Jurnal Politeknik Negeri Batam (PoliBatam)
Not a member yet
3001 research outputs found
Sort by
Application of ADASYN and Optuna in the XGBoost Algorithm for Stunting Detection
No full textThis study aims to develop an early detection model for childhood stunting risk using a machine learning approach based on Extreme Gradient Boosting (XGBoost), integrated with the Adaptive Synthetic Sampling (ADASYN) technique for data balancing and Optuna-based hyperparameter optimization. One of the main challenges in stunting prediction is class imbalance, where the number of stunting cases is significantly higher than non-stunting cases, thereby reducing the model’s ability to accurately identify the minority class. To address this issue, the study implements data deduplication, structured data splitting, and applies ADASYN exclusively to the training data to prevent data leakage and preserve the validity of the evaluation process. The proposed model (XGBoost with ADASYN and Optuna) is then compared with a baseline model that combines XGBoost and SMOTE. Experimental results show that the proposed model achieves an accuracy of 81.98%, a recall of 91.50%, and an F1-score of 89.14%, indicating improved sensitivity and a more balanced classification performance compared to the baseline. These findings demonstrate that the integration of ADASYN and Optuna-based hyperparameter optimization enhances model stability and generalization capability, making it a viable data-driven approach for stunting risk detection in environments with imbalanced class distributions.This study aims to develop an early detection model for childhood stunting risk using a machine learning approach based on Extreme Gradient Boosting (XGBoost), integrated with the Adaptive Synthetic Sampling (ADASYN) technique for data balancing and Optuna-based hyperparameter optimization. One of the main challenges in stunting prediction is class imbalance, where the number of stunting cases is significantly higher than non-stunting cases, thereby reducing the model’s ability to accurately identify the minority class. To address this issue, the study implements data deduplication, structured data splitting, and applies ADASYN exclusively to the training data to prevent data leakage and preserve the validity of the evaluation process. The proposed model (XGBoost with ADASYN and Optuna) is then compared with a baseline model that combines XGBoost and SMOTE. Experimental results show that the proposed model achieves an accuracy of 81.98%, a recall of 91.50%, and an F1-score of 89.14%, indicating improved sensitivity and a more balanced classification performance compared to the baseline. These findings demonstrate that the integration of ADASYN and Optuna-based hyperparameter optimization enhances model stability and generalization capability, making it a viable data-driven approach for stunting risk detection in environments with imbalanced class distributions
Optimizing Feature Extraction for Naïve Bayes Sentiment Analysis
No full textThe rapid growth of e-commerce platforms such as Tokopedia has generated a large volume of user reviews containing diverse opinions about products and services. These reviews reflect consumer perceptions and provide valuable insights for business decision-making. This study aims to enhance sentiment analysis performance by optimizing the Naïve Bayes algorithm through a comparison of two feature extraction techniques, namely Bag of Words (BoW) and Term Frequency–Inverse Document Frequency (TF-IDF). The dataset consists of 5,400 Tokopedia product reviews obtained from the Kaggle platform, which are categorized into positive and negative sentiments. The research process includes text preprocessing consisting of text cleaning, case folding, tokenization, stopword removal, and stemming, feature extraction using Bag of Words (BoW) and Term Frequency–Inverse Document Frequency (TF-IDF), handling data imbalance using the Synthetic Minority Over-sampling Technique (SMOTE), and model training using the Naïve Bayes. The dataset is divided into 80% training data and 20% testing data, and model performance is evaluated using accuracy, precision, recall, and F1-score. The results show that BoW achieved the highest accuracy of 93%, while TF-IDF reached 83%, indicating that BoW provides more effective feature representation and more stable performance for Naïve Bayes-based sentiment analysis on this dataset.The rapid growth of e-commerce platforms such as Tokopedia has generated a large volume of user reviews containing diverse opinions about products and services. These reviews reflect consumer perceptions and provide valuable insights for business decision-making. This study aims to enhance sentiment analysis performance by optimizing the Naïve Bayes algorithm through a comparison of two feature extraction techniques, namely Bag of Words (BoW) and Term Frequency–Inverse Document Frequency (TF-IDF). The dataset consists of 5,400 Tokopedia product reviews obtained from the Kaggle platform, which are categorized into positive and negative sentiments. The research process includes text preprocessing consisting of text cleaning, case folding, tokenization, stopword removal, and stemming, feature extraction using Bag of Words (BoW) and Term Frequency–Inverse Document Frequency (TF-IDF), handling data imbalance using the Synthetic Minority Over-sampling Technique (SMOTE), and model training using the Naïve Bayes. The dataset is divided into 80% training data and 20% testing data, and model performance is evaluated using accuracy, precision, recall, and F1-score. The results show that BoW achieved the highest accuracy of 93%, while TF-IDF reached 83%, indicating that BoW provides more effective feature representation and more stable performance for Naïve Bayes-based sentiment analysis on this dataset
