Jurnal Politeknik Negeri Batam (PoliBatam)
Not a member yet
3001 research outputs found
Sort by
Evaluating the Usability of Canva Among University Students in Pekanbaru Using the WEBUSE Method
No full textCanva is a design platform used to create social media graphics, presentations, posters, documents, and other visual content. This study aims to evaluate user satisfaction with the web-based version of Canva using the WEBUSE method, which covers four main aspects: Content, Organization and Readability, Navigation and Links, User Interface Design, and Performance and Effectiveness. Data were collected through an online questionnaire distributed to university students in the Pekanbaru area via WhatsApp and Instagram. A total of 65 respondents were obtained through the data collection and screening process. The evaluation results show that all aspects fall into the "Good" usability category, with the highest average score in User Interface Design (0.75) and the lowest in Performance and Effectiveness (0.70), resulting in an overall average score of 0.725. Therefore, Canva’s website is considered to have good usability according to user perceptions. This study is expected to serve as input for feature development and service quality improvement of Canva in the future.Canva is a design platform used to create social media graphics, presentations, posters, documents, and other visual content. This study aims to evaluate user satisfaction with the web-based version of Canva using the WEBUSE method, which covers four main aspects: Content, Organization and Readability, Navigation and Links, User Interface Design, and Performance and Effectiveness. Data were collected through an online questionnaire distributed to university students in the Pekanbaru area via WhatsApp and Instagram. A total of 65 respondents were obtained through the data collection and screening process. The evaluation results show that all aspects fall into the "Good" usability category, with the highest average score in User Interface Design (0.75) and the lowest in Performance and Effectiveness (0.70), resulting in an overall average score of 0.725. Therefore, Canva’s website is considered to have good usability according to user perceptions. This study is expected to serve as input for feature development and service quality improvement of Canva in the future
Comparative Analysis of Random Forest and XGBoost Models for Cervical Cancer Risk Prediction using SHAP-based Explainable AI
No full textCervical cancer remains one of the leading causes of cancer-related deaths among women, particularly in developing countries such as Indonesia. This study aims to develop an accurate and interpretable predictive model for cervical cancer risk using Random Forest (RF) and Extreme Gradient Boosting (XGBoost) algorithms. The dataset used is the Cervical Cancer Risk Factors from the UCI Repository, consisting of 858 patient records and 36 clinical and demographic features. The preprocessing stages include missing value imputation, class balancing using Synthetic Minority Oversampling Technique (SMOTE), and hyperparameter optimization through Randomized Search CV. Experimental results show that both models achieved high performance, with accuracy exceeding 96% and AUC above 0.95, while the XGBoost (Tuned + SMOTE) model slightly outperformed RF in detecting positive cases. The interpretability analysis using SHapley Additive exPlanations (SHAP) identified clinical features such as Schiller Test, Hinselmann Test, and Cytology Result as the most influential factors in the classification process, consistent with established clinical evidence. Therefore, the integration of XGBoost, SMOTE, and SHAP provides a predictive framework that is not only highly accurate but also clinically explainable, supporting the development of decision-support systems for early cervical cancer detection.Cervical cancer remains one of the leading causes of cancer-related deaths among women, particularly in developing countries such as Indonesia. This study aims to develop an accurate and interpretable predictive model for cervical cancer risk using Random Forest (RF) and Extreme Gradient Boosting (XGBoost) algorithms. The dataset used is the Cervical Cancer Risk Factors from the UCI Repository, consisting of 858 patient records and 36 clinical and demographic features. The preprocessing stages include missing value imputation, class balancing using Synthetic Minority Oversampling Technique (SMOTE), and hyperparameter optimization through Randomized Search CV. Experimental results show that both models achieved high performance, with accuracy exceeding 96% and AUC above 0.95, while the XGBoost (Tuned + SMOTE) model slightly outperformed RF in detecting positive cases. The interpretability analysis using SHapley Additive exPlanations (SHAP) identified clinical features such as Schiller Test, Hinselmann Test, and Cytology Result as the most influential factors in the classification process, consistent with established clinical evidence. Therefore, the integration of XGBoost, SMOTE, and SHAP provides a predictive framework that is not only highly accurate but also clinically explainable, supporting the development of decision-support systems for early cervical cancer detection
Generative AI Image Sentiment Analysis on Social Media X using TF-IDF and FastText
