Bulletin of Computer Science Research

Implementasi dan Optimasi Sistem Monitoring Kualitas Udara Berbasis Sensor Gas MQ-135 dan SHT21 dengan Metode IoT

2026-05-29T23:27:20+00:00

The decline in air quality caused by industrialization and urbanization requires an automated, real-time, and affordable monitoring method, since manual measurement by environmental health officers is still limited to periodic site visits. This study aims to implement and optimize an air quality monitoring system based on the Internet of Things (IoT) using the MQ-135 gas sensor and the SHT21 temperature-humidity sensor, targeting improved gas-sensor calibration accuracy through linear regression and faster, more reliable data transmission to a MySQL database. The system is designed to detect environmental parameters such as hazardous gas concentrations, temperature, and humidity, which are then transmitted in real-time to a web-based platform for visualization and data analysis. The implementation process involves integrating hardware, including the Arduino Uno and NodeMCU ESP8266 microcontrollers with the MQ-135 and SHT21 sensors, as well as developing software that enables information processing and IoT communication. Testing was conducted to evaluate sensor accuracy, data transmission stability, and system reliability under varying environmental conditions. The results demonstrate that the system can provide accurate information and respond effectively to changes in environmental parameters, sending notifications when gas concentrations exceed predefined thresholds. This system not only serves as a monitoring tool but also as an educational medium to raise awareness of the importance of maintaining air quality. The main contribution of this research is a self-hosted air quality monitoring architecture that combines a quantitatively validated MQ-135 gas-sensor calibration method with an integrated web- and Telegram-based early-warning system. Further development opportunities, such as incorporating predictive analytics and AI for more accurate air quality data analysis, emerge as outcomes of this research.

Evaluasi Kinerja U-Net ResNet34 dan MDSBN: Studi Komparatif untuk Segmentasi Naskah Kuno Indonesia

2026-07-01T02:44:54+00:00

The digitization of ancient documents is an important step in preserving historical and cultural information. However, the resulting images often suffer from degradation, such as stains, uneven background textures, faded ink, and low contrast, making text-background separation difficult. This study compares two deep learning architectures, namely U-Net ResNet34 and the Modified Deep Semantic Binarization Network (MDSBN), for the segmentation of Indonesian ancient documents. The dataset consists of Balinese palm-leaf manuscripts, Sundanese manuscripts, and additional ancient document images obtained from Wikimedia Commons. The experiments were conducted through a learning rate search and batch size sensitivity analysis, and the models were evaluated using the Dice Coefficient, Intersection over Union (IoU), Precision, Recall, and Root Mean Squared Error (RMSE). This study contributes through a controlled evaluation of both architectures using a consistent dataset, preprocessing pipeline, loss function, evaluation metrics, and computational environment, enabling performance differences to be analyzed more objectively. The results show that U-Net ResNet34 achieved its best performance using a learning rate of 5e-5 and a batch size of 16, with a test Dice score of 0.79338 and a test IoU score of 0.65752. It outperformed MDSBN, which achieved its best performance using a learning rate of 1e-6 and a batch size of 32, with a test Dice score of 0.75338 and a test IoU score of 0.60433. The functional advantage of U-Net ResNet34 is associated with the ability of its residual encoder to extract hierarchical features from complex textures and degradation patterns, while its skip connections help preserve the spatial details of thin text strokes. These characteristics make U-Net ResNet34 more adaptive to variations in degradation within the Indonesian ancient document dataset than the more compact MDSBN architecture.

Implementasi Algoritma Prophet dengan Grid Search Hyperparameter Tuning untuk Prediksi Konsumsi Energi Listrik Berbasis IoT

2026-07-01T14:32:26+00:00

?Real-time monitoring of household electricity consumption is not yet sufficient to support adaptive energy management. Therefore, an accurate yet easily interpretable prediction capability is required. This study implements the Prophet algorithm, an additive time series model based on trend and seasonal components, as the core method for predicting daily energy consumption in an Internet of Things (IoT)-based system with per-room granularity. Data were obtained from PZEM-004T sensors and NodeMCU ESP32 modules in three rooms over 30 days, processed through a two-stage grid search procedure for model hyperparameter optimization. The evaluation results show a testing MAPE of 1.05%–2.01% across the three rooms, all falling into the highly accurate category (<10%). Furthermore, the average MAPE difference between the training and testing data reached only 0.42 percentage points, indicating good model generalization without overfitting. Component decomposition analysis reveals that the consumption pattern is dominated by a stable linear trend with a low-amplitude weekly seasonal pattern (±0.06 kWh), thereby providing a higher level of interpretability compared to black-box models. The 30-day-ahead projection yields a total estimated consumption of approximately 384 kWh (~IDR 554,817) for the three rooms, which can be utilized as a basis for budget planning and adaptive electrical load management. The main contribution of this study is a transparent and reproducible Prophet tuning procedure for per-room electricity consumption data with limited historical volume, supported by metrological validation of the acquisition sensor as an input quality assurance step, a context that has not been widely explored in prior Prophet literature

Perancangan Infrastruktur Virtual Laboratory Multi-Layanan Berbasis Cloud Computing Menggunakan Metode NDLC

2026-07-10T04:07:19+00:00

The limited hardware specifications of student devices remain a major obstacle in conducting computation-based practical courses in higher education, particularly in practicums that require applications with high computational resource demands, such as programming software, network simulation tools, and various other supporting applications, which often cannot run optimally on student devices with low specifications. This condition causes practical learning to become inconsistent and highly dependent on the capability of each student's personal device. Based on this problem, this research aims to design and simulate a cloud computing-based Virtual Laboratory system that can be accessed through a web browser without requiring high hardware specifications on the user side, while also contributing an integrated virtual laboratory architecture that combines three cloud computing service models simultaneously, namely Infrastructure as a Service (IaaS), Platform as a Service (PaaS), and Software as a Service (SaaS). As a solution, IaaS is built using Proxmox Virtual Environment (VE) as a type-1 hypervisor, pfSense as a virtual firewall, Ubuntu Server as the service host, and Docker as the containerization platform, along with Cloudflare Tunnel using a Zero Trust architecture to ensure accessibility from outside the campus network without dependence on a static public IP. The system is equipped with multi-role authentication (student, lecturer, admin) based on Student ID Number (NIM) or National Lecturer ID Number (NIDN) verification and OTP verification, a QR code-based Attendance module integrated with the Timetable module, and a main portal built as an installable Progressive Web App (PWA) on user devices. The system was developed using the Network Development Life Cycle (NDLC) method. Preliminary functional testing results show that all services in the Programming Class and Network Class, including the Attendance and Timetable modules, were successfully accessed from outside the campus network, and the multi-role authentication system functioned as designed.

