Sinkron : jurnal dan penelitian teknik informatika

Comparative Study of Johnson and Bellman-Ford for Shortest Path in OpenFlow SDN

Afriza Tri Rizki — 2026-07-05

Software-Defined Networking (SDN) provides a highly programmable architecture specifically by decoupling the control plane from the data plane. The efficiency of the controller in mapping topologies and responding to link failures depends heavily on the shortest-path algorithm used. While previous studies have evaluated algorithms like Bellman-Ford in small scale SDN, empirical comparisons with hybrid algorithms like Johnson in medium-scale dense topologies remain significantly limited. This study aims to provide an empirical comparative evaluation of Johnson and Bellman-Ford algorithms on OpenFlow 1.3 using the RYU controller, analyzing scalability across ring (sparse) and full-mesh (dense) topologies from 10 to 50 nodes. The research methodology relies on experimental emulation using Mininet to test convergence time, throughput, and recovery time during dynamic link failures. The results indicate that in sparse ring topologies, both algorithms achieve similar convergence under 0,06 seconds. However, in dense 50 node full-mesh networks containing 2.450 links, Bellman-Ford demonstrates a faster average convergence of 37,93 seconds compared to Johnson's 47,44 seconds, primarily due to the absence of graph reweighting overhead, despite exhibiting higher variance. Both algorithms maintained stable throughput, and while recovery times generally met the near carrier-grade standard, some scenarios in dense networks reached 60 milliseconds, slightly exceeding the 50 ms threshold. This study evaluates recovery during single link failure scenarios. In conclusion, Bellman-Ford is highly recommended for dense data center infrastructures, while Johnson is optimal for sparse networks requiring instant route recovery.

Customer Complaint Classification at PT Pos Indonesia Manokwari Using Naive Bayes and Random Forest

Rizhmaria Ester Vieta Saphira — 2026-07-05

Customer complaints represent an important source of information for evaluating service quality and improving organizational performance. However, the increasing volume of complaints received by PT Pos Indonesia Manokwari makes manual complaint classification inefficient and time-consuming. This study aims to compare the performance of Naive Bayes and Random Forest algorithms for customer complaint classification using the Term Frequency–Inverse Document Frequency (TF-IDF) feature extraction method. The dataset consisted of 1,490 customer complaint records collected from the Customer Complaint Handling (CCH) system and categorized into twelve complaint classes. The research process included data cleaning, case folding, stopword removal, TF-IDF transformation, dataset splitting, model training, and performance evaluation. The classification models were evaluated using accuracy, precision, recall, F1-weighted score, F1-macro score, and 5-fold cross-validation. The experimental results showed that Random Forest achieved better performance than Naive Bayes. Random Forest obtained an accuracy of 87.92%, precision of 85.22%, recall of 87.92% an F1-weighted score of 86.30%, and an F1-macro score of 70.85%, while Naive Bayes achieved an accuracy of 84.90%, an F1-weighted score of 84.00%, and an F1-macro score of 48.41%. The cross-validation results produced an average accuracy of 71.81%. Although Random Forest achieved the highest hold-out accuracy, the cross-validation results indicate performance variation across different data partitions, which may be caused by class imbalance among complaint categories. These findings demonstrate that Random Forest is more effective for multiclass customer complaint classification and can support the development of automated complaint management systems at PT Pos Indonesia Manokwari.

