Jurnal CoSciTech (Computer Science and Information Technology)

Implementasi Asset Management Berbasis Android PT. PLN ICON PLUS

Yogi Pratama, Dewi Setiowati — Sun, 19 Apr 2026 00:00:00 +0700

Digital transformation urges companies to optimize asset management to improve operational efficiency. PT. PLN ICON PLUS, a subsidiary of PT PLN (Persero) specializing in information technology and network infrastructure, still uses manual asset management systems prone to errors and inefficiencies. This study aims to develop an Android-based asset management system integrated with a Laravel backend to enable real-time tracking and asset data updates. The methods used include Fishbone analysis, system design with UML, Android and Laravel-based application development, and black box testing. The implementation results show that the system effectively handles asset loan transactions, category and asset management, and user interactions through an interactive and user-friendly interface. This system improves data accuracy, reduces administrative burden, and facilitates data access for administrators and clients. In conclusion, the developed application successfully replaces manual systems with a more efficient and integrated digital solution.

Clickbait Text Classification with Deep Learning Hybrid LSTM-CNN Method

Aditya Darma, Sepni Sepni, Shelly Adelia Sary, Yayang Uswatun Hasanah — Sun, 19 Apr 2026 00:00:00 +0700

This study aims to determine the category of news titles by dividing them into two groups, namely clickbait and non-clickbait using the LSTM-CNN hybrid method. The dataset used consists of 14,878 data in two categories with 6,285 clickbait news data and 8,693 non clickbait news data obtained from the kaggle page. The research stages include data preprocessing through cleaning, tokenizing, stopword removal, stemming, and text representation using the Word2Vec algorithm. The dataset will then be separated into training and test data using a ratio of 80:20. The LSTM-CNN hybrid model is used because of CNN's advantage in extracting local features as well as LSTM's ability to understand sequential relationships between words. The model performance evaluation was conducted using confusion matrix, with the results of 77.07% accuracy, 70% recall, 73% precision, and 71% F1-score. The LSTM-CNN hybrid model showed better performance than the separate models with an increase in accuracy from 77% to 77.07%. This research shows that the LSTM-CNN model combination is effective in handling clickbait and non-clickbait news text classification, providing quite good results in improving the performance of the previous model.

Optimization of Backpropagation Algorithms for Enhancing Market Prediction Accuracy in Emerging Industries

Syaharuddin Syaharuddin — Sun, 19 Apr 2026 00:00:00 +0700

This research aims to systematically review and analyze the application of backpropagation algorithms in data-driven business market prediction, focusing on emerging industries. Using the Systematic Literature Review (SLR) method, this study examined research articles from the Dimensions and Scopus databases published over the last 10 years. The analysis synthesizes findings related to the effectiveness, challenges, and potential advancements of Backpropagation in improving market prediction accuracy. The results reveal that backpropagation models, particularly LSTM and MLP have shown significant promise in various sectors, including financial forecasting, customer behavior analysis, and sales prediction. However, challenges such as overfitting, high computational costs, and the integration of real-world market complexities remain. The study emphasizes the need for continuous optimization of these models, as well as improvements in data quality and computational efficiency. This research contributes valuable insights for enhancing predictive models in business market forecasting and offers directions for future studies to further refine the use of backpropagation in addressing market prediction challenges.

Impact of Retinal Image Preprocessing on Diabetic Retinopathy Classification Using Prototypical Networks

Abi Eka Putra Wulyono, Faisal Muttaqin, Budi Mukhamad Mulyo — Sun, 19 Apr 2026 00:00:00 +0700

Diabetic retinopathy is a diabetes complication that can lead to progressive retinal damage and permanent blindness. Early detection through automated fundus image classification is essential but challenged by varying image quality, background noise, and color dominance that reduces lesion visibility. Prototypical networks have demonstrated good performance in few-shot learning settings, yet specialized preprocessing is rarely explored. This study proposes a prototypical network enhanced with modified circle crop to remove irrelevant regions and enhanced green channel to improve microvascular lesion contrast. Experiments were conducted on the APTOS 2019 dataset consisting of 3,662 images, split into 2,929 training and 733 testing samples, using a 5-way 5-shot configuration. The proposed preprocessing increases accuracy from 64.53 percent to 71.35 percent and improves quadratic weighted kappa from 0.5712 to 0.6990. These results indicate that preprocessing enhances feature representation and classification performance under limited data conditions.

