Journal of Information System Research (JOSH) https://ejurnal.seminar-id.com/index.php/josh <p align="justify"><strong>Journal of Information System Research (JOSH)</strong>, is a research journal that contains articles in the field of Computer Science. JOSH is published 3 monthly (October, January, April, July) a year with&nbsp;ISSN <a href="http://issn.pdii.lipi.go.id/issn.cgi?daftar&amp;1568311829&amp;1&amp;&amp;">2686-228X (media online)</a>, based on LIPI No 0005.2686228X/JI.3.1/SK.ISSN/2019.09.&nbsp; Articles published go through a Blind Review process by Editorial and Reviewers. JOSH Journal, has been indexed on:&nbsp;<a href="https://scholar.google.com/citations?hl=id&amp;user=qjLmuXUAAAAJ">Google Scholar</a> |&nbsp;<a href="https://garuda.kemdikbud.go.id/journal/view/17991">Portal Garuda</a> |&nbsp;<a href="https://onesearch.id/Search/Results?lookfor=Journal+of+Information+System+Research+%28JOSH%29&amp;type=AllFields&amp;limit=20&amp;sort=relevance">Indonesia One Search</a> |&nbsp;<a href="https://index.pkp.sfu.ca/index.php/browse/index/10165">PKP Index&nbsp;</a> |&nbsp;<a href="https://www.scilit.net/sources/116098">SCILIT</a> |&nbsp;<a href="https://portal.issn.org/resource/ISSN/2686-228X">ROAD</a> | <a href="https://app.dimensions.ai/discover/publication?and_facet_source_title=jour.1422568">Dimensions</a> |&nbsp;<a href="https://sinta.kemdiktisaintek.go.id/journals/profile/8355">Science and Technology Index (SINTA) 4&nbsp;</a>|&nbsp;<a href="https://www.base-search.net/Search/Results?type=all&amp;lookfor=2686-228X&amp;ling=1&amp;oaboost=1&amp;name=&amp;thes=&amp;refid=dcresen&amp;newsearch=1">BASE </a>| <a href="https://www.worldcat.org/search?q=2686-228X&amp;qt=results_page">WorldCat.org</a>&nbsp;|&nbsp;<a href="https://explore.openaire.eu/search/dataprovider?datasourceId=issn__online::a1b58f65a8118c88674c4c6a4b386a64">OpenAIRE</a>.&nbsp;<br><strong>Journal of Information System Research (JOSH)</strong>,&nbsp;successful reaccreditation with a <strong>SINTA rating of 4</strong>&nbsp;through the Decree of the Director General of Strengthening Research and Development of the Ministry of Research, Technology and Higher Education based on number <a href="https://drive.google.com/file/d/1Lq3pCoZZmZwoZMSVsAuCM-0seprhkwee/view?usp=sharing">72/E/KPT/2024</a>, dated April 1, 2024 regarding the results Electronic Scientific Periodic Accreditation Period I 2024 from <strong>Volume 4 No 3 (2023)</strong> to <strong>Volume 9 No 2 (2028)</strong>.</p> en-US <p>Authors who publish with this journal agree to the following terms:</p> <ol> <li class="show">Authors retain copyright and grant the journal right of first publication with the work simultaneously licensed under&nbsp;<a href="http://creativecommons.org/licenses/by/4.0/" rel="license">Creative Commons Attribution 4.0 International License</a>&nbsp;that allows others to share the work with an acknowledgment of the work's authorship and initial publication in this journal.</li> <li class="show">Authors are able to enter into separate, additional contractual arrangements for the non-exclusive distribution of the journal's published version of the work (e.g., post it to an institutional repository or publish it in a book), with an acknowledgment of its initial publication in this journal.</li> <li class="show">Authors are permitted and encouraged to post their work online (e.g., in institutional repositories or on their website) prior to and during the submission process, as it can lead to productive exchanges, as well as earlier and greater citation of published work (Refer to&nbsp;<a href="http://opcit.eprints.org/oacitation-biblio.html" rel="license">The Effect of Open Access</a>).</li> </ol> seminar.id2020@gmail.com (Support Journal) mesran.skom.mkom@gmail.com (Mesran) Sun, 11 Oct 2026 07:42:17 +0700 OJS 3.1.1.4 http://blogs.law.harvard.edu/tech/rss 60 Sistem Lost and Found Berbasis Vision-Language Retrieval dengan Explainable AI untuk Lingkungan Smart Campus https://ejurnal.seminar-id.com/index.php/josh/article/view/10990 <p>Evidence for combined image-text retrieval in campus lost-and-found systems remains limited, and zero-shot CLIP's ability to distinguish item attributes and identities is uncertain. This study evaluates CLIP ViT-B/32 on an item-image collection covering eight categories. The controlled multimodal evaluation uses 404 images, 85 metadata-defined relevance groups as proxy ground truth, and 237 leave-one-out queries. Separate image-only validation uses 387 images from 79 provenance-based Object IDs after duplicate removal and quality control; each Object ID contains multiple images of the same physical item. In Object-ID validation, CLIP achieved Recall@1 of 0.674, Recall@10 of 0.935, and mAP of 0.419 (95% object-cluster bootstrap CI: 0.369–0.469). In multimodal evaluation, image-only achieved mAP 0.624, text-only 0.409, and image+text reached the best mAP of 0.732 at α=0.5, selected in four of five group cross-validation folds. Ablation showed that the single fused-index outperformed the three-index configuration (mAP 0.715 versus 0.692); therefore, three-index is not claimed as an accuracy improvement. Direct Indonesian input to English CLIP achieved mAP 0.163, compared with 0.363 using multilingual CLIP and 0.409 through translation. Attention Rollout is positioned only as a supporting visualization because its deletion-curve AUC of 0.871 lacks a random-occlusion baseline. These results indicate the benefit of image-text fusion under controlled evaluation but do not establish performance for free-form user descriptions or production readiness.