INTEGRATED PERFORMANCE AND IOT-BASED RELIABILITY ASSESSMENT OF OFF-GRID PV SYSTEMS

Mhd Adi Setiawan Aritonang — 2026-07-05

Island regions often face unreliable electricity supply due to limited grid infrastructure, resulting in high dependence on diesel generators with significant economic and environmental drawbacks. This study aims to evaluate the performance and reliability of an Internet of Things (IoT)-integrated off-grid solar photovoltaic (PV) system implemented in a small island setting. A 400 Wp off-grid PV system was deployed on Pulau Puteri, Indonesia, and equipped with an IoT-based monitoring platform to enable real-time acquisition of operational data. The methodology includes system design, installation, and continuous monitoring over a 30-day period with 5-minute data intervals. System performance was assessed using Performance Ratio (PR) and Capacity Factor (CF). The system produced 1.6–2.1 kWh/day with PR values of 0.75–0.82 and CF of 16.7%–21.9%, indicating efficient and stable operation under tropical conditions. IoT monitoring enabled continuous data acquisition and early anomaly detection, improving operational reliability. The results confirm that the system meets international performance benchmarks and provides a feasible solution for decentralized energy systems in remote islands. The study contributes an integrated framework combining real-time IoT monitoring and standardized performance evaluation using field data.

Tailscale-Based Overlay Network Architecture for Proxmox VE

Fiqih Akbari — 2026-07-09

Remote management of virtualized infrastructure introduces security risk when management services are exposed directly to the public internet. This risk is amplified when testbeds are intended to support sovereign edge computing workloads that require secure, isolated infrastructure. This study designs and evaluates a secure remote management architecture for a Proxmox VE node using a Tailscale overlay network and interface-specific firewall hardening, establishing a foundational infrastructure baseline for sovereign edge computing. The research follows Design Science Research supported by a network engineering evaluation procedure. The artefact was developed through problem identification, topology design, implementation, measurement, and evaluation. Data were collected from Tailscale status checks, Proxmox VE observation, ping latency testing, relay netcheck output, iptables verification, and external port scanning before and after firewall hardening. The Tailscale path achieved an average round-trip time of 0.434 milliseconds with zero packet loss, comparable to the public Internet Protocol path at 0.540 milliseconds with zero packet loss. Before hardening, public scanning detected management ports 22, 2222, and 8006. After applying interface-specific firewall rules, the external scan reported no open ports among the top 1000 ports, while private access to Proxmox VE through the Tailscale interface remained available. The proposed architecture demonstrates that overlay networking must be combined with firewall hardening to remove public management exposure without disrupting authorized remote administration. The result establishes a replicable foundational infrastructure baseline for sovereign edge computing, providing the first stage toward deployment of secure edge computing systems in resource-limited environments.

ERP-BIM Integration Impact on Post-Commissioning Payment Approval: A PLS-SEM Approach

I Gede Upeksa Negara — 2026-07-05

The construction industry continues to face inefficiencies in post-commissioning payment approval, resulting in delayed settlements, disputes, and increased project costs. Although Enterprise Resource Planning (ERP) and Building Information Modeling (BIM) have individually improved project management and information handling, their integrated use for financial process optimization remains insufficiently explored. This study aims to examine the effect of ERP-BIM integration on post-commissioning payment approval performance in construction projects and to identify the most influential determinants. A quantitative approach was employed using a structured questionnaire distributed to 150 construction professionals with direct experience in ERP and BIM implementation. The collected data were analyzed using Partial Least Squares Structural Equation Modeling (PLS-SEM) with SmartPLS. The results show that System Integration Quality has the strongest effect on Data and Workflow Efficiency (β = 0.356; p < 0.001) and also significantly influences Payment Approval Performance (β = 0.301; p < 0.001). ERP Functional Maturity also demonstrates a strong positive effect on Data and Workflow Efficiency (β = 0.291; p < 0.001) and Payment Approval Performance (β = 0.274; p < 0.001). In addition, Data and Workflow Efficiency significantly affects Payment Approval Performance (β = 0.332; p < 0.001) and mediates the relationship between the exogenous constructs and the endogenous variable. The model explains 62.4% of the variance in Data and Workflow Efficiency and 71.3% of the variance in Payment Approval Performance. These findings indicate that ERP-BIM integration has substantial potential to improve financial workflow effectiveness and strengthen payment approval processes in construction projects.