No full textThis research investigates public opinion on AI-generated images on Social Media X using machine learning-driven text classification. Three classification models were evaluated: Complement Naïve Bayes (CNB) utilizing TF-IDF features, Support Vector Machine (SVM) merging TF-IDF with FastText embeddings, and IndoBERT as a modern transformer-based baseline. A total of 1,958 Indonesian tweets were collected via web scraping with relevant keywords, followed by a pipeline involving text cleaning, manual labeling into positive, negative, and neutral categories, and data balancing using the Synthetic Minority Over-sampling Technique (SMOTE) for the classical models (with class weighting applied for IndoBERT). Results show that the SVM model outperformed the others, achieving 68.7% accuracy with average precision, recall, and F1-score of 0.69, 0.69, and 0.68, respectively; CNB attained 64.1% accuracy with average metrics of 0.64; while IndoBERT recorded 58.2% accuracy with average precision, recall, and F1-score of 0.58, 0.58, and 0.57. Confusion matrix analysis revealed SVM\u27s superior ability to distinguish positive and neutral sentiments in casual language, though IndoBERT demonstrated potential for capturing deeper semantic nuances despite underperforming due to dataset size and informal text. The findings highlight the efficacy of integrating statistical and semantic representations for improved sentiment analysis on unstructured, noisy social media data related to AI-generated imagery, while suggesting that transformer models like IndoBERT may benefit from larger datasets for optimal performance.This research investigates public opinion on AI-generated images on Social Media X using machine learning-driven text classification. Three classification models were evaluated: Complement Naïve Bayes (CNB) utilizing TF-IDF features, Support Vector Machine (SVM) merging TF-IDF with FastText embeddings, and IndoBERT as a modern transformer-based baseline. A total of 1,958 Indonesian tweets were collected via web scraping with relevant keywords, followed by a pipeline involving text cleaning, manual labeling into positive, negative, and neutral categories, and data balancing using the Synthetic Minority Over-sampling Technique (SMOTE) for the classical models (with class weighting applied for IndoBERT). Results show that the SVM model outperformed the others, achieving 68.7% accuracy with average precision, recall, and F1-score of 0.69, 0.69, and 0.68, respectively; CNB attained 64.1% accuracy with average metrics of 0.64; while IndoBERT recorded 58.2% accuracy with average precision, recall, and F1-score of 0.58, 0.58, and 0.57. Confusion matrix analysis revealed SVM\u27s superior ability to distinguish positive and neutral sentiments in casual language, though IndoBERT demonstrated potential for capturing deeper semantic nuances despite underperforming due to dataset size and informal text. The findings highlight the efficacy of integrating statistical and semantic representations for improved sentiment analysis on unstructured, noisy social media data related to AI-generated imagery, while suggesting that transformer models like IndoBERT may benefit from larger datasets for optimal performance
Method Design of an IoT-Based Automatic Pest Repellent System Prototype for Agriculture
No full textIndonesia, as an agricultural country, still faces serious challenges in the farming sector, particularly pest attacks from birds and insects that significantly reduce rice productivity and may lead to crop failure. The use of traditional methods and chemical pesticides is considered ineffective and has negative impacts on health and the environment. This study aims to design a prototype of an automated pest repellent system for agriculture based on the Internet of Things (IoT) that is environmentally friendly, energy-efficient, and easy to operate by local farmers. The research method employed a prototyping approach, which includes problem identification, hardware and software design, testing, and system evaluation. The device consists of a NodeMCU ESP32 microcontroller, a PIR sensor to detect pest movement, relay, ultrasonic speaker, electric net, and solar panel as the main power source. Testing on a miniature rice field model showed that the system could detect pest movement at a distance of approximately 5 meters and automatically activate the ultrasonic speaker with a range of 50–100 meters to repel birds, and the electric net to catch insects at night. Energy consumption is primarily supplied by the solar panel, and a fully charged battery can power the system for about 3 hours without sunlight. The detection success rate reached more than 85% with consistent actuator response. This system has proven to reduce pesticide dependency, is environmentally friendly, and has the potential to increase rice farming efficiency.Indonesia, as an agricultural country, still faces serious challenges in the farming sector, particularly pest attacks from birds and insects that significantly reduce rice productivity and may lead to crop failure. The use of traditional methods and chemical pesticides is considered ineffective and has negative impacts on health and the environment. This study aims to design a prototype of an automated pest repellent system for agriculture based on the Internet of Things (IoT) that is environmentally friendly, energy-efficient, and easy to operate by local farmers. The research method employed a prototyping approach, which includes problem identification, hardware and software design, testing, and system evaluation. The device consists of a NodeMCU ESP32 microcontroller, a PIR sensor to detect pest movement, relay, ultrasonic speaker, electric net, and solar panel as the main power source. Testing on a miniature rice field model showed that the system could detect pest movement at a distance of approximately 5 meters and automatically activate the ultrasonic speaker with a range of 50–100 meters to repel birds, and the electric net to catch insects at night. Energy consumption is primarily supplied by the solar panel, and a fully charged battery can power the system for about 3 hours without sunlight. The detection success rate reached more than 85% with consistent actuator response. This system has proven to reduce pesticide dependency, is environmentally friendly, and has the potential to increase rice farming efficiency