</p> Enrico Samuel Djimesha, Regi Fernando Najoan, Argha Orion Silitonga, Wilsen Grivin Mokodaser, George Morris William Tangka ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/10990 Sun, 11 Oct 2026 06:43:03 +0700 Klasifikasi Serangan DDOS Pada Arsitektur Software-Defined Network Menggunakan Stacked Ensemble Learning Berbasis Deep Neural Network https://ejurnal.seminar-id.com/index.php/josh/article/view/11099 <p>Software-Defined Networking (SDN) architecture offers a flexible network management paradigm through the centralization of the control plane. However, this centralization creates a single point of failure that is highly vulnerable to Distributed Denial of Service (DDoS) attacks, particularly those originating from Internet of Things (IoT) botnets. Traditional intrusion detection systems are no longer sufficient to recognize modern attack tactics, while the implementation of Machine Learning and Deep Learning is often hampered by high rates of false positives that degrade the Quality of Service (QoS) for legitimate users. This research proposes the development of a two-tier stacked ensemble learning-based DDoS detection framework to classify traffic with precision. The preprocessing stage drastically reduces data dimensions by extracting the top five features using the Gini Importance metric (covering 71.9% of the information variance). The classification architecture then integrates Random Forest (RF) and Support Vector Machine (SVM) as base learners, along with a Deep Neural Network (DNN) as a meta-learner tasked with synthesizing probability distributions. Experimental results demonstrate superior model performance with Accuracy, Precision, Recall, and F1-Score of 99.97%, as well as an ROC-AUC value of 0.99999976. This architecture successfully achieved zero false negatives and generated only five false alarms out of 19,750 test samples. The main contributions of this research lie in the proposal of a multi-level Stacked Ensemble architecture to distribute the detection load, the application of feature dimension reduction to optimize computation while reducing the False Positive Rate, and empirical evidence of the model’s robustness in maintaining the stability of the SDN network controller without degrading QoS.</p> Dicopran Sisco, Edi Surya Negara ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/11099 Sun, 11 Oct 2026 06:51:46 +0700 Integrasi Analisis Butir Soal dan Klasifikasi Kognitif Untuk Personalisasi Rekomendasi Remedial Adaptif Menggunakan K-Nearest Neighbor https://ejurnal.seminar-id.com/index.php/josh/article/view/11066 <p>In many secondary schools, classroom assessments still use tests that are written by hand, which makes it hard for teachers to ensure each question matches the right level of thinking from Bloom's taxonomy or to check if the questions are statistically reliable. This study tackles that issue by creating a web-based testing system for a madrasah tsanawiyah, which automatically sorts multiple-choice questions, checks their quality using psychometric methods, and offers custom learning suggestions to help students improve. Cognitive-level tags (C1 to C6) are assigned by matching keywords that describe actions, and a kNearest Neighbor classifier uses TF-IDF features to sort items into groups that show more or less complex thinking. Item quality is checked using Classical Test Theory module automatically computes the difficulty index and discrimination index of each item from students answer patterns, and materials that help improve learning are chosen based on each student's weak areas through content-based filtering. The system was developed using the Waterfall model and was tested on 600 items taken from three different subjects. The classifier achieved 86.67% accuracy when using k=11, which was verified through 10- fold cross-validation showing 81.83% ± 5.19%, and all 23 black-box test cases were successfully passed. The results show that using automatic item analysis along with adaptive remediation is possible to implement in assessment practices at the madrasah level. The main contribution of this study is the integration of automatic cognitive classification, CTT-based psychometric analysis, and content-based filtering remedial recommendation into a single real-time examination platform, a combination that has not been found in previous similar studies, which generally address only one of these three aspects separately.</p> Karimah Agustin, Alun Sujjada, Muhammad Ikhsan Thohir ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/11066 Sun, 11 Oct 2026 07:04:11 +0700 Penerapan Algoritma YOLOV5 Berbasis Computer Vision Untuk Pemantauan Real Time Child Part Assembly Bumper Pada Industri Otomotif https://ejurnal.seminar-id.com/index.php/josh/article/view/11131 <p>The assembly process of small-scale components, widely known as the bumper child part assembly stage in the automotive manufacturing industry, is highly vulnerable to human error and daily quality inspection inconsistencies. This vulnerability is primarily driven by excessive operational reliance on manual visual inspection conducted by human operators over extended periods. Such dependency frequently leads to defective products escaping to subsequent production lines, ultimately causing financial losses for the company. To address this critical issue, this study aims to design and implement an intelligent real-time monitoring system based on computer vision technology and the You Only Look Once version 5 (YOLOv5) deep learning algorithm to automatically detect anomalies in vehicle bumper component assembly. The image preprocessing stage is comprehensively performed using the OpenCV library and matrix manipulation to normalize variations in artificial lighting intensity on the production floor. The testbed in this study is focused on an