Pola Perilaku Pemain Roblox: Pemodelan Klasifikasi Berbasis Naïve Bayes

2026-07-13T16:46:49+00:00

The development of digital technology has driven the growth of online gaming as a medium for entertainment, social interaction, and creativity development. One platform that has grown rapidly is Roblox, which allows users to play, interact, and create digital content. This diversity of activities causes player behavior characteristics to become increasingly complex, making them difficult to identify manually. Therefore, a machine learning-based approach is needed to classify player behavior more objectively and systematicallys. This study aims to classify Roblox player behavior into four categories, namely active, casual, social, and creative players, using the Naïve Bayes algorithm. This algorithm was chosen because it has a simple and efficient computational process and is suitable for questionnaire data that has been transformed into numerical form. A total of 523 responses were successfully collected, and after the selection and preprocessing stages, 520 data points were obtained to be used as the research dataset. The data were processed through data cleaning, encoding, missing value handling, and dataset splitting using an 80% training data and 20% test data. The results showed that the model achieved an accuracy of 62.5%. Evaluation using precision, recall, and F1-score metrics revealed that The results showed that the model produced an accuracy of 62.5%, with a precision value of 63%, recall of 62%, and F1-score of 62%. Although the accuracy obtained is not yet high, these results indicate that the Naïve Bayes algorithm can be used as a baseline in classifying player behavior based on questionnaire data that has subjective and complex characteristics. The his study contributes by providing a baseline classification model for Roblox player behavior based on questionnaire data, along with insights into player characteristics that can serve as a reference for developers in understanding user behavior, thereby supporting the development of more adaptive features that better align with players' needs.

Analisis Prediksi Rasio Elektrifikasi Rumah Tangga Indonesia Menggunakan Algoritma Prophet Pendekatan Logistic Growth

2026-07-01T02:39:54+00:00

The Indonesian government continues striving for 100% household electricity access, yet regional disparities remain significant, particularly in eastern regions and New Autonomous Regions. Methodologically, forecasting the electrification ratio faces the challenge of data scarcity and the need to keep predictions below the absolute 100% ceiling. Conventional models and standard non-linear approaches such as penalized Logistic Regression have limitations in handling very small univariate time series and often fail to capture trends without a dynamic saturation point. This study therefore proposes the Prophet algorithm with a Logistic Growth approach to forecast the electrification ratio across 38 provinces for the 2026–2030 period. Prophet was selected for its robustness to minimal historical data and missing values, while Logistic Growth sets a logical maximum capacity (cap = 100.5%) so that predictions do not exceed the 100% asymptotic limit. The evaluation results show the model performs with precision in regions with mature historical data, evidenced by a MAPE of 0.41% and RMSE of 0.62 in DKI Jakarta. Conversely, predictions for DOB provinces such as Central Papua show high uncertainty, with errors reaching 34.30% due to inadequate initial data ranges. Projections through 2030 confirm that all provinces on Java remain stable at a 100.00% ratio, while an anomaly is detected in Southwest Papua, which is projected to decline sharply to 34.65%. The main contribution of this study is the first Prophet-Logistic Growth forecasting framework applied to 38 Indonesian provinces. This approach offers a mathematically stable forecasting framework as a basis for government decision-making on energy infrastructure allocation, particularly when combined with field-data verification in data-scarce regions.

Prediksi Time Series Permukaan Air Sungai Menggunakan Long Short-Term Memory-Autoregressive Integrated Moving Average

2026-07-13T16:41:04+00:00

River water level prediction is an essential component of flood early warning systems, as it provides information on potential water level increases before overflow occurs. The Kebonagung River in Sumenep Regency was selected as the study area due to its fluctuating water levels, which are influenced by variations in river discharge and water volume, thereby requiring a prediction model capable of capturing both linear and nonlinear patterns. The Autoregressive Integrated Moving Average (ARIMA) model is effective in modeling linear and seasonal time-series patterns, whereas the Long Short-Term Memory (LSTM) model is capable of learning nonlinear relationships and long-term temporal dependencies with high accuracy. This study aims to compare three prediction approaches: Long Short-Term Memory (LSTM), ARIMA-based discharge integrated with LSTM, and the Hybrid Autoregressive Integrated Moving Average–Long Short-Term Memory (ARIMA–LSTM) model for predicting river water levels using one year of historical data collected at one-hour intervals, consisting of water level, discharge, and water volume variables. Model performance was evaluated using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). The experimental results show that the LSTM model achieved an RMSE of 128.15, an MAE of 96.61, and a MAPE of 3.56%, while the ARIMA (discharge)-LSTM model obtained an RMSE of 127.63, an MAE of 96.33, and a MAPE of 3.55%. The Hybrid ARIMA–LSTM model produced the lowest RMSE of 127.06, although its MAE and MAPE were slightly higher at 96.81 and 3.59%, respectively. These findings indicate that integrating ARIMA and LSTM can reduce large prediction errors, resulting in a lower RMSE; however, the performance improvement over the other models is not substantial because all models exhibit relatively similar error levels. Therefore, the Hybrid ARIMA–LSTM model is better regarded as an approach that improves prediction stability rather than one that delivers a significant increase in predictive accuracy. The main contribution of this study lies in the evaluation of two ARIMA–LSTM integration strategies, namely the use of ARIMA-based discharge predictions as input features for the LSTM model and a residual-based Hybrid ARIMA–LSTM approach, using the same hydrological dataset to enable an objective comparison of the characteristics, strengths, and limitations of each approach. The findings are expected to provide a useful reference for the development of river water level prediction models to support time-series-based flood early warning systems.