Knowledge Discovery on E-Commerce Customer Churn Using Interpretable Machine Learning: A Comparative Study of SHAP-Based Classifiers
No full textCustomer churn remains one of the most pressing issues in the e-commerce sector, as it directly erodes revenue and reduces customer lifetime value. This study proposes an interpretable machine learning approach designed not only to predict churn but also to uncover practical insights that can inform retention strategies. The analysis draws on a publicly available dataset containing customer behavior and transaction records. Data preparation involved handling missing values, applying label encoding, and addressing class imbalance with SMOTE. Five classification models—Logistic Regression, Random Forest, XGBoost, Support Vector Machine, and Gradient Boosting—were trained on an 80:20 stratified split, with performance assessed through accuracy, precision, recall, F1-score, and AUC. Among these, XGBoost delivered the most consistent results, achieving 96% accuracy, 95% precision, 92% recall, and a near-perfect AUC of 0.999, followed closely by Random Forest. Logistic Regression produced the lowest AUC at 0.886. To ensure transparency in decision-making, SHAP (SHapley Additive exPlanations) was applied, revealing Tenure, Complain, and CashbackAmount as the most influential predictors. Longer customer relationships were linked to reduced churn risk, while frequent complaints and higher cashback usage indicated a greater likelihood of leaving. These findings contribute knowledge by blending robust predictive performance with interpretability, enabling e-commerce businesses to design more targeted and proactive customer retention measures
Sentiment Analysis of E-Commerce Product Reviews on Tokopedia Using Support Vector Machine
No full textThis research aims to analyze the performance of Support Vector Machine (SVM) algorithm in classifying sentiment of e-commerce product reviews on the Tokopedia platform using web scraping data of 571 reviews from the 2024 period. The data includes review text variables, publication dates, and usernames processed through text preprocessing (text cleaning, stopword removal, stemming with Sastrawi), auto-labeling using a lexicon-based approach, and TF-IDF feature extraction with optimal parameters (max_features=5000, ngram_range=(1,2)) resulting in 1,187 features. Data splitting was performed using stratified method with proportions of training (80%) and testing (20%) on 461 reviews from binary classification filtering (positive vs negative). The research results demonstrate that Support Vector Machine with linear kernel achieved excellent performance with accuracy 95.70%, precision 95.89%, recall 95.70%, and F1-score 94.89% on the testing set. Despite the imbalanced dataset characteristics (92.4% positive vs 7.6% negative), SVM effectively handled the classification task by identifying negative sentiment with 100% precision and 42.86% recall, demonstrating its robustness in handling skewed data distribution. TF-IDF feature analysis identified the highest discriminative words such as "suitable", "goods", and "good" that are relevant for classifying consumer sentiment towards e-commerce products. The results indicate that SVM algorithm is highly effective for sentiment classification of e-commerce product reviews, making it suitable for practical implementation in automated sentiment analysis systems for online marketplaces
Identification of Buzzers in Skincare Reviews Using a Lexicon-Based Sentiment Analysis Method
No full textAlong with the rapid development of digital technology, social media has become the main platform for consumers to share experiences about products, including skincare products. However, it is not uncommon for reviews provided by users to not reflect authentic experiences, but rather reviews created by certain parties, or buzzers, to manipulate public perception. The presence of buzzers in skincare reviews is important to consider, as they can affect consumer trust and influence purchasing decisions. This study aims to identify the presence of buzzers in skincare product reviews using a lexicon dictionary-based sentiment analysis. Of the 529 comments analyzed, 75 comments showed negative sentiment and 454 comments showed positive sentiment. The classification results revealed that 85.8% of the comments belonged to the non-buzzer category, while 14.2% were indicated as buzzers. Evaluation of the classification model showed high accuracy, reaching 93%, but performance in detecting buzzers was limited, with a recall metric of only 0.50. This shows that while the model managed to classify non-buzzer comments well, there are still difficulties in identifying buzzer comments, mostly due to data imbalance. This research emphasizes the importance of a proper analytical approach in detecting inauthentic reviews to ensure the information consumers receive remains accurate, transparent, and accountable.Along with the rapid development of digital technology, social media has become the main platform for consumers to share experiences about products, including skincare products. However, it is not uncommon for reviews provided by users to not reflect authentic experiences, but rather reviews created by certain parties, or buzzers, to manipulate public perception. The presence of buzzers in skincare reviews is important to consider, as they can affect consumer trust and influence purchasing decisions. This study aims