Jurnal Politeknik Negeri Batam (PoliBatam)
Not a member yet
3001 research outputs found
Sort by
Classification of Facial Acne Types Based on Self-Supervised Learning using DINOv2
No full textAcne is a common inflammatory skin condition that can affect an individual’s psychological well-being and overall quality of life. The inability to independently recognize specific types of acne often leads to the use of inappropriate skincare products. This situation highlights the need for an image-based classification system that can provide accurate visual identification. The self-supervised learning method Distillation with NO Labels, version 2 (DINOv2), is employed as a feature extractor to classify four types of acne—Acne fulminans, Acne nodules, Papules, and Pustules—using the “skin-90” dataset. The fine-tuning process is conducted through a Parameter-Efficient Fine-Tuning (PEFT) approach using Low-Rank Adaptation (LoRA) to adjust the model’s visual representations to the acne domain without updating all parameters in full, followed by integration with a classification head. The results show that the model achieves an accuracy of 90.70%, with precision, recall, and F1-score values of 90.64%, 90.68%, and 90.57%, respectively. The findings suggest that the proposed architectural design and training configuration are suitable for capturing relevant visual patterns of acne, while further validation is required to assess robustness across more diverse data distributions
Evaluating Image Recognition Accuracy in Explicit Content Detection: A Comparative Study with Indonesian Perceptions
No full textThis study evaluates image recognition accuracy in explicit content detection by using the Indonesian social context as a comparative reference. Google Vision SafeSearch is employed as a representative automated image recognition system widely used in online content moderation. Although such systems provide efficiency in detecting adult, violent, or racy content, challenges arise when their detection outputs must align with more conservative cultural and religious norms, such as those in Indonesia. A quantitative descriptive-comparative method was applied by testing six representative images based on SafeSearch explicit content categories (adult, racy, violence, medical, and spoof) and comparing the automated detections with Indonesian respondents’ perceptions collected through a Likert-scale questionnaire. Statistical analysis shows a significant difference between the system’s explicit content classifications and human perceptions, with respondents consistently rating explicitness higher than Google Vision API. Despite this difference, a strong Spearman rank correlation indicates that Google Vision SafeSearch is consistent in ranking explicit content levels, although still limited in capturing emotional intensity and cultural sensitivity. These findings highlight how Indonesian social and cultural norms shape the perception of explicit imagery, emphasizing the need for image recognition systems that incorporate local contextual factors
Comparison of Random Forest and LSTM for Tokopedia Sentiment Analysis
No full textTokopedia is one of the largest e-commerce platforms in Indonesia, where every transaction generates user reviews containing opinions about the products or services received. These reviews provide important information about product quality, but the very large quantity makes manual analysis inefficient. This study aims to automatically classify Tokopedia review sentiment and compare the performance of machine learning and deep learning methods. The dataset used was obtained from Kaggle and has undergone an initial cleaning stage, including removing irrelevant columns and manually labeling into two sentiment classes, positive and negative. The research methodology includes several stages, namely data preprocessing (cleaning, case-folding, stopword removal, tokenization, normalization, and stemming), feature extraction using TF-IDF for Random Forest and word embedding for LSTM, implementation of Random Forest and Long Short-Term Memory (LSTM) models, and model evaluation using confusion matrix. Experimental results show that LSTM provides the best performance with 94% accuracy, while Random Forest achieves 92% accuracy. These findings indicate that LSTM is more effective in understanding language context, resulting in more accurate sentiment classification and is useful for decision making in the e-commerce field.Tokopedia is one of the largest e-commerce platforms in Indonesia, where every transaction generates user reviews containing opinions about the products or services received. These reviews provide important information about product quality, but the very large quantity makes manual analysis inefficient. This study aims to automatically classify Tokopedia review sentiment and compare the performance of machine learning and deep learning methods. The dataset used was obtained from Kaggle and has undergone an initial cleaning stage, including removing irrelevant columns and manually labeling into two sentiment classes, positive and negative. The research methodology includes several stages, namely data preprocessing (cleaning, case-folding, stopword removal, tokenization, normalization, and stemming), feature extraction using TF-IDF for Random Forest and word embedding for LSTM, implementation of Random Forest and Long Short-Term Memory (LSTM) models, and model evaluation using confusion matrix. Experimental results show that LSTM provides the best performance with 94% accuracy, while Random Forest achieves 92% accuracy. These findings indicate that LSTM is more effective in understanding language context, resulting in more accurate sentiment classification and is useful for decision making in the e-commerce field
