Jurnal Politeknik Negeri Batam (PoliBatam)
Not a member yet
3001 research outputs found
Sort by
Improving Retrieval-Augmented Generation Performance Using the MAF-RAG Architecture, EVR–VOR Vector Retrieval, and Multi-Agent Fallback Reasoning
No full textRetrieval-Augmented Generation (RAG) AI chatbots have gained popularity for their effectiveness in producing accurate, fast, and reliable responses; however, they have faced critical challenges stemming from limited datasets, outdated documents, and noisy, unfiltered data. This study proposes a Multi-Agent Fallback in Retrieval Augmented Generation (MAF-RAG). This robust RAG system testing pipeline integrates three-phase retrieval, filtering, and re-ranking data, along with a multi-agent debating process to address these challenges. This study demonstrates MAF-RAG\u27s ability to perform under a constrained dataset, using a near-deployment dataset of 1,100 real-world documents. The pipeline utilizes 150 testing queries, carefully selected to reflect real-world RAG-based chatbot scenarios. A sentence-transformers/all-MiniLM-L6-v encoder encodes various chunks of documents into a 384-dimensional query vector embedding, ensuring an accurate relationship between testing queries and vectorized documents. The results show that the proposed MAF-RAG significantly outperforms the baseline system, achieving a mean F1-score of 0.556, an improvement of 18.8% over the Enhanced Baseline (mean F1-score = 0.469) and a 70.0% improvement over the Legacy Baseline (mean F1-score = 0.327). MAF-RAG also achieves the highest success rate, with 78% of the queries, while other baseline systems manage only 34% and 62%, respectively. MAF-RAG also reduces the failure rate by 42.1%, significantly increasing system reliability. Although MAF-RAG exhibits an increase in latency of 4.9%, these trade-offs are outweighed by the significant improvements in system reliability and performance. These findings highlight the contribution of this study: by implementing a robust retrieval testing pipeline, system accuracy can be improved, reducing the presence of noisy and unfiltered documents, and increasing system performance even when faced with challenging and varied datasets, making it a suitable solution for a RAG-based chatbot system that faces dataset challenges
Classification of Melinjo Fruit Ripeness Using a Convolutional Neural Network (CNN) Based on Digital Images
No full textThe subjective and ineffective manual sorting of melinjo fruit, a key ingredient in Indonesian cuisine, results in inconsistent quality. This study aims to create and evaluate an automated classification system for judging the ripeness of Gnetum gnemon fruit in order to solve these issues and offer a reliable and objective quality control method. The approach was to create a customized Deep Convolutional Neural Network (Deep-CNN). The model was trained and evaluated using a simple dataset of 5,718 images that were separated into three maturity levels: raw, semi-ripe, and fully ripe. Twenty percent of the dataset was used for testing, and the remaining 80 percent was used for training. Image preparation techniques like contrast enhancement and scaling to 250x250 pixels were applied in order to optimize the model\u27s input data. The evaluation was conducted using a test dataset consisting of 1,144 photos. After eight epochs of training with the Adam optimizer, the generated Deep-CNN model demonstrated remarkable efficacy with a final classification accuracy of 99.91%. The high level of performance that remained throughout the testing phase confirmed the model\u27s strong ability to accurately identify the ripeness levels of melinjo fruit. The previously unresolved issue of automated melinjo classification is addressed in this work with a tailored and remarkably accurate (99.91%) solution. Its primary advantage is that it provides a trustworthy and unbiased technical alternative to subjective hand sorting. This directly meets industry needs by offering a scalable method to improve operational effectiveness, standardize product quality, and increase the commercial value of melinjo fruit of agricultural products
Performance Evaluation of Multi-Cloud Failover Using Domain Name System
No full textThis research implements and analyzes a multi-cloud failover system using DNS failover via AWS Route53 and Nginx reverse proxy load balancers on Google Cloud (primary) and Herza Cloud (backup), with AWS EC2 as shared backend web servers. An Ubuntu control node orchestrates deployments across these providers, enabling automatic traffic rerouting from the primary to secondary load balancer upon failure detection via health checks. Performance testing employed wrk benchmarking (4 threads, 250 connections, 300s) and Python monitoring scripts under baseline and failover scenarios with DNS TTLs of 30s, 60s, and 120s. Baseline yielded 2,291.81 req/s throughput, 108.42ms average latency, and 231.15ms p99 latency. Failover results showed TTL 30s optimal for reliability (152.65s downtime, 48.62% failed requests, 30.53s average recovery), outperforming TTL 60s (243.92s downtime, 83.48% failures due to health check mismatch) and TTL 120s (186.88s downtime) and TTL 30s is recommended for high availability in low-budget SMEs, balancing reduced downtime against DNS overhead. However, this approach is limited to small-scale infrastructure