to identify the presence of buzzers in skincare product reviews using a lexicon dictionary-based sentiment analysis. Of the 529 comments analyzed, 75 comments showed negative sentiment and 454 comments showed positive sentiment. The classification results revealed that 85.8% of the comments belonged to the non-buzzer category, while 14.2% were indicated as buzzers. Evaluation of the classification model showed high accuracy, reaching 93%, but performance in detecting buzzers was limited, with a recall metric of only 0.50. This shows that while the model managed to classify non-buzzer comments well, there are still difficulties in identifying buzzer comments, mostly due to data imbalance. This research emphasizes the importance of a proper analytical approach in detecting inauthentic reviews to ensure the information consumers receive remains accurate, transparent, and accountable
Optimizing LoRa Gateway Placement for Marine Buoy Monitoring Using Particle Swarm Optimization (PSO)
No full textEffective marine environmental monitoring is critical for ensuring navigational safety, with LoRa technology emerging as a promising solution due to its long-range, low-power capabilities. However, the performance of LoRa networks heavily depends on strategic gateway placement, a task often performed manually, leading to suboptimal coverage. This study addresses this challenge by implementing and validating a Particle Swarm Optimization (PSO) algorithm to determine the optimal placement of gateways for a real-world network of 157 marine buoys in the Madura Strait. The PSO algorithm, configured with 30 particles and 100 iterations, was benchmarked against a baseline manual selection method based on geographic centrality. Results demonstrate a significant performance gain: the PSO-optimized configuration achieved 100% network coverage (157 buoys), a 34.2% increase over the 117 buoys covered by the manual method. These findings confirm that employing PSO for gateway placement substantially enhances network efficiency and data reliability, highlighting its value for creating robust and scalable marine IoT applications
Smart Waste Management Monitoring and Control Analysis Based on Objects Based on Smart Systems and Internet of Things
No full textGarbage is a problem that often becomes a trending topic in almost every country.throughout developing countries. The current condition of waste in our environment is still in a mixed condition, because the garbage has not been sorted. The minimum waste management information technology by officers also causes Waste management is slow, so that waste often piles up.The aim of this research is to develop a smart trash can that can sort metal, dry and wet waste automatically via Internet function of Things (IoT). The methodology used is Research and Development which can provide information when the trash can is full. This research was successful designing and implementing a prototype of a smart trash can based onInternet of Things (IoT) with the ability to sort waste into three categories The main components are metal, wet, and dry. The system utilizes proximity sensors inductive, soil sensor, and ultrasonic sensor HC-SR04 integrated with Blynk application for real-time monitoring of waste capacity. Algorithm Fuzzy logic is used so that the system is able to make adaptive decisions according to with the sensor condition. from the performance in the research Where the Accuracy of the system is 97.10%. The calculation is based on the number of correct predictions on the diagonal. main data divided by total data: true = 189 (Dry) + 187 (Wet) + 194 (Metal) = 570 out of a total of 587 samples, so 570/587 = 0.9710 (97.10%), with 17 error (error rate 2.90%). These values describe how much the accuracy and completeness of the model in recognizing each category of waste, with results consistently high (average 0.97)
Optimized LSTM with TSCV for Forecasting Indonesian Bank Stocks
No full textVolatility in financial markets presents complex forecasting challenges for investors, particularly within emerging economies such as Indonesia. This study proposes an optimized Long Short-Term Memory (LSTM) model for forecasting the stock prices of five significant Indonesian banks: BBCA, BBRI, BMRI, BBNI, and BBTN, utilizing daily OHLCV data (Open, High, Low, Close, Volume) and technical indicators from 2020 to 2025. The dataset comprises over 6,000 daily records, segmented using a sliding window approach to preserve temporal structure and enhance learning efficiency. Concurrently, the model architecture comprising dual LSTM layers with dropout regularization was refined through systematic hyperparameter tuning to enhance predictive performance. Model evaluation employed 5-fold Time Series Cross-Validation (TSCV), a sequential validation technique that mitigates data leakage and explicitly overcomes the limitations of conventional k-fold methods by preserving chronological integrity. Performance metrics included MSE, RMSE, MAE, R², and MAPE. The experiment results demonstrate the model’s robustness in capturing long-term dependencies within financial time series. BBCA and BMRI achieved superior accuracy (R² > 0.95), with BBCA recording the lowest MAPE of 2.34%. Despite market fluctuations, the model maintained consistent reliability across all test folds. This study overcomes a methodological limitation by integrating LSTM with TSCV in expanding markets, offering actionable insights for investors, analysts, and policymakers, and serving as a reference for adaptive AI-based, more informed forecasting tools. Moreover, the proposed framework holds promise for broader application across other financial sectors and regional markets with similar volatility characteristics