Komparasi Kinerja Algoritma Machine Learning dalam Memprediksi Posisi Finish Pembalap Formula 1

2026-06-27T17:52:07+00:00

Formula 1 is a motorsport competition with high complexity in determining drivers’ finish positions. Research on Formula 1 finish position prediction using regression approaches remains limited compared to classification, leaving its performance underexplored. This study aims to compare the performance of Ridge Regression, Support Vector Regression (SVR), and Random Forest Regression in predicting Formula 1 drivers’ finish positions, identify key factors, and predict finish positions at the 2026 Miami Grand Prix. The dataset, sourced from Kaggle, comprises historical Formula 1 data from 2021-2025 and data from the first three 2026 season races. This study applies preprocessing, feature engineering, feature selection, hyperparameter tuning, model evaluation, and interpretation of the best model using Permutation Feature Importance and SHAP. Results show that SVR performed best, with MAE 2.8685, RMSE 4.0919, R² 0.4946, MAPE 37.2987%, and Pearson Correlation Coefficient 0.7233. SVR’s MAE and MAPE outperformed Ridge Regression (MAE 3.1000; MAPE 56.8081%) and Random Forest Regression (MAE 3.1611; MAPE 57.6259%), although this MAPE remains high. Position Qualifying was the most influential feature, followed by Driver Rank Previous, Constructor Rank Previous, and Grid. In the 2026 Miami Grand Prix forecasting, SVR successfully predicted the race winner, although prediction gaps remained considerable for several drivers. This study contributes a comparative framework of linear and non-linear regression models for predicting Formula 1 finish positions, an aspect not widely explored, along with prediction interpretation using PFI and SHAP.

Analyzing Generation Z's Sentiment on Working Hours and Mental Health Using IndoBERT: Evidence from TikTok Discussions

2026-07-24T16:29:52+00:00

Working hours exceeding recommended limits may increase burnout, psychological stress, and sleep disturbances, particularly among Generation Z, who highly value mental health and work-life balance. Although public opinions on this issue are increasingly expressed on social media, most existing studies rely on conventional machine learning methods or lexicon-based labeling, which are less effective at capturing contextual meaning and linguistic nuance in informal social media text. To address this gap, this study proposes a transformer-based sentiment analysis framework that employs a fine-tuned BERT-based Multilingual model as the primary sentiment classifier, rather than merely as an automatic labeling tool, to analyze Generation Z's sentiment toward working hours and mental health based on TikTok comments. A total of 2,203 comments were collected through web scraping and processed using cleaning, normalization, tokenization, and BERT-based zero-shot sentiment labeling before being used to fine-tune the classification model. The model was evaluated using accuracy, precision, recall, and F1-score. The results indicate that negative sentiment dominated the dataset (41.13%), followed by positive (37.49%) and neutral (21.38%) sentiments. Frequently occurring terms, such as working hours, work, and resigning, suggest that users' concerns mainly relate to long working hours and the intention to leave their jobs. The fine-tuned model achieved excellent classification performance, although a gap between training and validation performance indicates the need for improved generalization. These findings demonstrate that BERT-based sentiment analysis can provide valuable insights for organizations, policymakers, and human resource practitioners in developing more flexible working-hour policies and mental health support programs tailored to Generation Z.

Rancang Bangun Aplikasi Quotes Berbasis Web Menggunakan Teknologi MERN (MongoDB, Express.js, React.js, Node.js)

2026-07-21T15:28:28+00:00

Quotes are among the most frequently sought information on the internet, whether for academic assignments, social media, or self-motivation. However, they are scattered across various sources without adequate category and meaning information, while available platforms are generally static and unable to generate new quotes when the desired one is unavailable. This research aims to design and build a web-based quotes application as a centralized platform for managing, searching, storing, and sharing quotes, while applying artificial intelligence to automatically generate and explain them. The application was developed using the MERN stack (MongoDB, Express.js, React.js, Node.js) with the Software Development Life Cycle (SDLC) waterfall model, through requirements analysis, Unified Modeling Language design, implementation, and testing. The artificial intelligence feature utilizes Large Language Model services via public APIs, namely Gemini as the primary provider and Groq as the fallback, while authentication uses a Google account through the OAuth 2.0 protocol. The resulting application supports browsing quotes without logging in, storing favorites, generating and explaining quotes with AI assistance, and quote moderation by an administrator. Functional black-box testing on 41 scenarios showed all valid, and usability testing using the System Usability Scale (SUS) with 25 respondents obtained an average score of 87.5, categorized as grade A+ (excellent). This research is expected to ease public access to inspirational quotes and serve as a reference for developing MERN-based web applications with artificial intelligence.

Rancang Bangun Sistem Informasi Manajemen Keberangkatan Travel Umroh Berbasis Web Menggunakan Metode Waterfall

2026-07-21T15:26:02+00:00

Umrah travel management at PT. Azza Wisata Mandiri is still conducted manually through paper archives and separate records, covering participant registration, document collection, file verification, and departure information delivery. These conditions lead to document accumulation, difficulties in data retrieval, recording error risks, and uneven information distribution to pilgrims. This study aims to design and develop a web-based umrah travel departure management information system that integrates the management of pilgrim data, documents, travel packages, payments, and departure schedules. The main contribution of this system lies in integrating online registration, document upload, administrative verification, umrah package management, payment monitoring, automatic notifications, and departure reporting into a single integrated platform. Unlike previous studies that only focused on registration aspects, information provision, or agency data management separately, this system integrates the entire umrah travel management flow from upstream to downstream. Black-box testing of all system functions showed that all features operated according to specified functional req

uirements. Furthermore, User Acceptance Testing (UAT) involving 1 owner, 2 staff, and 3 pilgrims obtained an overall score of 87.40%, classified in the Strongly Agree category. These results indicate that the system was well accepted by users and can support more centralized, accurate, and well-documented administration management and departure monitoring for umrah travel operations.

Penerapan Metode Random Forets dan Decision Tree Dalam Sistem Pendukung Keputusan Smartphone Gaming di bawah Harga Lima Juta

2026-07-10T04:12:43+00:00

The rapid advancement of gaming smartphone technology within an affordable price range (under five million rupiah) often makes it difficult for users to select a device that suits their specific needs and budget. This study aims to develop a Decision Support System (DSS) for classifying and recommending gaming smartphones based on criteria such as chipset, RAM, ROM, battery capacity, screen size, and price. The methodology employed is Knowledge Discovery in Databases (KDD), encompassing selection, cleaning, transformation, data mining, and evaluation. The classification process compares the performance of the Decision Tree and Random Forest algorithms. Testing results indicate that the Random Forest algorithm achieves an accuracy of 100%, outperforming the alternative. The system proves effective and accurate in providing optimal gaming smartphone recommendations. The study's primary contribution is the development of a hardware performance classification model integrated into an interactive web ecosystem, offering consumers a practical solution to translate technical specifications into specific device recommendations

Optimasi Klasifikasi Kelayakan Perizinan Frekuensi Radio Menggunakan Varian Naïve Bayes