Analysis of Public Sentiment Towards the Free Nutritious Meals Program on TikTok Social Media Using the K-Nearest Neighbor Algorithm
No full textThe Free Nutritious Meals Program is currently one of the most talked about public policies, generating a wide range of responses from the public. One of the most active discussion forums is the social media platform TikTok, given that it has a large number of users and a relaxed and informal style of language. This study aims to examine public sentiment toward the MBG program through TikTok user comments, while also testing the performance of the K-Nearest Neighbor (KNN) algorithm in classifying sentiment as positive or negative. Research data was collected by crawling comments on several TikTok videos discussing Free Nutritious Meals during the period from September to November 2025. A total of 1,000 comments were obtained and then processed through data cleaning stages, such as data cleaning, case folding, normalization, tokenization, stopword removal, and stemming. To convert the text into numerical form, the Term Frequency–Inverse Document Frequency (TF-IDF) method was used. Meanwhile, sentiment labeling was done manually to maintain the quality of the training data. Model performance was evaluated using a confusion matrix with accuracy, precision, recall, and F1-score indicators. The test results showed that the best accuracy rate, which was 70.50%, was obtained at a K value of 4. From the sentiment analysis conducted, negative comments were found to outnumber positive sentiments. The criticism that emerged generally related to food quality and safety, inequality in program distribution, and a lack of transparency in information provided to the public. This study shows that the KNN algorithm is quite capable of being used for sentiment analysis on TikTok comment data, although it still has limitations in understanding the variety of informal language often used by users. Therefore, the results of this study are expected to provide public opinion-based input for policymakers, as well as a foundation for the development of sentiment analysis methods that are more suited to the characteristics of social media in future studies
Dynamics and Control of Human Papillomavirus (HPV) Infection Using an SVITR Compartmental Model
No full textHuman papillomavirus (HPV) remains a significant public health concern due to its high transmissibility and associated health risks. This study underscores the pivotal role of vaccination in reducing HPV transmission, while also highlighting the limitations of relying solely on vaccination for infection control. In this study, we present a deterministic compartmental model to investigate the transmission dynamics of Human Papillomavirus (HPV). The model stratifies the population into five compartments: susceptible individuals S(t), Vaccinated individuals V(t), HPV Infected individuals I(t), treated HPV-infected individuals T(t) and recovered individuals R(t). We establish the existence and uniqueness of the model solution and also examine the existence of disease-free and endemic equilibrium and analyze their stability properties. Numerical simulations were performed to explore the temporal evolution of the compartments, assess the sensitivity of key parameters, and investigated the behaviour of the basic reproduction number R_0. Our findings were that a comprehensive strategy, incorporating both preventive vaccination and therapeutic management, is essential for achieving sustainable control of HPV spread. Strengthening these measures, alongside reducing transmission through demographic interventions, offers the best way for long-term management of the infection. These results provide insights into the impact of vaccination and treatment strategies on HPV transmission and highlight critical factors for public health
Public Sentiment Analysis on Demonstration Actions Using IndoBERT Based on Transfer Learning
No full textSentiment analysis based on language modeling plays a crucial role in mapping public perception of socio-political dynamics in Indonesia. This study aims to evaluate public sentiment toward the House of Representatives of the Republic of Indonesia (DPR RI) in response to the August 2025 demonstrations using the IndoBERT model based on transfer learning. The dataset comprises 1,815 Indonesian-language opinion texts classified into positive and negative sentiments. Due to a substantial class imbalance dominated by negative opinions, a hybrid sampling strategy combining oversampling and undersampling was employed to obtain a balanced dataset of 650 samples per class. The research methodology included text preprocessing, an 80:20 training–testing split, and fine-tuning the IndoBERT-base-p1 model. Experimental results indicate that the proposed model achieves robust and balanced performance, with an overall accuracy of 85%. Precision and F1-score for both sentiment classes reached 0.85, while recall values were 0.86 for negative sentiment and 0.85 for positive sentiment, demonstrating the model’s ability to