Random Forest-Based Prediction of Self-Reported Headache Complaint Indicators Among College Students Using Daily Activity and IoT Sensor Data

Adrian Halomoan Parerus Simbolon — 2026-07-05

Headache complaints among college students may be associated with daily activity patterns and environmental conditions. This study aimed to model self-reported headache complaint indicators using daily activity questionnaire data and Internet of Things environmental data without positioning the output as a clinical diagnosis. Environmental data were recorded using BME280, BH1750, and MQ-135 sensors, while daily activity data were collected using a self-report questionnaire. Sensor readings were aggregated by date and integrated with questionnaire responses to form 305 records from 59 respondents. Random Forest was optimized using Randomized Search CV and evaluated against Decision Tree and K-Nearest Neighbors under three feature scenarios Internet of Things features, daily activity features, and combined features. SMOTE was applied only to the training data, and model differences were assessed using the McNemar test and Wilcoxon signed-rank test. Random Forest achieved the highest overall performance in the daily activity questionnaire scenario, with 81.52% accuracy, 84.96% F1-score, and 85.32% mean cross-validation F1-score. In the combined scenario, Random Forest obtained 77.17% accuracy and 81.08% F1-score. Statistical testing showed significant differences only in selected McNemar comparisons, while Wilcoxon tests on cross-validation F1-scores were not significant across all comparisons. Daily activity data were more informative than date-level environmental sensor data in this dataset. The findings should be interpreted as exploratory numerical performance results rather than evidence of clinical causality or universal model superiority.

Comparative Sentiment Analysis of GrabFood Reviews Using BiLSTM and BiGRU

Jakasurya Siswoyo — 2026-07-05

The exponential growth of user-generated reviews on digital platforms has made manual sentiment interpretation of Online Food Delivery (OFD) services increasingly impractical. GrabFood, operating within the Grab ecosystem, has accumulated over 16.1 million reviews on the Google Play Store, necessitating an automated and scalable approach to sentiment monitoring. Conventional labeling approaches, including star-rating proxies and lexicon-based annotation, are inadequate for capturing contextual nuance, negation, and informal linguistic patterns prevalent in Indonesian-language OFD reviews. Furthermore, limited research has systematically compared BiLSTM and BiGRU architectures within a transformer-assisted labeling framework for Indonesian OFD sentiment analysis. This study aims to implement RoBERTa-based automatic sentiment labeling and to comparatively evaluate BiLSTM and BiGRU models for three-class sentiment classification of GrabFood reviews. A corpus of 265,500 raw reviews was collected via web scraping, filtered to 17,709 reviews through rigorous preprocessing, and annotated using the w11wo/indonesian-roberta-base-sentiment-classifier. Random Oversampling was applied to address class imbalance. BiLSTM and BiGRU models were trained and benchmarked against Support Vector Machine (SVM) and Naïve Bayes baselines. BiLSTM achieved 86% accuracy while BiGRU attained 85%, both substantially outperforming SVM (82%) and Naïve Bayes (77%). However, BiGRU demonstrated superior convergence speed and more stable per-class performance, particularly on the neutral category (F1: 51% vs. 50%). Transformer-assisted automatic labeling combined with bidirectional recurrent architectures constitutes an effective and scalable pipeline for Indonesian OFD sentiment classification, with neutral sentiment remaining the primary classification challenge.

K-Means and Fully Connected Neural Network for Child Nutritional Status Classification