2026-07-29T17:00:36+00:00

Efficient management of the radio frequency spectrum requires a fast and accurate license evaluation system to avoid signal interference and ensure regulatory compliance. High submission volumes often cause processing bottlenecks in public licensing services. This study aims to optimize the classification of Radio Station License (ISR) applications into Granted and Rejected classes using variants of the Naïve Bayes algorithm. A real-world dataset comprising 38,954 application records with 26 technical, administrative, and geographic features was used. Data preprocessing involved removing post-decision leakage features, median imputation, label encoding, and feature scaling/transformation. Three algorithm variants were evaluated: Standard Gaussian Naïve Bayes, Gaussian Naïve Bayes with PowerTransformer, and Complement Naïve Bayes. Empirical results show that Gaussian Naïve Bayes with PowerTransformer achieved the highest overall accuracy of 78.32% and a high precision of 88.63% (5,112 True Negatives and 990 True Positives). Conversely, Complement Naïve Bayes effectively handled class imbalance by achieving a significantly higher recall of 73.71% and an F1-Score of 64.21%. In conclusion, transforming non-Gaussian technical features improves precision, while Complement Naïve Bayes optimizes candidate identification for approval. These findings demonstrate that optimized Naïve Bayes variants provide an effective and transparent model for automated decision support systems in spectrum licensing.

Perancangan dan Implementasi Sistem Microlearning Menggunakan Arsitektur Model-View-ViewModel (MVVM)

2026-07-26T18:08:36+00:00

The development of mobile microlearning applications requires a software architecture that supports consistent state management, separation of concerns among components, and application maintainability. Although previous studies on microlearning have largely focused on pedagogical aspects, research addressing software architecture, particularly Model-View-ViewModel (MVVM), in mobile learning applications remains relatively limited. This study aims to design, implement, and evaluate a mobile microlearning system using the Model-View-ViewModel (MVVM) architecture. The system is developed using the Flutter framework on the client side and Laravel as the backend, communicating through a REST API within a client-server architecture. MVVM is applied to separate presentation logic, state management, and the user interface, thereby supporting application modularity, maintainability, and consistent data flow. The study adopts the Design Science Research (DSR) paradigm and Rapid Application Development (RAD) as the software development approach. Functional testing using Black-Box Testing showed that all test scenarios produced the expected results. Evaluation of the artifact within the case study environment also demonstrated that the system supports integrated microlearning content delivery, assignment management, and learning progress tracking. The findings indicate that applying the MVVM architecture supports a more modular application structure and consistent state management in a mobile microlearning system.

Otomatisasi Deteksi Dan Mitigasi Webshell PHP Menggunakan Algoritma Random Forest

2026-07-23T15:21:42+00:00

PHP webshells pose a serious threat to web-based applications, enabling attackers to gain unauthorized access and remotely control servers. Varied obfuscation techniques make these webshells difficult to detect using conventional keyword-based scanning. This research aims to develop a real-time PHP webshell detection and mitigation system based on File Integrity Monitoring (FIM) using the Random Forest algorithm. Feature extraction is derived from VLD opcode function calls (with a regex fallback), PHP variable frequency, webshell signatures, and statistical features (entropy and compression ratio). Processing efficiency is maintained through a streaming method for large files (? 5 MB) and a Watchdog-based FIM mechanism optimized with path filtering, event debouncing, and an asynchronous queue. The research contributes a hybrid feature extraction mechanism combining internal instruction indicators (opcodes) and statistical randomness metrics (entropy) to recognize complex obfuscation patterns and integrates a Random Forest classification model with an active FIM-based mitigation architecture capable of autonomous, automated response actions. Testing on 1,207 samples from a balanced dataset (20% test set) showed the Random Forest model achieving an overall accuracy of 97%. Specifically, the webshell class achieved 96% precision, 99% recall, and a 97% F1-score (555 True Positives and 29 False Negatives), while the normal class achieved 99% precision, 95% recall, and a 97% F1-score (616 True Negatives and 7 False Positives). The study is limited to monitoring PHP source code files, relies on a server environment with the VLD extension installed, and focuses mitigation actions on real-time file quarantine. Functional testing confirmed the system's ability to perform monitoring, classification, quarantine outside the web root, and Telegram Bot notifications automatically and in real-time.

Evaluasi Kinerja Pendekatan Latent Semantic Indexing dan Kombinasi Latent Semantic Indexing - K Nearest Neighbor pada Klasifikasi Dokumen Beban Kerja Dosen

2026-07-21T15:48:38+00:00

The Lecturer Workload (BKD) document is an academic administrative document categorized based on higher education's Tridharma (triple dharma) activities, making it highly valuable as a dataset for document classification research. However, the characteristics of BKD documents which feature diverse academic terms and an imbalanced class distribution pose a distinct challenge in the classification process. This study aims to evaluate and compare the performance of two classification approaches based on Latent Semantic Indexing (LSI): LSI + Cosine Similarity and LSI + K-Nearest Neighbor (KNN). As a benchmark, TF-IDF + KNN is utilized as a baseline method to analyze the impact of LSI on classification performance. The dataset consists of 352 Postgraduate BKD documents from Universitas Negeri Medan, which underwent text extraction, preprocessing (case folding, cleaning, tokenization, stopword removal, stemming, normalization, and BKD keyword enrichment), TF-IDF weighting, and dimensionality reduction using LSI. Evaluation was conducted using Stratified 3-Fold Cross Validation with Accuracy, Precision, Recall, F1-Score, ROC-AUC, and Average Precision (AP) as metrics. The results indicate that the LSI + Cosine Similarity approach delivers the best performance, achieving an Accuracy of 86.3%, F1-Weighted of 0.865, AUC-Macro of 0.965, and AP-Macro of 0.887. Meanwhile, LSI + KNN achieved 84.6% Accuracy, and the TF-IDF + KNN baseline reached 82.6% Accuracy. These findings demonstrate that semantic representation using LSI successfully enhances the quality of BKD document classification, while the Cosine Similarity-based approach proves more stable than KNN on datasets characterized by varied academic terminology and imbalanced class distributions.