assembly workstation equipped with ring illumination and a controlled conveyor speed to ensure visual data validity. This artificial intelligence system is then integrated into an interactive web monitoring interface that provides instantaneous feedback along with automated inspection logging into a database. Empirical test results in an actual production environment demonstrate that the system achieves an average detection accuracy of 98.6%, a precision of 98.9%, a recall of 99.8%, and an F1-score of 99.3%. The inference time is remarkably short at 12.5 milliseconds, enabling a processing speed of up to 65 frames per second. The principal contribution of this research lies in the development of an integrated inspection framework that synergizes a lightweight YOLOv5 detection model with a real-time Fatigue Analytics instrument and automated database logging. This integrated solution functions not only as a passive visual quality verifier but also as an empirical ergonomics diagnostic tool that suppresses component defect rates below two percent in actual manufacturing environments.</p> Abby Muhammad Titan; Dendy K. Pramudito, Muhamad Fatchan ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/11131 Sun, 11 Oct 2026 00:00:00 +0700 Perbandingan Algoritma Random Forest dan SVM untuk Prediksi Tren Harga Saham BBCA https://ejurnal.seminar-id.com/index.php/josh/article/view/11198 <p>A lack of financial literacy and a fear of investment risk often lead university students to engage in stock speculation without objective decision-making tools. This study aims to develop a machine learning-based predictive model to classify daily stock price movement trends for PT Bank Central Asia Tbk (BBCA) during the 2020-2026 period, a popular investment choice among students. A comparative experimental method is employed to evaluate and compare the performance of the Random Forest (RF) algorithm against the Support Vector Machine (SVM) algorithm using a Radial Basis Function (RBF) kernel. SMA-10, RSI-14, MACD, and Signal Line technical indicators were used to extract 1,151 rows of historical data from 2020 to 2026. The data was then divided chronologically using an 80:20 ratio. According to experimental data, the SVM model outperformed Random Forest (accuracy: 52.41%; precision: 43.15%) in terms of global accuracy (57.88%) and precision (47.69%). Nevertheless, the SVM model turned out to be unduly cautious, producing a poor recall of 24.22% and, as a result, losing many chances to spot rising (bullish) patterns. Random Forest, on the other hand, showed better balanced generalization ability, with an F1-score of 45.99% and a recall of 49.22%. This study introduces the integration of momentum indicators to mitigate FOMO among novice investors. As a result, Random Forest is suggested as the decision engine architecture for creating a system that helps students make investment decisions.</p> Muhammad Hasan Rizky, Dian Hafidh Zulfikar, Hermanto Hermanto ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/11198 Sun, 11 Oct 2026 07:41:29 +0700 Implementasi SaaS Multi-Tenant Berbasis Database-per-Tenant dan Metode FEFO pada Sistem Inventori dan Point of Sale UMKM Bakery https://ejurnal.seminar-id.com/index.php/josh/article/view/11024 <p>Bakery micro, small, and medium-sized enterprises manage products with limited shelf lives, requiring accurate inventory control and product rotation based on expiration dates. Toko Kue Khaerul Akbar continues to face several operational challenges, including manual stock recording, discrepancies in stock-taking results, inadequate prioritization of products nearing expiration, and delays in generating operational reports. This study develops an inventory and Point of Sale system based on the Software as a Service model by implementing a database-per-tenant architecture and the First Expired, First Out method. The system was developed using the Extreme Programming method through four sprints, comprising the development of the multi-tenant foundation, batch-based inventory management, FEFO integration with the POS system, and the Owner Central feature. The FEFO method prioritizes batches according to the earliest expiration date, using the received date as a secondary criterion when expiration dates are identical. Unit testing was conducted on 12 test cases organized into four scenario groups, with all cases successfully passing. In addition, black-box testing involving 31 scenarios achieved a fulfillment rate of 100%. Isolation testing confirmed that cross-tenant access was successfully denied. An evaluation involving 10 respondents yielded a UAT score of 4.4 out of 5.0 and a SUS score of 81.7 out of 100. The developed system improves inventory traceability, prioritizes products according to expiration proximity, provides real-time operational reports, and enables centralized monitoring of multiple business units while preserving the isolation of operational data across tenants. The contribution of this study is the integration of a database-per-tenant SaaS architecture, batch-based FEFO allocation directly connected to POS checkout, and Owner Central for cross-unit monitoring while maintaining operational data isolation for each tenant.