Analysis of SMOTE and Random Search on Machine Learning Algorithms for Stroke Disease Diagnosis
No full textStroke is a critical medical condition in which false negative predictions may lead to delayed treatment and increased mortality. Therefore, predictive models in the medical domain should prioritize sensitivity (recall) in addition to overall accuracy. This study analyzes the impact of the Synthetic Minority Over-sampling Technique (SMOTE) and Random Search hyperparameter optimization on five machine learning algorithms—Random Forest, XGBoost, Support Vector Machine (SVM), Logistic Regression, and CatBoost—for stroke disease diagnosis. Two experimental scenarios were conducted, namely models trained without SMOTE and models trained with SMOTE applied only to the training data to prevent data leakage. Model performance was evaluated using accuracy, precision, recall, and F1-score, with particular emphasis on recall due to its clinical relevance. In clinical practice, low recall may lead to false negative predictions, where high-risk stroke patients are not identified by the system, potentially resulting in delayed medical intervention. Therefore, recall is emphasized as the primary performance metric in this study. Experimental results demonstrate that SMOTE consistently improves recall across all models, while Random Search further enhances performance. CatBoost achieved the best performance with an accuracy of 96.61%, recall of 97%, and F1-score of 97%. Despite its superior performance, potential overfitting risks are critically discussed. These findings indicate that the proposed approach produces a clinically relevant decision-support model for stroke risk prediction.Stroke is a critical medical condition in which false negative predictions may lead to delayed treatment and increased mortality. Therefore, predictive models in the medical domain should prioritize sensitivity (recall) in addition to overall accuracy. This study analyzes the impact of the Synthetic Minority Over-sampling Technique (SMOTE) and Random Search hyperparameter optimization on five machine learning algorithms—Random Forest, XGBoost, Support Vector Machine (SVM), Logistic Regression, and CatBoost—for stroke disease diagnosis. Two experimental scenarios were conducted, namely models trained without SMOTE and models trained with SMOTE applied only to the training data to prevent data leakage. Model performance was evaluated using accuracy, precision, recall, and F1-score, with particular emphasis on recall due to its clinical relevance. In clinical practice, low recall may lead to false negative predictions, where high-risk stroke patients are not identified by the system, potentially resulting in delayed medical intervention. Therefore, recall is emphasized as the primary performance metric in this study. Experimental results demonstrate that SMOTE consistently improves recall across all models, while Random Search further enhances performance. CatBoost achieved the best performance with an accuracy of 96.61%, recall of 97%, and F1-score of 97%. Despite its superior performance, potential overfitting risks are critically discussed. These findings indicate that the proposed approach produces a clinically relevant decision-support model for stroke risk prediction
Evaluation of Histogram-Based Image Enhancement Methods for Facial Images in Drowsy Driver Using No-Reference Metrics
No full textLow-light facial images suffer significant quality degradation, leading to performance degradation in surveillance and face recognition systems, where conventional enhancement methods often produce over-enhancement or unnatural noise artifacts. This study compares three histogram equalization methods, namely HE, AHE, and CLAHE, for low-light facial image enhancement, with evaluation using no-reference quality assessment metrics, including NIQE, LOE, and Entropy, as well as visual analysis and histogram distribution. The results showed that AHE produced the lowest NIQE (4.96 ± 1.38) and the highest entropy (7.86 ± 0.11) but had significant noise artifacts, HE produced an overly even distribution with NIQE of 6.34 ± 1.41, while CLAHE showed the most balanced performance with the lowest LOE (0.07 ± 0.02) and the best visual quality when using the optimal clip limit in the range of 1.2-2.0, providing an optimal trade-off between contrast enhancement, naturalness preservation, and artifact minimization with computational efficiency below 1 ms
Comparative Analysis of BERT and LSTM Models for Sentiment Classification of Mobile Game User Reviews