Evaluasi Model Latent Dirichlet Allocation untuk Topic Modelling pada Ulasan SIMKOPDES Menggunakan Coherence Score dan Perplexity

2026-07-17T16:26:12+00:00

The digital transformation of Koperasi Desa/Kelurahan Merah Putih is supported through the development of the SIMKOPDES Mobile application as a digital service platform for cooperative members. As the number of users continues to grow, reviews posted on Google Play Store have become an important source of information for evaluating application quality and user satisfaction. The large volume of unstructured textual reviews makes manual analysis inefficient. This study aims to identify the dominant topics in SIMKOPDES Mobile user reviews using Latent Dirichlet Allocation (LDA) and evaluate the quality of the generated topic models based on Coherence Score and Perplexity. User reviews were collected through web scraping, followed by text preprocessing consisting of case folding, tokenization, normalization, stopword removal, and stemming. LDA models were then developed using three different numbers of topics and evaluated using Coherence Score and Perplexity, while Intertopic Distance Map and Word Cloud visualizations were employed to support topic interpretation. The results indicate that the one-topic model achieved the highest Coherence Score of 0.419677, whereas the three-topic model produced the lowest Perplexity value of ?6.2917. Although these evaluation metrics yielded different optimal results, the three-topic model was selected because it generated more distinct, representative, and semantically interpretable topics related to cooperative membership and application benefits, account registration and verification, and user experience. This study demonstrates that selecting the optimal LDA model should not rely solely on statistical evaluation metrics but should also consider semantic interpretability, providing a more comprehensive approach for topic modelling of digital application reviews.

Pengembangan Sistem Informasi Presensi Siswa Berbasis QR Code Menggunakan Framework Laravel untuk Meningkatkan Efisiensi Administrasi Sekolah

2026-07-24T16:24:18+00:00

Digital transformation in education encourages schools to improve attendance management that still relies on manual recording. Manual processes may cause delayed recapitulation, recording errors, data loss, and limited real-time attendance monitoring. This study aims to develop ScanHadir, a web-based student attendance information system using QR Code technology at SMP Kristen Rurukan. The system was developed using the Waterfall method, covering requirements analysis, design, implementation, and testing. Laravel 11, Livewire, Filament, Tailwind CSS, and MySQL/MariaDB were used to implement the application. The resulting features include role-based authentication, student and class data management, learning schedules, unique QR Code generation, attendance scanning and validation, permission requests, attendance correction, and automatic reporting. Black Box Testing indicates that the main functions produce outputs according to the specified functional requirements. The contribution of this study is the integration of QR Code-based attendance recording and school attendance administration in a single interconnected web application.

Implementasi Sistem Manajemen Servis Bengkel Motor Berbasis Android dengan Fitur Monitoring Progres Servis Menggunakan Metode Prototype

2026-08-01T16:52:22+00:00

Digitalization of services in motorcycle repair workshops has become increasingly important to improve operational efficiency and service quality. However, in many workshops, the management of customer data, vehicle records, service transactions, spare parts usage, and the delivery of service progress information are still carried out manually or through instant messaging applications. This condition results in service data not being centrally documented, vehicle service histories being difficult to retrieve, inefficient communication of service progress to customers, and customers being unable to monitor the progress of their vehicle repairs in real time. This study aims to develop an Android-based motorcycle workshop management application equipped with a service progress monitoring feature to integrate the entire service process into a single system. The contribution of this study is the integration of workshop operational management and customer service progress monitoring into a single Android application. The study adopted the Research and Development (R&D) approach, while the Prototype model was employed as the system development method. The application was validated using Black Box Testing to verify that each system function met the specified requirements and the System Usability Scale (SUS) to evaluate usability from the users' perspective. The developed application integrates customer management, vehicle records, service documentation, service progress monitoring, inventory and service management, payment transactions, thermal receipt printing, and email receipt delivery into a single platform. All Black Box Testing scenarios were declared valid, while the SUS evaluation produced an average score of 77.03, which falls into the Acceptable category with Grade B. These results indicate that the application has good usability, is well accepted by users, and supports the transparency of service information for customers

Perbandingan Performa Custom CNN, MobileNetV2, dan InceptionV3 untuk Identifikasi Penyakit Daun Tanaman

2026-07-26T18:21:46+00:00

Batch-size sensitivity in deep learning architectures is often assumed to behave uniformly, yet this study finds an opposite-direction pattern between architectures trained from scratch and those relying on Transfer Learning. Using a large-scale multi-commodity dataset (72,000 images, 72 disease classes from 21 commodities), Custom CNN improves consistently as batch size grows, whereas MobileNetV2 and InceptionV3 built on ImageNet pretrained weights via two-phase Feature Extraction and Fine-Tuning achieve their best performance at the smallest batch size (8), consistent with the sharp-versus-flat minima theoretical framework. The pattern is confirmed statistically using two-proportion z-tests (p<0.0001 across all Custom CNN batch-size comparisons). At its optimal configuration, MobileNetV2 significantly outperforms InceptionV3 (test accuracy 97.57% vs. 96.35%; z=4.24, p<0.0001) while being far more compact (±2.96 million parameters, roughly 8 times smaller). Explainable AI analysis (Confusion Matrix, t-SNE, and Grad-CAM) confirms this advantage stems from more separable class-level feature representations and attention concentrated on lesion regions rather than dataset artifacts, making MobileNetV2 at a small batch size a strong candidate for mobile deployment in smart farming.

Klasifikasi Anemia pada Data Tidak Seimbang Menggunakan Random Forest, SMOTETomek, dan GridSearchCV

2026-06-22T08:56:58+00:00

This study applies a Machine Learning approach to support preliminary anemia identification using image-color information and hemoglobin levels. The main issue in the dataset is an imbalanced class composition, where normal records are more numerous than anemia records and may cause the classifier to learn the majority-class pattern more strongly. To address this issue, Random Forest is integrated with SMOTETomek and GridSearchCV. The Kaggle Anemia Diagnosis dataset contains 104 records, four numerical attributes (%Red Pixel, %Green Pixel, %Blue Pixel, and Hb), and one anemia classification target. The original class distribution consists of 78 normal records and 26 anemia records. SMOTETomek is applied to the training data to balance the classes by generating synthetic minority samples and removing Tomek Links, while GridSearchCV searches for the most appropriate Random Forest hyperparameters. Model performance is assessed using 10-Fold Cross Validation and the Wilcoxon Signed-Rank Test. The proposed model obtains 97.09% accuracy, 97.50% precision, 91.67% recall, and 93.24% F1-score. The Wilcoxon p-value of 0.03125 for accuracy and F1-score indicates a significant difference at the 0.05 level. These findings indicate that the integrated Random Forest, SMOTETomek, and GridSearchCV approach improves anemia classification on imbalanced data.