</p> Rizki Wahyu, Jefry Sunupurwa Asri, Nizirwan Anwar, Adi Widiantono ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/11024 Sun, 18 Oct 2026 21:28:23 +0700 Klasifikasi Hierarkis dengan Kalibrasi Probabilitas untuk Penyakit Daun Singkong Menggunakan Fusi Fitur dan Random Forest https://ejurnal.seminar-id.com/index.php/josh/article/view/11238 <p>Field-image classification of cassava leaf disease requires trustworthy probabilities and a mechanism for withholding uncertain decisions, rather than label accuracy alone. This study evaluates HSV, GLCM, and MobileNetV3-Small embedding fusion in a two-stage hierarchical Random Forest with probability calibration and selective prediction. A total of 7,508 CCMT images were evaluated using outer stratified 5-fold cross-validation. Model comparisons report mean ± standard deviation across five outer folds. End-to-end MobileNetV3-Small achieved 93.90±0.73% accuracy, 94.64±0.59% balanced accuracy, and 94.23±0.73% Macro F1, exceeding the Hierarchical Hybrid Random Forest at 88.28±1.12%, 89.66±0.76%, and 88.79±1.07%, respectively. The 1.03-percentage-point Macro F1 difference between the hierarchical and Flat Hybrid models was not significant at the 5% level (two-sided Wilcoxon, p=0.125). A separate calibration and selective-prediction experiment divided each outer-train fold into 70% model-fit, 15% calibrator-fit, and 15% threshold-selection partitions, all disjoint from the outer-test fold. Simplex-temperature yielded the lowest ECE (0.54%), whereas sigmoid performed better on log loss and Brier score. Thresholds selected on the held-out threshold-selection partition yielded 75.87% outer-test coverage, 95.40% accepted accuracy, and 95.58% accepted Macro F1. Selective accuracy applies only to accepted samples, not all data or guaranteed future performance; the contribution is evaluation of the coverage–risk trade-off, not improved intrinsic accuracy.</p> Regi Fernando Najoan, Enrico Samuel Djimesha, Argha Orion Silitonga, Wilsen Grivin Mokodaser, George William Tangka ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/11238 Sun, 18 Oct 2026 00:00:00 +0700 Smart Assistant System for Blind Users Based on Vision-Language Models and Contextual Location Assistance https://ejurnal.seminar-id.com/index.php/josh/article/view/9646 <p>In this study, we designed and developed an AI-based Smart Assistant for visually impaired users at Panti Sosial GMIM Baertemeus. The proposed smart assistant system integrates a Vision-Language Model (VLM) to identify and describe visual objects, and Contextual Location Assistance (CLA) to provide directional and positional information based on geographic data. This research followed the Prototyping method which consists of some user needs that had been identified, such as system design, prototype development and testing with visually impaired users. The implementation results show that the proposed system can recognize objects and their surroundings and deliver contextual voice feedback adapted to the user’s location. The CLA feature also allows administrators to monitor user positions in real time via a web dashboard. This system improves the independence and safety of visually impaired individuals in daily activities and contributes to the development of inclusive AI-based assistive technologies through the integration of VLM and CLA for accessibility enhancement.</p> Clay Aiken Mangeber Jr, Angeli Marselin Kolmate, Semmy Wellem Taju, Regi Fernando Najoan ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/9646 Sun, 18 Oct 2026 21:52:47 +0700 User Flow Analysis of the Ruangguru App Using Heuristic Evaluation and Cognitive Walkthrough Methods https://ejurnal.seminar-id.com/index.php/josh/article/view/10218 <p>This study evaluates the effectiveness of user flow in the Ruangguru application by integrating Heuristic Evaluation (HE) and Cognitive Walkthrough (CW). HE was conducted by two UI/UX experts, while CW involved 62 students from lower secondary, upper secondary, and higher education levels. The research stages comprised problem identification, literature review, HE/CW scenario development, data collection, analysis, and formulation of design recommendations. HE identified 23 violations of Nielsen’s principles with an average severity rating of 2.74. The most critical issues occurred in System Status Visibility (SR = 3.2), Consistency and Standards (SR = 3.0), and Recognition rather than Recall (SR = 2.8), particularly during registration, content search, and payment transactions. CW produced an overall usability score of 4.08/5, with 63% of respondents willing to reuse the application. Interface Visibility achieved the highest score (4.10), whereas Action–Goal Association recorded the lowest score (3.81). HE exposed structural interface deficiencies, while CW revealed cognitive barriers during task execution. The combined analysis generated five recommendations: design-system standardisation, clearer status indicators, more intuitive labels, structured payment-method grouping, and contextual onboarding. This study contributes an integrated Heuristic Evaluation and Cognitive Walkthrough framework that links expert-identified interface problems with user-experienced cognitive barriers, providing a clear basis for prioritizing user-flow improvements in educational technology applications.</p> Florentina Yuni Arini, Elang Restu Kembara, Fakhri Syihamul Umam, Hasya Anaqi Baskoro, Rosita Anggun Denistyansyah, Adjie Satria Pratama ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/10218 Sun, 18 Oct 2026 22:06:06 +0700 Optimasi Parameter Extreme Learning Machine pada Prediksi Penjualan Kopi Guna Efisiensi Stok UMKM https://ejurnal.seminar-id.com/index.php/josh/article/view/9345 <p>Pait Karanganyar Coffee Shop does not currently implement a structured forecasting technique to accurately estimate and predict demand for its coffee products. Prior studies demonstrate that the Extreme Learning Machine (ELM) method offers superior accuracy and highly efficient computation time compared to traditional models. For instance, a separate study using ARIMA achieved a Mean Absolute Percentage Error (MAPE) of 17.05%, while Long Short-Term Memory (LSTM) recorded