Convolutional Neural Network with InceptionV3 Architecture for Food Image Classification Based on Origin of Java and Sumatra Regions

Diva Nayla Khasanah, Rahmad Firdaus — Sun, 27 Apr 2025 00:00:00 +0700

This study aims to improve the accuracy of classifying traditional food images based on the regions of Java and Sumatra using the Convolutional Neural Network (CNN) algorithm with the InceptionV3 architecture. Traditional foods from these two regions are often difficult to distinguish due to visual similarities. The dataset consists of 472 food images processed through segmentation, augmentation, and rescaling. The InceptionV3 model was selected for its ability to capture complex visual patterns with high efficiency. The training process employed the Adam optimizer, a learning rate of 0.001, and a 50% dropout regularization technique to prevent overfitting. The model was evaluated using accuracy, precision, recall, and F1-score metrics. The results show that the model achieved an accuracy of 90.42%.precision of 91.07%, recall of 92.72%, and F1-score of 90%, significantly improving compared to previous research, which only achieved an accuracy of 64% using CNN without a specific architecture. This study is expected to contribute to the preservation of local culinary culture and support the promotion of tourism and technology-based culinary industries in Indonesia.

Comparison of YOLOv8n and YOLOv8s Model Performance in Object Detection Using Evaluation Metrics and Confidence Scores

Doms Upuy — Sun, 27 Apr 2025 00:00:00 +0700

Deep learning-based object detection has developed rapidly and is widely used in various computer vision applications. One of the most widely used methods is the You Only Look Once (YOLO) algorithm, which is capable of real-time object detection with a high degree of accuracy. This study aims to analyze the performance comparison of two variants of the YOLOv8 model, namely YOLOv8n and YOLOv8s, in detecting objects using evaluation metrics and confidence scores. The dataset used consists of 5000 images, which are divided into training data (70%), validation data (20%), and testing data (10%). Model performance evaluation is carried out using several object detection metrics, namely precision, recall, and mean Average Precision (mAP), as well as additional analysis in the form of computation time and confidence scores to assess the stability of the model's predictions. The results show that the YOLOv8n model achieved a precision value of 0.9313, while the YOLOv8s model achieved a recall value of 0.8415 and a mean Average Precision (mAP0.5) of 0.9055, which is slightly higher than YOLOv8n with a mAP0.5 value of 0.9009. In terms of computational efficiency, YOLOv8n has a faster training time of around 2670 seconds, compared to YOLOv8s which takes around 4477 seconds. In addition, the YOLOv8s model shows a higher confidence score, which indicates a better level of prediction confidence in detecting objects. Based on these results, it can be concluded that YOLOv8n is superior in computational efficiency, while YOLOv8s provides more stable and accurate detection performance. The results of this study are expected to serve as a reference in selecting the optimal object detection model for various computer vision-based applications

Klasifikasi Kendaraan Bermotor Berdasarkan Jumlah Gandar Menggunakan Adaptive Minimal Ensemble

Abdurrahman Al Hakim, Faisal Muttaqin, Hendra Maulana — Sun, 27 Apr 2025 00:00:00 +0700

The increasing volume of motor vehicles requires automated monitoring for the classification of heavy vehicle categories (Category I–V) based on the number of axles using side-view cameras. This process represents a complex fine-grained visual classification challenge due to the similar body shapes of trucks. To address the dilemma between the need for high accuracy and computational efficiency, this study implements an Adaptive Minimal Ensemble (AME) architecture that adaptively combines small-scale models. The model is evaluated using a confusion matrix along with accuracy, precision, recall, and F1-score metrics. The testing results demonstrate that a single EfficientNetV2-S model is only able to achieve a maximum accuracy of 83% and exhibits significant limitations in extracting crucial distinguishing features, leading to misclassification of Category 4 and 5 vehicles. In contrast, the AME architecture, which utilizes the two best-performing EfficientNetV2-S base models, successfully achieves a substantial performance improvement with 95% accuracy, 95.21% precision, 95% recall, and a 94.99% F1-score. In conclusion, the adaptive layer mechanism in AME is proven to be highly effective in compensating for the individual prediction weaknesses of its base models, resulting in a significantly more precise vehicle classification monitoring system.