identify both classes effectively without bias toward the majority class. Despite the dominance of negative sentiment in the original dataset, the application of data balancing techniques successfully mitigated class imbalance effects, enabling fair and proportional sentiment classification. These findings confirm that the IndoBERT-based transfer learning approach is effective in capturing public sentiment related to mass demonstrations and can provide valuable, data-driven insights for policymakers in understanding societal concerns in the digital era.Sentiment analysis based on language modeling plays a crucial role in mapping public perception of socio-political dynamics in Indonesia. This study aims to evaluate public sentiment toward the House of Representatives of the Republic of Indonesia (DPR RI) in response to the August 2025 demonstrations using the IndoBERT model based on transfer learning. The dataset comprises 1,815 Indonesian-language opinion texts classified into positive and negative sentiments. Due to a substantial class imbalance dominated by negative opinions, a hybrid sampling strategy combining oversampling and undersampling was employed to obtain a balanced dataset of 650 samples per class. The research methodology included text preprocessing, an 80:20 training–testing split, and fine-tuning the IndoBERT-base-p1 model. Experimental results indicate that the proposed model achieves robust and balanced performance, with an overall accuracy of 85%. Precision and F1-score for both sentiment classes reached 0.85, while recall values were 0.86 for negative sentiment and 0.85 for positive sentiment, demonstrating the model’s ability to identify both classes effectively without bias toward the majority class. Despite the dominance of negative sentiment in the original dataset, the application of data balancing techniques successfully mitigated class imbalance effects, enabling fair and proportional sentiment classification. These findings confirm that the IndoBERT-based transfer learning approach is effective in capturing public sentiment related to mass demonstrations and can provide valuable, data-driven insights for policymakers in understanding societal concerns in the digital era
Performance Analysis of KNN and BERT Algorithms for Classifying Student Sentiments Towards Campus Services
No full textThis study addresses the limitations of campus service evaluation processes that are still conducted manually and are unable to optimally process students’ textual opinions. The objective of this research is to analyze and compare the performance of the K-Nearest Neighbor (KNN) and BERT algorithms in classifying student sentiments toward campus services. The research stages include text preprocessing, the generation of IndoBERT embeddings for the KNN model, and fine-tuning IndoBERT for direct sentiment classification. The dataset consists of student evaluation texts from the Faculty of Engineering at UNSULBAR, labeled as negative, neutral, and positive sentiments. Model evaluation is performed using accuracy, precision, recall, and F1-score metrics. The results show that the basic KNN model achieves an accuracy of 79%, while KNN with hyperparameter tuning improves performance to 86%. The BERT model delivers the best performance, achieving an accuracy of 88.68%, precision of 87.87%, recall of 90.19%, and an F1-score of 88.79%. These findings indicate that transformer-based approaches, particularly IndoBERT, are more effective in understanding the contextual nuances of student language than traditional methods, and are therefore more recommended for sentiment analysis implementation in campus service evaluation.This study addresses the limitations of campus service evaluation processes that are still conducted manually and are unable to optimally process students’ textual opinions. The objective of this research is to analyze and compare the performance of the K-Nearest Neighbor (KNN) and BERT algorithms in classifying student sentiments toward campus services. The research stages include text preprocessing, the generation of IndoBERT embeddings for the KNN model, and fine-tuning IndoBERT for direct sentiment classification. The dataset consists of student evaluation texts from the Faculty of Engineering at UNSULBAR, labeled as negative, neutral, and positive sentiments. Model evaluation is performed using accuracy, precision, recall, and F1-score metrics. The results show that the basic KNN model achieves an accuracy of 79%, while KNN with hyperparameter tuning improves performance to 86%. The BERT model delivers the best performance, achieving an accuracy of 88.68%, precision of 87.87%, recall of 90.19%, and an F1-score of 88.79%. These findings indicate that transformer-based approaches, particularly IndoBERT, are more effective in understanding the contextual nuances of student language than traditional methods, and are therefore more recommended for sentiment analysis implementation in campus service evaluation
Assessing Academic Website Quality Using the WebQual 4.0 Framework