a 7.5% MAPE. Based on these advantages, the ELM algorithm was selected for this study. This research utilizes this artificial neural network method to assist in predicting daily coffee beverage sales. Structurally, the ELM model is comprised of an input layer, an output layer, and a single connecting hidden layer. A dataset containing 215 daily sales data samples was collected at Pait Coffee Shop from April to October 2025 for testing purposes. The results and error values analysis using the MAPE method demonstrated an initial average error of 1.06%. Furthermore, a rigorous comparison between the actual and predicted data revealed that the optimal configuration of input features and hidden neurons achieved an average MAPE value of 3.39%. Ultimately, this research contributes to providing an efficient inventory forecasting model for retail coffee shop MSMEs to accurately minimize raw material waste. With this best configuration, an error rate measured using MAPE calculation was 3.39%. This research contributes to providing an efficient inventory forecasting model for retail coffee shop MSMEs to minimize raw material waste accurately.</p> Heribertus Ary Setyadi, Pudji Widodo, Supriyanta Supriyanta ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/9345 Fri, 23 Oct 2026 22:31:02 +0700 Pemodelan Spasial-Temporal Kejadian Banjir Menggunakan Random Forest Regression dan Long Short-Term Memory untuk Mitigasi https://ejurnal.seminar-id.com/index.php/josh/article/view/11277 <p>This study evaluates a spatial-temporal flood-event modeling framework using Random Forest (RF) and Long Short-Term Memory (LSTM). RF estimates the number of flood events in 2025 across 19 districts using five disaster indicators from 2024 and mean elevation. Dry-season intake discharge was excluded because the available dam inventory does not provide a consistent series of observed discharge values. RF was evaluated using leave-one-out cross-validation (LOOCV). Reproduction of the BPBD + elevation model yielded MAE 2.122, RMSE 2.650, and R² -0.531. LSTM modeled monthly rainfall from two BMKG stations for August 2024–July 2026. The three-month look-back configuration achieved MAE 117.34 mm, RMSE 157.09 mm, and R² −0.313, compared with a persistence baseline MAE of 153.85 mm and RMSE of 216.14 mm. Given only 19 spatial units and 24 monthly rainfall observations, RF and LSTM are framed as a decision-support framework rather than a validated operational prediction system. The findings highlight the importance of out-of-sample evaluation, data-availability audits, and transparent reporting of limitations. The contribution of this study is to evaluate RF using LOOCV and topographic-feature comparisons, report prediction errors at the district level, compare LSTM with a persistence baseline, and audit hydrological data limitations.</p> Mochamad Amry Assiva, Hadi Tanuji, Eka Renny Noviati Wahyuni, Muhammad Fa'iq Robbani ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/11277 Fri, 23 Oct 2026 22:44:32 +0700 Analisis Pengelompokan Spasial dan Profil Digitalisasi Usaha Mikro, Kecil, dan Menengah Menggunakan DBSCAN https://ejurnal.seminar-id.com/index.php/josh/article/view/11269 <p>This study addresses the uneven adoption of digital technology among Micro, Small, and Medium Enterprises (MSMEs) in Depok City and the limited information linking their geographic distribution with their digitalization characteristics. Such information is important for planning area-based business support, particularly when training and mentoring resources are limited. This study aims to identify the spatial distribution patterns of MSMEs in Depok City using the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm and describe the digitalization profiles of the resulting groups. An exploratory quantitative approach was applied to a dataset of 50 MSMEs across 11 districts. This number was used for exploratory analysis based on the available dataset and is not intended to represent the entire MSME population of Depok City. Clustering was performed exclusively using latitude and longitude coordinates with the Haversine distance metric, while digitalization was examined after cluster formation (post-hoc profiling) through indicators of social media, marketplace, and website adoption. K-Distance Graph analysis and sensitivity testing selected Eps = 2.1 km and min_samples = 4. DBSCAN identified four clusters containing 45 MSMEs, while five observations were classified as noise. Evaluation produced a Silhouette Score of 0.3223 and a Davies-Bouldin Index of 0.6698, indicating that cluster separation should be interpreted cautiously. The highest average digitalization score was found in Cluster 0 (1.72), while the lowest was found in Cluster 3 (0.20). The Kruskal–Wallis test indicated a statistically significant difference in digitalization score distributions across the four main clusters (H = 14.479; p = 0.0023). These findings provide preliminary insight into variations in digital channel adoption among geographically clustered MSMEs, but they do not establish that location causes differences in digitalization. In practice, the results may support needs assessment, collective training for geographically proximate businesses, and the prioritization of MSME mentoring. Further research should use larger and more representative samples and expand digitalization indicators to measure the intensity and quality of digital channel use.