No full textSentiment classification of user reviews for mobile games that rely on direct advertising (direct ads) is crucial for understanding player perceptions and improving user experience. This study aims to compare the performance of two deep learning architectures, Long Short-Term Memory (LSTM) and multilingual Bidirectional Encoder Representations from Transformers (BERT) in classifying sentiment in reviews into three categories, positive, negative, and neutral. The dataset used consists of reviews from games employing direct ads, which underwent rule-based labeling and text preprocessing. The LSTM model was built from scratch using a custom embedding layer, while the multilingual BERT model was fine-tuned using a transfer learning approach. Evaluation was conducted based on accuracy, precision, recall, and F1-score metrics. Experimental results show that multilingual BERT achieves superior validation loss compared to LSTM (0.37 vs. 0.44). BERT also outperforms LSTM significantly in terms of F1-score and its ability to understand multilingual linguistic context. However, LSTM demonstrates advantages in computational efficiency and training speed. These findings offer practical recommendations for developers in selecting an appropriate sentiment analysis model based on accuracy requirements and resource availability.Sentiment classification of user reviews for mobile games that rely on direct advertising (direct ads) is crucial for understanding player perceptions and improving user experience. This study aims to compare the performance of two deep learning architectures, Long Short-Term Memory (LSTM) and multilingual Bidirectional Encoder Representations from Transformers (BERT) in classifying sentiment in reviews into three categories, positive, negative, and neutral. The dataset used consists of reviews from games employing direct ads, which underwent rule-based labeling and text preprocessing. The LSTM model was built from scratch using a custom embedding layer, while the multilingual BERT model was fine-tuned using a transfer learning approach. Evaluation was conducted based on accuracy, precision, recall, and F1-score metrics. Experimental results show that multilingual BERT achieves superior validation loss compared to LSTM (0.37 vs. 0.44). BERT also outperforms LSTM significantly in terms of F1-score and its ability to understand multilingual linguistic context. However, LSTM demonstrates advantages in computational efficiency and training speed. These findings offer practical recommendations for developers in selecting an appropriate sentiment analysis model based on accuracy requirements and resource availability
Software Development for Swimmer Performance Prediction System Based on Physical Characteristics using XGBoost
No full textSwimmer performance assessment in Indonesia still largely depends on coaches’ intuition, which may lead to subjective decisions and inconsistencies in training program planning, particularly in environments where frequent changes in coaches and sports administrators occur. The lack of structured and data-driven performance assessment tools further limits the continuity and objectivity of athlete development. This study aims to develop a web-based system capable of predicting swimmers’ performance potential by estimating race times based on physical characteristics using the XGBoost model. The proposed system is designed to support coaches in identifying athlete performance potential in a more objective and data-driven manner. Model evaluation results indicate that the XGBoost model achieved an R² value of 0.9190, demonstrating a very high level of prediction accuracy, with an average prediction time of 7.036 seconds. Software testing results confirm that the system operates as intended and is able to present prediction outputs in the form of estimated swimming time, performance percentage, and performance classification into four categories: Very High, High, Medium, and Low. Furthermore, usability evaluation using the USE method yielded excellent results, with an average score of 88.16%
The Comparative Analysis of K-Nearest Neighbors Algorithm and Random Forest Regressor for House Price Prediction in Bandung City
No full textThe rapid population growth and continuous urban expansion in Bandung have contributed to volatile and escalating housing prices, creating significant challenges for market transparency and affordability. This study aims to develop and evaluate machine-learning models to predict house prices in the Bandung region using a publicly available dataset consisting of 7,609 property records. Following the CRISP-DM methodology, the research includes data exploration, preprocessing (outlier handling using IQR, one-hot encoding, and feature standardization), model training, and performance evaluation. Two regression models K-Nearest Neighbors (KNN) Regressor and Random Forest (RF) Regressor—were compared through systematic hyperparameter tuning using Grid Search and Random Search techniques. The experimental results show that the Random Forest Regressor achieves the best performance with an R² score of 0.7838 and a mean absolute error (MAE) of approximately Rp 399.7 million, outperforming the optimized KNN model. Feature importance analysis also indicates that land area, building area, and location are the most influential predictors of property prices. The findings highlight the effectiveness of ensemble methods in handling complex real-estate data and demonstrate the potential of machine-learning-based predictive tools to support buyers, sellers, and policymakers in making informed and data-driven decisions in the Bandung housing market.The rapid population growth and continuous urban expansion in Bandung have contributed to volatile and escalating housing prices, creating significant challenges for market transparency and affordability. This study aims to develop and evaluate machine-learning models to predict house prices in the Bandung region using a publicly available dataset consisting of 7,609 property records. Following the CRISP-DM methodology, the research includes data exploration, preprocessing (outlier handling using IQR, one-hot encoding, and feature standardization), model training, and performance evaluation. Two regression models K-Nearest Neighbors (KNN) Regressor and Random Forest (RF) Regressor—were compared through systematic hyperparameter tuning using Grid Search and Random Search techniques. The experimental results show that the Random Forest Regressor achieves the best performance with an R² score of 0.7838 and a mean absolute error (MAE) of approximately Rp 399.7 million, outperforming the optimized KNN model. Feature importance analysis also indicates that land area, building area, and location are the most influential predictors of property prices. The findings highlight the effectiveness of ensemble methods in handling complex real-estate data and demonstrate the potential of machine-learning-based predictive tools to support buyers, sellers, and policymakers in making informed and data-driven decisions in the Bandung housing market