Rancang Bangun Sistem Reservasi Wisata Alam Berbasis Website dengan Integrasi Algoritma Neural Networks dan Validasi QR-Code

2026-08-04T12:14:21+00:00

The management of tourist destinations in the current digital era faces complex challenges related to operational efficiency and visitor management optimization. Often, tourist destinations experience serious obstacles in predicting accurate visitor volumes, which impacts the unreadiness of infrastructure as well as inefficiencies in the time-consuming manual ticket validation process. This study aims to design and implement a comprehensive website-based reservation system for Syaakirah The View by integrating the Neural Networks algorithm as a smart solution for forecasting future visitor numbers. In developing the system, the methodology used is the Waterfall model. This structured approach is chosen to ensure that each development stage ranging from requirements analysis, system design, implementation, to testing runs systematically and is well-documented. To ensure the reliability and validity of the prediction results, the Neural Networks model is trained using the 10-fold cross-validation technique on historical tourist visit datasets for the 2025–2026 period. This dataset has previously gone through a series of data preprocessing stages, namely data cleaning and normalization, to address noise or inconsistent data, allowing the model to learn visitor behavior patterns more accurately. The results of the study show a significant contribution to the operational activities of the tourist destination. First, the system successfully implements an online reservation feature integrated with QR-Code validation; this mechanism is proven to accelerate ticket verification time by up to 70% compared to previous conventional or manual methods. Second, the constructed Neural Networks model achieves a very satisfactory prediction accuracy level, with a Mean Absolute Percentage Error (MAPE) value of 3.41% or equivalent to an accuracy level of 96.59% on the test data. This achievement empirically proves that the model is capable of forecasting visitor numbers with a very low error rate without experiencing overfitting, thereby providing a more data-driven and precise decision-making foundation for the management of Syaakirah The View.

Kerangka Hybrid Adaptif SVR–PSO untuk Prediksi Tingkat Pengangguran Terbuka

2026-08-04T12:18:13+00:00

Accurate prediction of the Open Unemployment Rate (OUR) is essential for supporting evidence-based employment policies; however, conventional prediction models often struggle to capture nonlinear relationships within panel socioeconomic data. This study proposes an Adaptive Hybrid Support Vector Regression–Particle Swarm Optimization (SVR–PSO) Framework for panel-based unemployment prediction in Central Java. The framework was evaluated using socioeconomic panel data integrating cross-sectional and temporal dimensions, consisting of 315 observations from 35 regencies and cities over the 2017–2025 period. Socioeconomic indicators, including labor force participation, education, minimum wage, human development, regional economic output, poverty, population density, population, and population growth, were used as predictor variables. Particle Swarm Optimization was employed to optimize the SVR hyperparameters, while prediction performance was assessed using MAE, RMSE, MAPE, and R². The proposed framework achieved an MAE of 0.7715, RMSE of 0.9264, MAPE of 16.78%, and R² of 0.6956, outperforming the baseline SVR model by reducing MAE, RMSE, and MAPE by 17.77%, 22.33%, and 11.31%, respectively, while increasing R² by 40.41%. These results demonstrate that the proposed Adaptive Hybrid SVR–PSO Framework improves predictive accuracy and model generalization across heterogeneous socioeconomic panel observations, providing a robust approach for panel-based unemployment prediction and supporting data-driven regional labor market planning.

Integrasi IndoBERT dan NSGA-II untuk Penjadwalan HEMS Berbasis Preferensi Pengguna

2026-08-01T16:40:17+00:00

The increasing use of Internet of Things (IoT) based devices in smart homes is driving the need for a Home Energy Management System (HEMS) capable of optimal and adaptive energy scheduling. However, most HEMS systems still use statistical user preference parameters, thus lacking the ability to understand natural language user instructions. This research proposes the integration of IndoBERT and the Non-dominated Sorting Genetic Algorithm II (NSGA-II) to generate HEMS device scheduling based on Indonesian language user preferences. The research method utilizes user preferences using the IndoBERT indobenchmark/indobert-base-p1 model as a pre-trained feature extractor to identify intents and entities such as devices, operating times, priorities, reminders, and environmental conditions. The extracted results are then converted into preference vectors, comfort weights, and dynamic constraints as a modification of the NSGA-II optimization. To mitigate the impact of friction in the NLP extraction results, the IndoBERT output is processed using a confidence-based weighting mechanism before being used in optimization. The IndoBERT model evaluation results showed Accuracy, Precision, Recall, and F1-Score values ??of 70.49%, with user-level validation reaching 93%. The IndoBERT–NSGA-II integration was able to produce HEMS scheduling that considers energy efficiency, photovoltaic energy utilization, and user comforts.

Aplikasi Web Pembelajaran BISINDO Interaktif dengan Deteksi Gestur Real-Time

2026-08-04T12:05:38+00:00

Indonesian Sign Language (BISINDO) is the primary language of the Deaf community in Indonesia; however, access to BISINDO learning remains limited, particularly in North Sulawesi. This research develops an interactive web-based BISINDO learning application, BISINDO.app, providing real-time feedback on users' sign gestures via webcam. The system uses MediaPipe Hand Landmarker to extract 21 hand landmarks, classified using low-parameter deep learning models: a one-dimensional Convolutional Neural Network (Conv1D) for static gestures and a Bidirectional Long Short-Term Memory (BiLSTM) network for dynamic gestures. All inference runs client-side via TensorFlow.js, so the application requires no additional installation. The application uses React, Node.js/Express, and SQLite, secured with JSON Web Token (JWT) authentication, with content validated together with North Sulawesi's Indonesian Sign Language Center (PUSBISINDO). Training results show accuracies on the validation set (not yet evaluated on an independent test set) of 97.64% (alphabet, 26 classes), 98.01% (static numbers, 11 classes), 87.69% (dynamic numbers, 10 classes), and 99.26% (basic words, 17 classes), with model weights ranging from 165 KB to 650 KB; evaluation on an independent test set is left for future work. A prediction stabilization algorithm using a 0.8 confidence threshold, three-tick consistency, and Exponential Moving Average smoothing produces stable, responsive feedback. Functional testing across 14 scenarios confirms all features work as specified, and a usability evaluation with 13 respondents yielded an average SUS score of 83.19 (Excellent). The main contribution of this research is the integration of real-time client-side gesture recognition with an interactive feedback mechanism within a single web-based BISINDO learning platform. This research is expected to improve BISINDO learning accessibility and support the Deaf community's social inclusion in Indonesia.