</p> Haris Satria Putra, Yoshua Ardy Putra ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/11269 Thu, 29 Oct 2026 21:23:00 +0700 Penerapan POS Berbasis Web dengan Mekanisme Rollback untuk Efisiensi Transaksi UMKM https://ejurnal.seminar-id.com/index.php/josh/article/view/11299 <p>Kiokovi is an F&amp;B micro-enterprise still relying on manual transaction recording, making it vulnerable to calculation errors and stock data inconsistencies. Previous studies on Point of Sale (POS) systems for MSMEs have often overlooked transactional integrity in data recovery functions (such as automatic stock rollback during cancellations). This study addresses this gap by implementing a web-based cashier application using PHP, JavaScript, and MySQL. The Waterfall method was utilized as a structured framework for the initial digitalization phase, strictly focusing on database integrity before future system expansion. The developed system, Kiokovi POS, applies CASCADE relations to automate stock adjustments and recovery. System validation involved Black Box Testing, encompassing boundary testing and error handling scenarios. The implementation yielded measurable efficiency: transaction processing time decreased from an average of 420 seconds to 45 seconds. Technically, the system achieved 100% automation in stock recovery during transaction modifications, secured user data via encryption, and ensured the accuracy of operational data for the owner.</p> Elank Prabaswara Moeviedhie putra, Toni Wijanarko Adi Putra, Arsito Ari Kuncoro ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/11299 Thu, 29 Oct 2026 21:42:50 +0700 True Structured Network Surgery of EfficientNet-B3 with Focal Loss for Highly Imbalanced Dermoscopic Image Analysis https://ejurnal.seminar-id.com/index.php/josh/article/view/11161 <p>Skin tumor classification relies heavily on deep convolutional neural networks, which are computationally expensive for edge deployment. This study proposes a resource-efficient pipeline using true structured network surgery on an EfficientNet-B3 architecture. To address the extreme class imbalance in the ISIC dataset, we implemented minority class oversampling alongside a custom Binary Focal Loss function. Our experiments evaluated 0%, 20%, and 50% structural channel pruning utilizing L<sub>2</sub>-norm magnitude importance. The baseline model achieved an Area Under the Curve (AUC) of 0.9314 with 10.70 million parameters. At 50% pruning, the parameter footprint dropped by 74% to 2.79 million, and floating-point operations (FLOPs) decreased by 71% to 0.55 G. Despite these reductions, latency metrics indicated a hardware execution plateau on an NVIDIA RTX 3060 Laptop GPU, stabilizing at approximately 12.93 ms per image. By dynamically optimizing the decision threshold using Youden's J statistic, the 50% pruned model maintained a high malignant sensitivity (recall) of 81.59%, despite an overall accuracy reduction to 76.69%. These results demonstrate that combining physical tensor pruning with focal loss yields highly compressed models capable of robust malignant lesion recall. Furthermore, post-hoc visual interpretability via Grad-CAM verified that these compressed architectures retained their clinical diagnostic integrity by accurately localizing genuine pathological features. The primary contributions of this research include the demonstration of true structural channel pruning to reduce model parameters by 74% for edge deployment, the preservation of an 81.59% malignant sensitivity in a highly compressed network using Binary Focal Loss and dynamic thresholding via Youden's J statistic, and the verification of post-pruning diagnostic integrity using Grad-CAM. This pipeline presents a viable pathway for deploying lightweight, accurate dermatological diagnostic tools in resource-constrained clinical environments.</p> Argha Orion Silitonga ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/11161 Thu, 29 Oct 2026 22:06:37 +0700 Komparasi Kinerja Decision Tree dan Random Forest untuk Menentukan Kelayakan Kandidat Peserta Lomba Robotik https://ejurnal.seminar-id.com/index.php/josh/article/view/11037 <p>The selection process for robotics competition candidates at PT. Sonnic Robotic still relies on instructors’ subjective assessments and lacks standardized criteria, potentially leading to suboptimal decisions. This study aims to compare the performance of the Decision Tree and Random Forest algorithms in classifying the eligibility of robotics competition candidates. The dataset consists of 99 active elementary school students and nine variables, including attendance percentage, monthly scores over five periods, and creativity, subject comprehension, and independence, assessed using a 1–4 rating scale. The data are normalized using Min-Max Normalization, and total scores are then calculated based on expert judgment weights to determine the Eligible and Not Eligible labels using a threshold of 75. Modeling is conducted using Altair AI Studio with 10-fold cross-validation. The comparative results show that Random Forest achieves an accuracy of 91.89%, weighted precision of 92.79%, weighted recall of 91.92%, and a weighted F1-score of 91.93%. Meanwhile, Decision Tree achieves an accuracy of 90.89%, weighted precision of 91.99%, weighted recall of 90.25%, and a weighted F1-score of 90.91%. These results indicate that Random Forest performs better across all the evaluated metrics, although the difference in accuracy between the two algorithms is relatively small. However, this study is limited by its sample size of only 99 students. Therefore, the superiority of Random Forest cannot yet be considered statistically significant, and the generalizability of the findings needs to be further evaluated using larger and more diverse datasets. The contribution of this study is to provide a comparative evaluation of both algorithms and identify a classification approach with the potential to support a more objective, measurable, and consistent selection process for robotics competition candidates.