Implementasi LoRa pada Monitoring Tempat Pembakaran Sampah berbasis Website

Muhammad Raehan Maulana, Hafiz Muhardi, Dwi Marisa Midyanti — Sun, 27 Apr 2025 00:00:00 +0700

Waste accumulation is a common problem caused by low public awareness of waste management and suboptimal waste handling by sanitation workers. Waste incineration can serve as a solution, but it has the potential to cause undesirable consequences if not properly supervised; therefore, a monitoring system capable of tracking the waste incineration process in real-time is necessary. With the rapid advancement of technology, wireless technology can be utilized to address this issue. However, most wireless technologies still rely on internet networks, making them less effective in areas with limited internet connectivity. Therefore, this study proposes the use of LoRa technology as a solution for data transmission without reliance on an internet network. The developed system can monitor waste incineration sites using MQ-2 sensors and flame sensors in real-time via a website, utilizing the LoRa SX1278 data transmission system. Implementation results show that the monitoring system can detect residual smoke and fire from waste incineration sites and monitor them up to a distance of 150 meters. Signal quality in LoRa SX1278 transmission is expressed in RSSI (Received Signal Strength Indicator) units, with an “excellent” signal category achieved up to a distance of 40 meters in the tests conducted.

Implementation of Random Forest Algorithm in a Web Application as an Early Detection Tool for Diabetes

Habibul Fauzan, Elin Haerani, Fitra Kurnia, Novi Yanti — Mon, 28 Apr 2025 00:00:00 +0700

Diabetes is a chronic metabolic disease and one of the leading causes of death worldwide, with the number of sufferers projected to reach 1.3 billion by 2050. Delayed diagnosis remains a primary challenge, as nearly half of those affected are unaware of their condition in the early stages, thereby increasing the risk of fatal complications. Data mining approaches using classification algorithms have been widely utilized for early screening. However, the development of medical record models is often hindered by imbalanced data, which causes models to be biased toward the majority class and reduces detection sensitivity for the minority class (patients with diabetes). Furthermore, there is a lack of research integrating these predictive models into responsive application interfaces for end-users. Consequently, this study implements Random Forest optimized with the SMOTE (Synthetic Minority Over-sampling Technique) into a web-based application to serve as a practical early detection tool. Random Forest was selected for its ability to handle complex data and reduce the risk of overfitting. The research stages include data preprocessing, balancing training data using SMOTE, model parameter adjustment through hyperparameter tuning with Grid Search, and the development of a client-server architecture using AstroJS and Flask. The evaluation results demonstrate that the use of SMOTE significantly improves the model's ability to identify the minority class. The model achieved a Recall of 75.0% and an overall accuracy of 95.8%, effectively minimizing False Negative errors. The developed application was verified through Black Box Testing and was declared successful as a responsive and accessible early detection tool for both healthcare professionals and the general public.

Efficiency Analysis of the U-Net Architecture with a MobileNetV2 Encoder for Coffee Leaf Rust Segmentation.