IoT Application Development for Marine Debris Management in 3T Islands: Supporting a Circular Economy and Community Empowerment
No full textMarine debris is a serious problem, especially in the outermost, foremost, and least developed (3T) islands of Indonesia, where limited infrastructure and low public awareness are the main obstacles to effective waste management. This study aims to design, develop, and evaluate an Internet of Things (IoT)-based application integrated with a community-based web platform to support circular economy practices and community empowerment in marine debris management. The research method used is Research and Development (R&D) adapted from Borg & Gall, starting from the needs analysis stage to dissemination. An IoT module equipped with ultrasonic and GPS sensors is used to detect container capacity and location in real-time. Performance testing results show a response time of 1.8 seconds, a data transmission success rate of 98.7%, and a capacity detection accuracy of 96.2%, which meets the established technical standards. User acceptance testing using the Technology Acceptance Model (TAM) involving 15 respondents resulted in an average Perceived Usefulness (PU) score of 4.40 and Perceived Ease of Use (PEOU) of 4.23. Pearson\u27s correlation analysis showed an r value of 0.84 (p = 0.0001), indicating a very strong and significant positive relationship between ease of use and perceived usefulness. This finding confirms that the developed system is technically reliable, easy to use, and capable of promoting environmental sustainability and economic opportunities in the 3T island communities
From Speech to Summary: A Pipeline-Based Evaluation of Whisper and Transformer Models for Indonesian Dialogue Summarization
No full textThe rapid increase in online meetings has produced massive amounts of undocumented spoken content, creating a practical need for automatic summarization. For Indonesian, this task is hindered by a dual-faceted resource scarcity and a lack of foundational benchmarks for pipeline components. This paper addresses this gap by creating a new synthetic conversational dataset for Indonesian and conducting two systematic, discrete benchmarks to identify the optimal components for an end-to-end pipeline. First, we evaluated six Whisper ASR model variants (from tiny to turbo) and found a clear, non-obvious winner: the turbo (distil-large-v2) model was not only the most accurate (7.97% WER) but also one of the fastest (1.25s inference), breaking the expected cost-accuracy trade-off. Second, we benchmarked 13 zero-shot summarization models on gold-standard transcripts, which revealed a critical divergence between lexical and semantic performance. Indonesian-specific models excelled at lexical overlap (ROUGE-1: 17.09 for cahya/t5-base...), while the multilingual google/long-t5-tglobal-base model was the clear semantic winner (BERTScore F1: 67.09)
Comparison of Naïve Bayes and Support Vector Machine for Sentiment Classification of Acne Skincare Reviews
No full textThe increasing popularity of skincare products for acne-prone skin had led to a surge in online consumer reviews, which are characterized by informal language, domain-specific terminology, and imbalanced sentiment distribution, posing challenges for sentiment classification tasks. This study aims not only to compare the performance but also to analyze the generalization behavior of two popular machine learning algorithms, Naïve Bayes and Support Vector Machine (SVM), for sentiment classification of skincare product reviews specifically targeting acne-prone skin. A comprehensive methodology was employed, including thorough text preprocessing, feature extraction using Term Frequency-Inverse Document Frequency (TF-IDF) with n-gram representation, and data balancing through Synthetic Minority Over-sampling Technique (SMOTE). The study utilized a dataset of 4,004 labeled reviews categorized into positive and negative sentiments. The models were evaluated using stratified 5-Fold cross-validation to ensure robust and fair assessment. Results indicate that Naïve Bayes slightly outperforms SVM on the testing set, achieving the highest accuracy of 91.14% compared to 90.64% for SVM. While SVM demonstrated higher performance during training, its testing performance suggested a tendency toward overfitting, whereas Naïve Bayes exhibited more stable generalization on unseen data. Further qualitative insight analysis revealed that product effectiveness and user experience are the primary drivers of consumer sentiment, while competitive analysis highlighted distinct brand perception patterns across skincare categories. These findings indicate that simpler probabilistic models such as Naïve Bayes can provide robust and reliable performance for sentiment analysis in specialized and imbalanced skincare review datasets