</p> Aura Nurshifa, Entin Sutinah, Nani Agustina ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/11037 Thu, 29 Oct 2026 22:22:00 +0700 Comparing Machine Learning and BERT for Classifying Grammatical Error Types in Indonesian EFL Essays https://ejurnal.seminar-id.com/index.php/josh/article/view/11296 <p>Automated writing feedback tools are usually trained on learner corpora from outside Indonesia, so their accuracy on Indonesian students' grammatical errors is unknown. This study compares Multinomial Naive Bayes, a linear Support Vector Machine (SVM), Random Forest, and a fine-tuned BERT model for classifying nine grammatical error types. The models were trained on 48,011 error edits from the BEA-2019 W&amp;I+LOCNESS and FCE corpora, re-annotated with ERRANT 3, and tested in-domain on 5,129 edits and cross-corpus on 1,900 edits from 66 essays by 33 Indonesian university students in the ICNALE Edited Essays. BERT achieved the highest macro-F1 on both test sets, 0.917 in-domain and 0.849 on the Indonesian essays, and McNemar tests confirmed its advantage over each classical model. All models lost 0.035 to 0.068 macro-F1 under the corpus shift. Noun number errors formed 24% of the Indonesian errors, three times their training share, while word form and pronoun were the hardest classes. This study contributes a cross-corpus benchmark of classical and Transformer-based error-type classifiers on Indonesian students' essays, a reproducible pipeline that labels both corpora with the same ERRANT version, and a class-level reliability profile for classroom feedback. For English teachers in Indonesia, the results support explicit, contrastive practice on articles and plural marking and the use of BERT for immediate feedback on reliably recognised classes, while its word form and pronoun suggestions are confirmed by the teacher.</p> Meily Ivane Esther Neman, George Morris William Tangka, Regi Fernando Najoan, Enrico Samuel Djimesha, Daniel Semuel Warouw ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/11296 Sat, 31 Oct 2026 00:00:00 +0700 Implementasi Metode Agile SCRUM Dalam Perancangan dan Pembangunan Sistem Informasi Internal Messenger Berbasis Android https://ejurnal.seminar-id.com/index.php/josh/article/view/7925 <p>The need for a secure, controlled, and regulation-compliant internal communication medium is essential for companies, especially considering that the use of third-party messenger applications poses risks such as limited control, potential data breaches, and non-compliance with internal policies. This study aims to design and develop a custom Android-based internal messenger application specifically for PT. Bank Aceh Syariah KCP Cunda. The development process utilized the Agile Scrum methodology, which supports iterative and adaptive development to meet user needs and project dynamics. The development stages include needs analysis, system design, implementation of key features such as user authentication, messaging, internal announcements, user management, and system testing. The result is an internal messenger application that improves communication efficiency among employees, accelerates information dissemination, and enhances data security and compliance with company regulations. The implementation of this application provides the company with full control over features and managed data, while also offering scalability for future development. Thus, this application serves as a strategic solution to support the company’s digital transformation, strengthen competitiveness, and foster sustainable innovation.</p> Muhammad Zuhdi, Dahlan Abdullah, Zalfie Ardian ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/7925 Sat, 31 Oct 2026 00:00:00 +0700 Pengembangan Sistem Informasi Promosi Produk UMKM Konveksi Berbasis Website https://ejurnal.seminar-id.com/index.php/josh/article/view/11316 <p>Advancements in information technology offer opportunities for Micro, Small, and Medium Enterprises (MSMEs) to expand their promotional reach through digital media. Tailors operating as MSMEs in Selagik Village require promotional media capable of presenting structured business information such as tailor profiles, service expertise, and the sewn products they offer. This study aims to develop a promotional information system for garment-making MSMEs in Selagik Village using a website, thereby helping tailors showcase their businesses and facilitating communication between tailors and visitors. The study employs the Research and Development (R&amp;D) method utilizing the Waterfall development model. System testing covered Functional Suitability, Black Box Testing, Performance Efficiency, and Interaction Capability. The Functional Suitability test yielded a result of 92.36%; 27 Black Box test scenarios were successfully executed; and the Interaction Capability test yielded a result of 85.56% based on evaluations by 50 respondents. Meanwhile, the Page Speed Insights test yielded an average score of 73.67 on mobile devices and 97.33 on desktops. The results indicate that the developed promotional information system successfully fulfills its primary function of supporting the promotion of garment-making MSMEs in Selagik Village, while also facilitating the dissemination of business information and communication with consumers.