Muhammad Adeva, Faisal Muttaqin, Budi Mukhamad Mulyo — Mon, 28 Apr 2025 00:00:00 +0700

Coffee Leaf Rust (Hemileia vastatrix) poses a serious threat to Robusta coffee productivity. Manual identification is often slow and subjective, while standard Deep Learning segmentation methods like U-Net with VGG16 encoder bear heavy computational loads (~24.89 million parameters), hindering deployment on resource-constrained devices. This study aims to optimize computational efficiency by proposing a Lightweight U-Net architecture based on the MobileNetV2 encoder. The model's performance was comparatively evaluated against the VGG16 baseline using the PlantSeg public dataset. Experimental results show that MobileNetV2 integration successfully reduced model size massively by 96% (to ~0.95 million parameters) and accelerated inference time by ~20% (76.28 ms). Although there was a slight F1-Score decrease of 0.3% compared to the baseline, the proposed architecture offers the best trade-off between efficiency and accuracy, making it a viable solution for mobile implementation

Development Of Mobile Library Applications In The Lancang Kuning University Environment

Rahmat Hidayat, Yufi Darmayunata — Mon, 28 Apr 2025 00:00:00 +0700

Libraries play an essential role in supporting academic activities in higher education. However, the limitations of web-based systems have hindered students of Universitas Lancang Kuning from optimally accessing digital collections via mobile devices. This study aims to develop a mobile library application for Android integrated with SLiMS as a data management API. The research employed the prototype method, involving requirements analysis, system design using UML, interface design, prototype development, and user feedback testing. The results show that the application runs properly on the Android platform, integrates with SLiMS for data communication, and enables users to access book collections and read digital books directly. The application performance depends on internet stability, but its features support efficiency and accessibility of library services. Therefore, this mobile application is expected to be a solution for the digitalization of library services at Universitas Lancang Kuning.

Application of Classification Algorithms in Data Mining in Various Case Studies: Literature Review

Hibatullah Naufal Ramadhan, Hasanatul Fu'adah Amran — Mon, 28 Apr 2025 00:00:00 +0700

Data mining is one of the most widely used approaches in the field of Information Technology to extract knowledge from large data sets. One of the main techniques in data mining is the classification method, which aims to predict a particular class or category based on available attributes. Various classification algorithms such as Naïve Bayes, Decision Tree (C4.5), Random Forest, Support Vector Machine, and Artificial Neural Network have been applied in various research domains, including health, education, government, agriculture, and cybersecurity. Differences in data characteristics and methods used cause variations in performance in each study. Therefore, this study aims to conduct a literature review on the application of data mining with classification methods in various data prediction and classification cases. The research method used is a literature review by examining eleven scientific articles from accredited national journals. The results of the study show that the Naïve Bayes and Decision Tree algorithms are the most frequently used methods due to their ease of implementation and interpretation, while Random Forest and Support Vector Machine tend to provide more stable performance on data with high complexity.

Implementing Victoriametrics as A Time Series Database For Kubernetes Cluster Monitoring

Roby Yasir Amri, Nungky Awang Chandra, Mohamad Yusuf — Mon, 28 Apr 2025 00:00:00 +0700

Infrastructure monitoring is a critical component in managing Kubernetes clusters, particularly for ensuring service availability and analyzing system performance. As the complexity and scale of infrastructure increase, monitoring systems are required to efficiently handle large volumes of metric data. This study aims to analyze the performance of VictoriaMetrics as a time-series database within Kubernetes monitoring systems and compare it with Prometheus based on resource usage. The research employs a quantitative approach with benchmark experiments conducted under three load scenarios: 500, 750, and 1000 target hosts. The analyzed parameters include CPU usage, memory consumption, and storage capacity. The results indicate significant differences in resource efficiency, where VictoriaMetrics maintains CPU usage between 2–10% across all scenarios, substantially lower than Prometheus, which reaches 12–24%. In terms of memory consumption, VictoriaMetrics requires only 21–27%, whereas Prometheus increases to 41–67%. For storage usage, VictoriaMetrics consumes 5–13 GB, while Prometheus requires 13–45 GB. These findings are expected to serve as a reference for organizations in selecting an appropriate monitoring solution that aligns with their Kubernetes infrastructure scale and requirements.