Analisis Sentimen Berbasis Aspek pada Opini Publik Pascabencana Hidrometeorologi Menggunakan Model Pre-Trained IndoBERT

2026-08-13T11:14:42+00:00

Aceh Province is highly vulnerable to disasters, particularly hydrometeorological events like floods and landslides. Aceh Disaster Management Agency (BPBA) data shows 418 disasters occurred in 2023 with total losses reaching Rp430 billion; similarly, in 2024, most incidents were hydrometeorological disasters affecting tens of thousands of families. This high frequency triggered a social media reaction surge, producing data volumes unanalyzable via conventional sentiment methods. This study applies Aspect-Based Sentiment Analysis (ABSA) and the pre-trained IndoBERT model (indobenchmark/indobert-base-p1) to analyze public opinion across five disaster management aspects: logistics and assistance, evacuation and shelters, government coordination, infrastructure and reconstruction, and healthcare. A total of 1,000 opinions were gathered from X (Twitter) and YouTube (November 2025–April 2026), pre-processed via indoNLP, yielding 3,319 labeled opinion-aspect pairs. The model used sentence-pair classification, fine-tuned with CrossEntropy Loss (learning rate 2e-5, batch size 16, 3 epochs), evaluated via accuracy, precision, recall, and F1-score (target >85%). Testing shows IndoBERT achieved 91.26% accuracy and an 86.03% F1-score, exceeding targets. Sentiment distribution reveals healthcare triggered the most positive sentiment (89.82%), reflecting public appreciation for medical teams. Conversely, negative sentiment dominated evacuation and shelters (34.45%), indicating public dissatisfaction with evacuation procedures and shelter availability post-disaster in Aceh.

Analisis Sebaran Spasial dan Profil Pemancar Radio Menggunakan Algoritma K-Means dan DBSCAN

2026-08-01T17:26:06+00:00

The rapid growth of radio transmitters has caused frequency spectrum congestion and high license rejection rates in urban areas, while conventional manual licensing evaluation focused on single-criterion analysis struggles to map potential interference accumulation and detect isolated locations. This research aims to analyze the spatial distribution and technical profile of radio transmitters using K-Means and Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithms. A comprehensive dataset comprising 38,776 radio station license records was processed using Z-score standardized spatial attributes (geographic coordinates) and technical parameters (operating frequency, Effective Radiated Power, antenna height, and bandwidth). First-stage K-Means modeling (k=4) segmented four macro profiles: Cluster 0 represents urban high-frequency microwave backhauls (mean frequency 14.72 GHz) with the lowest licensing approval rate (22.28%); Cluster 1 encompasses low-frequency mobile cellular infrastructure (mean frequency 1.95 GHz) with the highest approval rate (83.79%); while Cluster 2 and Cluster 3 cover eastern terrain backhauls (mean frequency 9.22 GHz, approval rate 26.69%) and western mountainous link networks (mean frequency 9.94 GHz, approval rate 25.03%). Furthermore, second-stage spatial clustering using DBSCAN with a 3 km Haversine search radius identified 221 local high-density hotspots (primary mega-hotspot encompassing 14,864 transmitters) and isolated 1,581 noise points (4.08%) representing remote locations. The integration of K-Means and DBSCAN effectively reveals hidden technical patterns and geographic density variations, providing a data-driven zoning framework for automated radio frequency spectrum licensing evaluation.

Client-Side Real-Time Inference untuk Efisiensi Bandwidth: Komparasi Lightweight Super-Resolution ESPCN dan FSRCNN Berbasis Browser pada Web Retail

2026-07-13T16:43:52+00:00

Indonesia's rapidly expanding e-commerce industry demands high-quality product imagery, yet directly distributing high-resolution images imposes a significant bandwidth burden particularly on fluctuating mobile network infrastructures. Prior research on client-side visual processing has focused on video streaming, while Lightweight Super-Resolution (LSR) architecture development has been conducted exclusively in native computing environments leaving an empirical gap in browser-based LSR deployment for retail web platforms. This study bridges that gap by implementing and comparing two LSR architectures ESPCN and FSRCNN executed directly within the browser using TensorFlow.js with WebGL acceleration on the CV Citra Wyrus Sakti retail web platform. Evaluation was conducted on 28 industrial product image samples across three dimensions: bandwidth efficiency, cross-device inference time, and visual restoration quality. Results demonstrate that this approach reduces average bandwidth consumption by 50.1%, directly translating to a Lighthouse Performance score increase from 79 to 99 and a Largest Contentful Paint (LCP) reduction from 3.8 seconds to 0.8 seconds. ESPCN proved 1.84× faster than FSRCNN on laptops and 1.52× faster on mobile devices, while also outperforming in restoring edge sharpness and texture detail without excessive smoothing artifacts. The main contribution of this study is twofold: it provides one of the first empirical comparisons of client-side LSR inference on real-world e-commerce product imagery, and it delivers a reusable browser-based deployment framework combining TensorFlow.js, WebGL acceleration, and an HR–LR preprocessing pipeline that Indonesian retail platforms can adopt without native-side infrastructure changes. ESPCN is recommended as the optimal architecture for balancing bandwidth efficiency with e-commerce visual aesthetics within browser environments.

Komparasi Akurasi Model SARIMA, LSTM, dan GRU dalam Peramalan Indeks Harga Saham S&P 500

2026-08-10T17:08:13+00:00

Movement of S&P 500 stock price index is non-linear and influenced by numerous macroeconomic factors, making closing price forecasting a critical challenge for investors and portfolio managers in investment decision-making. This study aims to determine whether deep learning models outperform classical statistical models in forecasting non-linear patterns in the daily closing price data of the S&P 500 index using a public dataset. The forecasting models employed are Seasonal Autoregressive Integrated Moving Average (SARIMA), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). The three models were independently constructed and trained, then evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination (R-squared). Results show that GRU delivered the best performance, with an MAE of 87.04, RMSE of 100.72, MAPE of 2.94 percent, and R-squared of 0.836, followed by LSTM with an MAE of 105.89, RMSE of 132.13, MAPE of 3.55 percent, and R-squared of 0.718. SARIMA performed considerably worse, with an MAE of 418.77, RMSE of 610.20, MAPE of 13.72 percent, and a negative R-squared of -5.02, indicating its inability to capture the non-linear patterns and strong trends in the data. These findings confirm that deep learning models, particularly GRU, outperform classical statistical models in forecasting highly volatile stock indices, while also offering better computational efficiency than LSTM.