The Integration of AHP and Rank Order Centroid in a Decision Support System for Selecting Social Media for MSMEs
No full textThis study aims to develop a Decision Support System (DSS) to assist Micro, Small, and Medium Enterprises (MSMEs) in selecting the most optimal social media platform for promotional activities. In the digital era, choosing an appropriate platform is a critical factor in enhancing marketing effectiveness; however, many MSMEs face challenges in making informed decisions due to limited analytic capabilities and resources. To address this issue, the proposed system integrates the Analytic Hierarchy Process (AHP) to determine the relative importance of decision criteria and the Rank Order Centroid (ROC) method to assign weights to the alternatives. The evaluation criteria include audience reach, cost efficiency, and user engagement, which are considered essential factors in digital marketing strategies for MSMEs. The results indicate that Instagram achieved the highest score of 0.208 and is recommended as the most suitable social media platform for MSME promotion. TikTok ranked second with a score of 0.082, followed by Facebook with a score of 0.041. Furthermore, user validation testing demonstrates that the system is well accepted by MSME practitioners, as it provides recommendations that are accurate, structured, and easy to use.This research contributes by offering a technology-based decision-making solution that enhances the effectiveness of digital marketing strategies for MSMEs. The developed DSS serves as a practical and relevant tool to support promotional decision-making and to address the challenges of social media utilization in today’s competitive digital landscape.This study aims to develop a Decision Support System (DSS) to assist Micro, Small, and Medium Enterprises (MSMEs) in selecting the most optimal social media platform for promotional activities. In the digital era, choosing an appropriate platform is a critical factor in enhancing marketing effectiveness; however, many MSMEs face challenges in making informed decisions due to limited analytic capabilities and resources. To address this issue, the proposed system integrates the Analytic Hierarchy Process (AHP) to determine the relative importance of decision criteria and the Rank Order Centroid (ROC) method to assign weights to the alternatives. The evaluation criteria include audience reach, cost efficiency, and user engagement, which are considered essential factors in digital marketing strategies for MSMEs. The results indicate that Instagram achieved the highest score of 0.208 and is recommended as the most suitable social media platform for MSME promotion. TikTok ranked second with a score of 0.082, followed by Facebook with a score of 0.041. Furthermore, user validation testing demonstrates that the system is well accepted by MSME practitioners, as it provides recommendations that are accurate, structured, and easy to use.This research contributes by offering a technology-based decision-making solution that enhances the effectiveness of digital marketing strategies for MSMEs. The developed DSS serves as a practical and relevant tool to support promotional decision-making and to address the challenges of social media utilization in today’s competitive digital landscape
Comparison of LightGBM and CatBoost Algorithms for Diabetes Prediction Based on Clinical Data
No full textDiabetes Mellitus presents a global health challenge necessitating accurate early detection to prevent fatal complications. However, clinical data often exhibit imbalanced class distributions, hindering standard prediction models from effectively detecting positive patients. This study aims to compare the performance of two modern Gradient Boosting algorithms, LightGBM and CatBoost, in predicting diabetes risk. Random Forest and Logistic Regression algorithms were included as baseline models to benchmark effectiveness. To address data imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was applied during the training data preprocessing stage. The dataset was sourced from the Kaggle public repository (Diabetes Prediction Dataset), comprising 100,000 patient medical records with clinical attributes such as age, body mass index (BMI), and HbA1c levels. Performance evaluation utilized Accuracy, Precision, Recall, F1-Score, and Area Under the Curve (AUC) metrics. Experimental results demonstrated a tight competition, where LightGBM achieved the highest Accuracy of 97.16%. However, CatBoost demonstrated superior sensitivity (Recall) of 69.71% and the highest F1-Score of 80.48%. This makes CatBoost the most reliable model in minimizing False Negatives compared to LightGBM and Random Forest, whereas Logistic Regression showed the lowest performance. Furthermore, interpretability analysis using SHAP (SHapley Additive exPlanations) revealed that HbA1c and blood glucose levels were the most dominant features in detection, validating the model\u27s alignment with clinical diagnosis. This study concludes that the CatBoost algorithm combined with SMOTE offers a more sensitive, transparent, and efficient diabetes prediction for medical screening.Diabetes Mellitus presents a global health challenge necessitating accurate early detection to prevent fatal complications. However, clinical data often exhibit imbalanced class distributions, hindering standard prediction models from effectively detecting positive patients. This study aims to compare the performance of two modern Gradient Boosting algorithms, LightGBM and CatBoost, in predicting diabetes risk. Random Forest and Logistic Regression algorithms were included as