No full textThe rapid growth of digital technologies in higher education has encouraged universities to improve academic service delivery through integrated online platforms. This study aims to measure the quality of the E-PPT (Pusat Pelayanan Terpadu Berbasis Elektronik) website of the Faculty of Computer Science, Universitas Sriwijaya, using the WebQual 4.0 model. A quantitative descriptive method was adopted, involving 354 valid respondents drawn from a total faculty population of 3,058 students across four academic levels (Diploma, Undergraduate, Master’s, and Doctoral). Data were collected via an online survey and analyzed using Microsoft Excel and SPSS. The website achieved an overall mean score of 3.60, indicating a good level of quality. Information Quality showed the highest performance (3.68), followed by Usability (3.58) and Service Interaction Quality (3.56). A supplementary correlation analysis also confirmed positive associations among the WebQual dimensions and overall website quality. These results suggest that the website delivers accurate information and is easy to use, although improvements are needed in responsiveness and interaction quality
Optimization of Support Vector Machine Model Performance in Image Classification through Dimension Reduction with Principal Component Analysis (PCA)
No full textThis study examines how to optimize a Support Vector Machine (SVM) model using a dimensionality reduction method called Principal Component Analysis (PCA) to classify images with multiple dimensions. The dataset used is Chessman images with an initial number of features of 12,288. PCA was applied with the aim of retaining 99% of the total variation, resulting in 312 principal components. The results show a significant improvement in computational efficiency: training time was drastically reduced from 29.85 seconds to just 0.17 seconds (168 times faster), and memory usage decreased from 25.83 MB to 0.66 MB (97% more efficient). Although the accuracy experienced a small decrease, namely from 31.58% to 31.22%, PCA still functions as a noise filter that helps improve performance, especially in classes with complex visual patterns, such as an increase in the F1-score of the "Rook" class from 0.32 to 0.37. The conclusions of this study indicate that PCA provides important efficiency improvements without significantly sacrificing classification performance.This study examines how to optimize a Support Vector Machine (SVM) model using a dimensionality reduction method called Principal Component Analysis (PCA) to classify images with multiple dimensions. The dataset used is Chessman images with an initial number of features of 12,288. PCA was applied with the aim of retaining 99% of the total variation, resulting in 312 principal components. The results show a significant improvement in computational efficiency: training time was drastically reduced from 29.85 seconds to just 0.17 seconds (168 times faster), and memory usage decreased from 25.83 MB to 0.66 MB (97% more efficient). Although the accuracy experienced a small decrease, namely from 31.58% to 31.22%, PCA still functions as a noise filter that helps improve performance, especially in classes with complex visual patterns, such as an increase in the F1-score of the "Rook" class from 0.32 to 0.37. The conclusions of this study indicate that PCA provides important efficiency improvements without significantly sacrificing classification performance
A Systematic Review of Post-Quantum Cryptography for Healthcare Data Protection: Performance, Readiness, and Deployment Challenges
No full textThe traditional cryptographic methods used to protect healthcare data, especially for the long-term storage of medical imaging records, are becoming increasingly threatened by the quick development of quantum computing. The purpose of this study is to assess the challenges, efficacy, and preparedness of integrating Post-Quantum Cryptography (PQC) into healthcare information systems. Twenty peer-reviewed studies published between 2020 and 2025 were analysed following the Preferred Reporting Items for Systematic Reviews and Meta Analyses (PRISMA) protocol. The review was conducted using a systematic research design that included qualitative thematic synthesis, predetermined eligibility criteria, and database searching. According to the results, lattice-based PQC schemes, specifically, CRYSTALS-Kyber for encryption and CRYSTALS-Dilithium for authentication, show great promise because of their effectiveness, resilience, and suitability for decentralized architectures like blockchain and Internet-of-Medical-Things environments. Nonetheless, the review points out a notable deficiency of empirical assessment in actual healthcare settings, particularly with regard to cloud-based platforms and Picture Archiving and Communication Systems utilized in medical imaging processes. Scalability limitations, intricate key-management specifications, system interoperability restrictions, and the requirement for conformity with regulatory and compliance frameworks are some of the major issues noted. The results indicate that lattice-based PQC schemes have great promise, deployment readiness remains largely at the conceptual and experimental stage, particularly for cloud-based PACS environments. Real-world implementation validation in a healthcare setting has not been achieved