F1-score metrics. Analytical results reveal DeepSeek achieved an accuracy of 43.41%, exhibiting a considerably superior ability to comprehend instructions compared to Qwen, which attained a maximum accuracy of only 20.35% and often yielded near-random predictions. An in-depth error analysis indicates that this performance gap is likely attributable to factors such as pre-training data bias and tokenization mismatches with the Indonesian language. This research offers empirical evidence regarding the comparative strengths and weaknesses of DeepSeek and Qwen, providing a diagnostic benchmark that underscores the significance of instruction tuning and robust multilingual representation for Indonesian NLP tasks