baseline models to benchmark effectiveness. To address data imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was applied during the training data preprocessing stage. The dataset was sourced from the Kaggle public repository (Diabetes Prediction Dataset), comprising 100,000 patient medical records with clinical attributes such as age, body mass index (BMI), and HbA1c levels. Performance evaluation utilized Accuracy, Precision, Recall, F1-Score, and Area Under the Curve (AUC) metrics. Experimental results demonstrated a tight competition, where LightGBM achieved the highest Accuracy of 97.16%. However, CatBoost demonstrated superior sensitivity (Recall) of 69.71% and the highest F1-Score of 80.48%. This makes CatBoost the most reliable model in minimizing False Negatives compared to LightGBM and Random Forest, whereas Logistic Regression showed the lowest performance. Furthermore, interpretability analysis using SHAP (SHapley Additive exPlanations) revealed that HbA1c and blood glucose levels were the most dominant features in detection, validating the model\u27s alignment with clinical diagnosis. This study concludes that the CatBoost algorithm combined with SMOTE offers a more sensitive, transparent, and efficient diabetes prediction for medical screening
Forecasting Export Values in West Sumatra Using Backpropagation Neural Network
No full textExport value is an important indicator in supporting regional economic growth. However, its movement tends to be volatile and non-linear, making it difficult to forecast using conventional statistical methods such as ARIMA. This study aims to forecast the export value of West Sumatra Province using an Artificial Neural Network (ANN) with the Backpropagation algorithm. The data used consist of monthly export values from January 2006 to October 2025 obtained from Badan Pusat Statistik (BPS) of West Sumatra Province. The data were normalized and modified using the rolling window method, then divided into training and testing datasets. Several network architectures were evaluated through a trial-and-error process with variations in the number of neurons in the hidden layer. The best model was achieved with the BPNN(12,12,1) architecture, yielding a Mean Square Error (MSE) of 0.0236 and a Mean Absolute Percentage Error (MAPE) of 25.31%. The results indicate that the model is capable of capturing non-linear patterns and reasonably following the trend of the actual data. The selected model was then used to perform short-term forecasting of export values for the period from November 2025 to March 2026. The findings demonstrate that the Backpropagation Neural Network algorithm is effective for forecasting export values in West Sumatra Province. This study contributes theoretically by enriching the application of artificial intelligence in regional economic forecasting and practically by supporting data-driven policy formulation for export strategies in West Sumatra.Export value is an important indicator in supporting regional economic growth. However, its movement tends to be volatile and non-linear, making it difficult to forecast using conventional statistical methods such as ARIMA. This study aims to forecast the export value of West Sumatra Province using an Artificial Neural Network (ANN) with the Backpropagation algorithm. The data used consist of monthly export values from January 2006 to October 2025 obtained from Badan Pusat Statistik (BPS) of West Sumatra Province. The data were normalized and modified using the rolling window method, then divided into training and testing datasets. Several network architectures were evaluated through a trial-and-error process with variations in the number of neurons in the hidden layer. The best model was achieved with the BPNN(12,12,1) architecture, yielding a Mean Square Error (MSE) of 0.0236 and a Mean Absolute Percentage Error (MAPE) of 25.31%. The results indicate that the model is capable of capturing non-linear patterns and reasonably following the trend of the actual data. The selected model was then used to perform short-term forecasting of export values for the period from November 2025 to March 2026. The findings demonstrate that the Backpropagation Neural Network algorithm is effective for forecasting export values in West Sumatra Province. This study contributes theoretically by enriching the application of artificial intelligence in regional economic forecasting and practically by supporting data-driven policy formulation for export strategies in West Sumatra
Analysis of Digital Readiness in the Social Assistance Distribution System with the Unified Theory of Acceptance and Use of Technology (UTAUT)
Full text linkThe adoption of digital systems for social assistance distribution has become increasingly vital in enhancing efficiency and accessibility. This study examines the acceptance of such a system using the Unified Theory of Acceptance and Use of Technology (UTAUT) model, analyzing six key constructs: Performance Expectancy (PE), Effort Expectancy (EX), Social Influence (SI), Facilitating Conditions (FC), Behavioral Intention (BI), and Actual Use (AU). A total of 150 respondents participated in the survey, providing insights into their perceptions of the system. The findings indicate that Performance Expectancy (4.2) received the highest mean score, demonstrating that users perceive the system as beneficial in improving efficiency. Effort Expectancy (4.0) suggests that the system is easy to use, while Social Influence (3.8) highlights the moderate role of external encouragement. Facilitating Conditions (3.9) reveal the availability of infrastructure but also suggest areas for improvement. Additionally, Behavioral Intention (4.1) and Actual Use (4.0) indicate strong user commitment toward system utilization. The study contributes to the understanding of digital technology adoption in social welfare programs and provides recommendations for optimizing system implementation. Future research should explore the long-term impact of digital adoption, assess its effectiveness in different demographic groups, and integrate qualitative insights to deepen the understanding of user experiences. Additionally, expanding the analysis to include external factors such as policy support, economic conditions, and digital literacy could further enhance the model’s applicability