Implementation of Blockchain in Medical Data Security: a Systematic Review of Publish or Perish

Mon, 28 Apr 2025 00:00:00 +0700

The digitalization of healthcare services offers substantial opportunities to improve efficiency and quality of care; however, it also introduces significant challenges related to the security, privacy, and integrity of medical data. The literature indicates that conventional Electronic Health Record (EHR) systems remain vulnerable to data breaches, information manipulation, and system failures due to their centralized architecture. This review examines the development of blockchain technology as a potential solution for modern healthcare data management. The study employs a systematic literature review using the Publish or Perish approach, with Google Scholar as the data source to identify relevant scientific articles, which were subsequently screened based on predefined inclusion and exclusion criteria and analyzed qualitatively. Owing to its characteristics of decentralization, immutability, strong cryptography, and the use of smart contracts for access control, blockchain offers significant improvements in medical data security, transparency, and interoperability. Its applications have been reported in EHR systems, telemedicine, pharmaceutical supply chains, medical imaging, and clinical trial data management. Nevertheless, several limitations continue to hinder widespread adoption, including scalability issues, computational overhead, integration complexity with legacy systems, the transparency–privacy trade-off, and regulatory challenges such as compliance with data protection laws and international standards. Future research trends point toward the integration of blockchain with artificial intelligence, the Internet of Medical Things (IoMT), and federated learning, as well as the development of lightweight blockchain solutions for resource-constrained environments. Overall, blockchain demonstrates considerable potential to strengthen the security and reliability of healthcare information systems; however, its implementation requires a gradual, standardized, and regulation-compliant approach.

Tinjauan Penerapan CNN dan YOLO Pada Pengolahan Citra Agraria Medis Insdutri Cerdas

Robby Saputra, Yan Rianto, Muhammad Romadhona Kusuma — Wed, 30 Apr 2025 00:00:00 +0700

The Rapid growth of Artificial Intelligence (AI), particularly Deep Learning, is driving significant transformations in digital image processing in the argricultural, medical, and smart industrial sectors. Two approaches are most dominant in this research Convolutional Neural Network (CNN) for image classification and You Only Look Once (YOLO) for real-time object detection. The purpose of this reasearch is to systematically review the application, performance, and defense of CNN and YOLO in various domains with different data characteristics. The method used is a Systematic Literatur Review (SLR) of the latest relevant scientific publications, focusing on evaluation matrics such as accuracy, pression, recall, F1-score, and Mean Average Precision (mAP). The review results show that CNN excels in image classification tasks with a high level of accuracy, especially on data with relatively stabel visual patterns, while YOLO is more effective in applications that demand inference speed and direct object detection. However, several major limitations were found, including decreased performance in extreme lighting conditions, complex backgrounds, small objects, and visual similarity between classes. It is concluded that the choice of architecture must be adjusted to the characteristics of the data and application needs,

Pemodelan dan Prediksi Tingkat Kemiskinan Provinsi Sumatera Barat Menggunakan Support Vector Machine

Melani Septina Putri, Satrio Junaidi, Ainil Mardiyah — Sun, 04 May 2025 00:00:00 +0700

This research is motivated by the problem of poverty distribution in West Sumatra Province, which still varies between regions. The objectives of this study are to build a prediction model using the Support Vector Machine (SVM) algorithm, evaluate the model's performance, and implement the prediction results in the form of an interactive dashboard to support local government decision-making. The study uses secondary data from the Central Statistics Agency (BPS) of West Sumatra Province for the period 2015–2024, covering 19 districts/cities. The dependent variable is the percentage of poor people (P0), while the independent variables consist of seven socio-economic indicators. The method used refers to the CRISP-DM stages. In the data preparation stage, missing values are handled using median imputation, outliers are handled using winsorizing, standardization is carried out using Z-Score, and the addition of a one-period lag variable (P0_lag1). The data is divided into training data (2015–2022) and test data (2023–2024), with parameter optimization using GridSearchCV and TimeSeriesSplit. The results showed that the Support Vector Regression (SVR) model with a radial basis function (RBF) kernel provided the best performance with parameters C=1000, epsilon=0.05, and gamma=0.001. This model produced an MAE value of 0.32, RMSE of 0.36, and R² of 0.98. The implementation of the prediction results in the Streamlit dashboard for the 2025–2030 period showed a downward trend in poverty levels in most regions. This model is considered effective as a basis for planning and evaluating data-based poverty alleviation policies.