</p> Evi Wulandari, Baiq Desi Dwi Arianti, Jamaluddin Jamaluddin ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/11316 Sat, 31 Oct 2026 00:00:00 +0700 Penerapan Verifikasi Bertingkat pada Pelaporan Aksi Massa Menggunakan Geolocation API dan Algoritma Haversine https://ejurnal.seminar-id.com/index.php/josh/article/view/11009 <p>Information on mass demonstrations across digital channels is fragmented and lacks location-based checking, while open crowdsourcing is prone to repeated reports from the same source. This study aims to design, implement, and evaluate the functionality and usability of a multi-level verification mechanism in PetaAksi, a mass demonstration reporting application, supporting credibility assessment of public reports. The mechanism applies a volunteered geographic information quality assurance approach through three levels: location-origin validation using the Geolocation API, grouping of reports within a 100-meter radius using the Haversine Algorithm with support levels computed from unique sources rather than report counts, and administrator moderation as the final decision. Support levels only assist moderators and do not determine report truthfulness. The system was developed as a Progressive Web App using the ADDIE model. Black-box testing of 28 scenarios matched all expected outputs, and six verification scenarios showed that the system forms report groups, distinguishes supporting sources, and raises support levels when different sources appear within the same radius. The System Usability Scale with 32 respondents yielded a mean score of 74.76 (Grade B). The study used controlled scenarios, lacks field testing, does not measure fake-report detection accuracy, and relies on browser-based source differentiation.</p> Muhammad Iqbal Fadhilah, Yudi Ahmad Hambali, Ani Anisyah ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/11009 Sat, 31 Oct 2026 00:00:00 +0700 Penerapan Soft Voting Ensemble Berbasis LightGBM, XGBoost, dan CatBoost untuk Deteksi Website Phishing Berdasarkan Fitur URL dan Domain https://ejurnal.seminar-id.com/index.php/josh/article/view/11080 <p>Phishing websites are crafted to resemble legitimate services in order to obtain user credentials and other sensitive information. Blacklist-based mechanisms frequently fail to recognise newly emerging sites, while high reported accuracy remains insufficient to characterise operational risk, since misclassifying a malicious site as legitimate carries a greater consequence than the opposite error. This study applies a Soft Voting Ensemble method that combines the class probabilities of LightGBM, XGBoost, and CatBoost, and compares it against individual models, a Stacking Ensemble, and a Multilayer Perceptron. The evaluation employs a public dataset comprising 11,055 instances and 30 predictive features describing URL, domain, certificate, and page content characteristics, with phishing designated as the positive class. Median imputation and MinMax scaling were placed within a pipeline to mitigate data leakage, after which the data were divided using a stratified 90:10 split. Model performance was assessed on the test set as well as through stratified five-fold cross-validation. On the test set, XGBoost attained the highest accuracy of 98.01% with a phishing F1-score of 97.73%, whereas Soft Voting reached 97.83% accuracy and a phishing F1-score of 97.52%, with 472 true positives, 18 false negatives, 6 false positives, and 610 true negatives. Under cross-validation, Soft Voting recorded the highest mean accuracy of 97.25% and mean phishing F1-score of 96.86%. These findings indicate that model superiority may shift as the evaluation scheme changes. The principal contribution lies in a consistent and traceable evaluation design, so that small performance margins are not readily interpreted as generally valid superiority.</p> Rozan Sanura Albary ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/11080 Sat, 31 Oct 2026 00:00:00 +0700 Prediksi Risiko Putus Kuliah Mahasiswa Berdasarkan Atribut Sosio-Ekonomi Menggunakan Ensemble Learning: Studi Eksperimental pada Dataset UCI https://ejurnal.seminar-id.com/index.php/josh/article/view/10259 <p>Student dropout is an important indicator of the success of higher education institutions. Delays in identifying students who are potentially at risk of dropping out may result in academic and financial interventions being provided only after problems have developed. This study aims to compare the performance of three ensemble learning algorithms, namely Random Forest, AdaBoost, and Gradient Boosted Trees (GBT), in predicting student dropout status based on socio-economic attributes. The study uses the secondary Predict Students Dropout and Academic Success dataset from the UCI Machine Learning Repository. After removing students with the Enrolled status, 3,630 records with two classes, Dropout and Graduate, were obtained. The modeling process was conducted using RapidMiner with an 80:20 data split and evaluated using accuracy, sensitivity, specificity, and Area Under the Curve (AUC). The results show that GBT achieved the best performance with an accuracy of 84.16%, sensitivity of 77.46%, specificity of 88.46%, and AUC of 0.912. Information Gain analysis shows that Tuition fees up to date has the highest information value of 0.2145, followed by Scholarship holder at 0.1582 and Debtor at 0.0984. The contribution of this study is to provide empirical evidence on the comparative performance of three ensemble learning algorithms using socio-economic attributes and to identify the most relevant attributes for classifying student dropout status. The findings can serve as a basis for developing an early warning system to support higher education institutions in monitoring students and prioritizing interventions for those potentially at risk of dropping out.</p> Sugeng Hendra Wijaya, Novi Andayaningtyas ##submission.copyrightStatement## http://creativecommons.org/licenses/by/4.0 https://ejurnal.seminar-id.com/index.php/josh/article/view/10259 Sat, 31 Oct 2026 00:00:00 +0700