Named Entity Recognition for Medical Records of Heart Failure Using a Pre-trained BERT Model
Full text linkThis study aims to develop a Named Entity Recognition (NER) model based on a pre-trained BERT model for medical records of heart failure patients. The focus of this research is to classify essential medical entities from unstructured medical record texts. The classification covers four categories: objective data (patient identity, laboratory test results, and objective examination data), subjective data (patient complaints), prescriptions, and diagnoses (diagnosis codes and descriptions). The methodology employs Natural Language Processing (NLP) techniques using Transformer-based architectures, such as Bidirectional Encoder Representation from Transformers (BERT). The developed model is evaluated based on entity label prediction accuracy and medical entity classification performance. The results indicate that the BERT-based NER model performs well, achieving an entity prediction accuracy of 84.82%. Furthermore, the model effectively classifies medical entities from input texts in alignment with expected medical entities. This research is expected to contribute significantly to medical data management, assist healthcare professionals in clinical decision-making, and serve as a reference for the development of AI-based healthcare technology in Indonesia.This study aims to develop a Named Entity Recognition (NER) model based on a pre-trained BERT model for medical records of heart failure patients. The focus of this research is to classify essential medical entities from unstructured medical record texts. The classification covers four categories: objective data (patient identity, laboratory test results, and objective examination data), subjective data (patient complaints), prescriptions, and diagnoses (diagnosis codes and descriptions). The methodology employs Natural Language Processing (NLP) techniques using Transformer-based architectures, such as Bidirectional Encoder Representation from Transformers (BERT). The developed model is evaluated based on entity label prediction accuracy and medical entity classification performance. The results indicate that the BERT-based NER model performs well, achieving an entity prediction accuracy of 84.82%. Furthermore, the model effectively classifies medical entities from input texts in alignment with expected medical entities. This research is expected to contribute significantly to medical data management, assist healthcare professionals in clinical decision-making, and serve as a reference for the development of AI-based healthcare technology in Indonesia
Influence Literacy Finance Shariah and Use Syariah Fintech Lending on the Development of MSMEs
Full text linkMicro, Small, and Medium Enterprises (MSMEs) play an important role in supporting the national economy, including in Medan City. This study aims to analyze the influence of Islamic financial literacy and Islamic fintech lending on the development of MSMEs in Medan City. Islamic financial literacy includes MSME actors\u27 understanding of Islamic financial principles, such as the prohibition of usury and gharar, while Islamic fintech lending is a technology-based financing alternative that is in accordance with Islamic principles. This study uses a quantitative approach with multiple linear regression analysis methods. Data were obtained through questionnaires distributed to 86 MSME actors selected using purposive sampling techniques. The results of the study indicate that Islamic financial literacy significantly affects the ability perpetrator MSMEs in manage finance And determine source appropriate financing. In addition, sharia fintech lending has a significant positive impact on the ability of MSMEs to obtain fast, safe and halal financing. Simultaneously, both variables contribute positively to the development of MSMEs. This finding confirms that collaboration between Islamic financial literacy and financial technology can be an effective solution to overcome the challenges faced by MSMEs
Implementation of LSTM Method on Tidal Prediction in Semarang Region
Full text linkSemarang is the capital of the Central Java province, located in the north and directly adjacent to the Java Sea. Having an almost flat land condition with a slope of about 0-2%, Semarang City has the opportunity to experience tidal flooding. The occurrence of tides does not have a fixed period. So, it is necessary to predict the height of the tide and the ebb of the seawater. Thus, this research aims to predict tides in the Semarang area using the LSTM method. The data used is tidal data in Semarang waters from 2020 to 2024. The advantage of the LSTM method is its ability to effectively remember time series data or data with long-term dependence. LSTM can store past information using special cells contained in its structure. This research on tidal prediction using the LSTM method with 70% training data trial batch size 32 and epoch 200 obtained the smallest error value, namely the MAE value of 0.0388 and MAPE of 0.0313 which is the best LSTM result