Evaluasi Ketergunaan Flap Gate Berbasis SUS: Perbandingan Staf Kependidikan dan Tenaga Kebersihan

2026-08-02T16:03:31+00:00

The implementation of a flap gate-based physical access control system at the Rectorate Building of UIN Raden Intan Lampung aims to ensure building security. However, neglecting ergonomic aspects risks triggering systemic security breaches due to security-bypassing behaviors, such as tailgating by staff and leaving emergency gates open due to jammed trolleys. This study aims to quantitatively evaluate and compare the usability of the flap gate device based on two internal user groups' perceptions. The evaluation utilized the System Usability Scale (SUS) through a standardized 10-item 5-point Likert scale questionnaire distributed to 30 respondents, selected via purposive sampling for Administrative Staff and a census for Cleaning Staff. The quantitative analysis yielded a global average SUS score of 69.00 (Acceptable/Marginal High, Grade C, Good). Comparative analysis identified a sharp perception gap between Administrative Staff (72.12; Grade B/C) and Cleaning Staff (62.75; Grade D/OK), alongside a drastic decline in the learnability dimension to 60.83. These findings prove that operational bottlenecks do not merely compromise comfort but undermine the building's primary security function. This research offers empirical, theoretical, and practical contributions by proving the causality between ergonomic barriers and security breaches, as well as formulating a hybrid mitigation roadmap. Reconfiguring gate delay intervals, adding lower infrared sensors, and providing a dedicated logistics lane serve as urgent solutions to maintain security integrity.

Implementasi Forward Chaining dan Certainty Factor pada Sistem Pakar Diagnosis Penyakit Kulit Berbasis Android

2026-07-29T17:02:43+00:00

Skin diseases are common health problems in the community and frequently present overlapping symptoms, making independent early identification difficult. Limited access to dermatologists encourages self-medication without adequate information, which may worsen the condition. This study applies the Forward Chaining and Certainty Factor methods to SkinCheck, an Android-based expert system application for the early diagnosis of skin diseases. Forward Chaining serves as a data-driven inference engine to determine disease candidates from the symptom facts selected by the user, while Certainty Factor measures the confidence level of the diagnosis by combining expert confidence weights using the CF Combine formula. The system's knowledge base consists of 20 skin diseases, 52 symptoms, and 59 diagnostic rules validated by a medical professional, complemented by a diagnosis history feature and automatically generated health education articles based on a Large Language Model. The agreement between manual calculations and the application's output indicates that the inference process and confidence level calculations operate consistently with the algorithm design. Functional testing using Black Box Testing indicates that all main features operate validly, while System Usability Scale testing involving 30 respondents obtained an average score of 71.67, falling within the Acceptable category with Grade Scale C. The main contribution of this study lies in applying the combined Forward Chaining and Certainty Factor methods on the Android platform covering 20 skin diseases to provide an early diagnosis accompanied by a measurable confidence level, developing a knowledge base validated by medical professionals, and integrating a diagnosis history feature alongside automated health education articles powered by a Large Language Model (LLM) to enhance user health literacy.

Prediksi Risiko Demam Berdarah Dengue Menggunakan Algoritma Random Forest Berbasis GIS di Kota Tangerang

2026-08-09T15:17:26+00:00

Dengue Hemorrhagic Fever (DHF) remains a major public health challenge in Indonesia, as its transmission is influenced by various factors, including climate conditions, environmental characteristics, population density, and spatial characteristics. Early identification of high-risk areas is needed to support more effective dengue prevention and control strategies. This study aims to develop a Geographic Information System (GIS)-based dengue risk prediction model using the Random Forest algorithm in Tangerang City. The study utilized secondary data consisting of historical dengue case records obtained from the Tangerang City Health Office, rainfall, temperature, humidity, healthcare facilities, and administrative boundary data for the 2023–2025 period. The dengue case data consisted of recorded cases from the Tangerang City Health Office and underwent completeness and consistency checks. The research process included data preprocessing, feature engineering, spatial data integration, model development, and evaluation using Spatial Cross-Validation to reduce potential bias caused by spatial autocorrelation and assess the model's generalization capability. The time-based evaluation resulted in an Area Under the Curve (AUC) of 0.973 and an F1-score of 0.889, while the Spatial Cross-Validation evaluation produced an AUC of 0.956 and an F1-score of 0.892. These results indicate that the model maintained good performance when evaluated across different geographic areas. In addition, the developed GIS-based dashboard provides interactive visualization of dengue risk distribution to support the identification of priority areas and data-driven decision-making for dengue prevention and control in Tangerang City.

Rancang Bangun Sistem Pengelolaan UMKM Berbasis Website Terintegrasi Untuk Mendukung Tata Kelola dan Promosi Digital

2026-07-21T15:31:12+00:00

Data management and promotion for MSMEs in Taba Baru Urban Village are currently handled manually, making business information difficult to access widely and leaving data verification processes unintegrated. While previous studies have developed web-based MSME information systems, they generally lack mechanisms for verification by village authorities and fail to integrate the management of MSME data, products, and reporting into a single platform. This study aims to design and build a web-based MSME management system that integrates account verification, data change verification, product management, and reporting to support the digital promotion of MSMEs. The system was developed using the Waterfall model of the System Development Life Cycle (SDLC), comprising requirements analysis, design, implementation, testing, and maintenance phases. Technologies used include PHP, HTML, CSS, Bootstrap, and MySQL. The resulting system involves three main actors the public, MSME owners, and village staff and provides features for MSME registration, profile and product management, data verification, product catalogs, visitor statistics, and MSME reporting. The novelty of this research lies in the development of a system that integrates registration, account verification, data change verification, product management, report generation, and information presentation into a single web-based platform, thereby supporting both administrative governance at the village level and the digital promotion of MSMEs. Black Box Testing conducted across 18 test scenarios demonstrated that all system functions operated according to functional requirements with a 100% success rate. These results indicate that the system effectively supports integrated MSME data management, verification processes, information provision, and digital promotion within the Taba Baru Urban Village.

Penerapan Data Mining dengan Metode Association Rule Menggunakan Algoritma Apriori Untuk Analisis Pola Barang Minus Bernilai Tinggi

2026-08-02T08:35:18+00:00

Indomaret stores, as part of the modern retail sector, face a high risk of inventory discrepancies, particularly high-value stock losses that may cause significant financial impacts. In practice, audit data are often analyzed manually, making it difficult to identify association patterns among missing items. This study aims to implement data mining using the association rule method with the Apriori algorithm to analyze patterns of high-value missing items at Indomaret stores (case study: IC Team Ajat Sugiantoro). The dataset was obtained from inventory audit results during April to June 2025, focusing on NK-category high-loss items. The analysis was conducted by generating frequent itemsets and association rules based on a minimum support threshold of 30% and a minimum confidence threshold of 70%. The results indicate that the Apriori algorithm can effectively identify relationships among missing items that frequently occur together, which can be used to prioritize stock monitoring and support preventive actions against inventory discrepancies. In addition, this research developed a web-based system that enables users to upload audit data, set support and confidence parameters, and automatically display the association rule results in a more systematic and efficient manner.