Rismayanti — 2026-07-09

Stunting remains a persistent child nutrition problem because delayed growth is closely related to long-term health, cognitive, and productivity risks. Manual interpretation of anthropometric measurements using the World Health Organization Z-score standard is clinically valid, yet it becomes inefficient and error-prone when routine records are processed in large numbers. This study develops a child nutritional status classification model by combining K-Means clustering and a fully connected neural network for early identification of stunting, underweight, and wasting. The dataset consisted of toddler anthropometric records from 2021-2024 with sex, age, body weight, and body height attributes. The data were cleaned, standardized, transformed into Z-score indicators, and grouped into 27 clusters representing possible combinations of nutritional status profiles. Cluster membership was then used with Zlen, Zwei, and Zwfl features in a multi-head fully connected neural network. Evaluation on 82 held-out samples showed accuracy values of 91.46% for stunting, 93.90% for underweight, and 98.78% for wasting. Weighted precision, recall, and F1-score were consistently high across the three outputs, while the training curves indicated stable learning without strong overfitting. The proposed hybrid model improves the reliability of child nutrition classification and can support a web-based decision support system for data-driven nutritional screening and intervention planning.

Classification of Traditional Balinese Kites Using CNN for Cultural Preservation

Ni Wayan Sumartini Saraswati — 2026-07-05

The digital preservation of cultural heritage has become increasingly important in sustaining local traditions amid rapid modernization. Balinese traditional kites represent a distinctive form of intangible cultural heritage with unique visual characteristics; however, their identification and classification are still largely based on subjective expertise. This research develops a Convolutional Neural Network (CNN)-based model for image classification to automatically recognize three primary types of Balinese traditional kites: Bebean, Janggan, and Pecukan. Beyond technical implementation, this research contributes to the development of a culturally specific visual dataset, addressing the limited representation of local heritage objects in mainstream computer vision research, which is predominantly based on global datasets of generic objects. A balanced dataset of 2,400 images was constructed and evaluated using 5-Fold Cross Validation to assess model stability and generalization capability. The proposed CNN model achieved an average validation accuracy of 91.5%, with balanced precision, recall, and F1-score across folds. Further evaluation on an independent test set of 282 images resulted in an accuracy of 87.94%, indicating a generalization gap of approximately 4%, which remains within an acceptable range. The results demonstrate that CNN-based classification can effectively support structured digital documentation of traditional kites. This study highlights the potential of computer vision not only as a technical tool, but also as a strategic approach to advancing data-driven cultural preservation and expanding AI applications within localized cultural contexts.

Hybrid Deep Learning Model for Coffee Leaf Disease Detection Using CNN DeiT

Jepri Banjarnahor — 2026-07-05

Coffee production plays a crucial role in the agricultural economy; however, its productivity is significantly affected by plant diseases that are difficult to detect at early stages. Accurate disease identification remains challenging due to subtle visual differences and high intra-class variability in leaf symptoms. To address this problem, this study proposes a hybrid deep learning framework that integrates Convolutional Neural Networks (CNN) and Data-efficient Image Transformers (DeiT) for automated coffee leaf disease classification. The proposed architecture leverages CNN to capture fine-grained local features, while DeiT models global contextual relationships through self-attention mechanisms, enabling a more comprehensive feature representation.

The model is trained and evaluated on a dataset of 6,048 labeled images across four classes: Healthy, Rust, Red Spider, and Leaf Miner. Experimental results demonstrate that the proposed CNN–DeiT model outperforms baseline CNN and Transformer-based approaches, achieving an accuracy of 93.1%, an F1-score of 92.3%, and a ROC-AUC of 95.6%. Robustness analysis shows that performance degradation remains limited (1.6%–3.4%) under various perturbation conditions, while out-of-distribution evaluation indicates strong generalization capability with only a minor accuracy decrease. These findings confirm that the hybrid CNN–Transformer architecture effectively enhances classification performance, robustness, and generalization. This study contributes to the advancement of deep learning methodologies in agricultural image analysis by providing a robust and scalable framework for plant disease classification, with potential applications in precision agriculture and data-driven crop management.

Intelligent Fault Diagnosis in Multi-Setpoint Water Level Systems Using LSTM-Autoencoder

Muhammad Giriarda Abrari — 2026-07-05

Fault detection in multi-setpoint industrial process control systems is complicated by the fact that normal sensor behavior shifts substantially across operating points, making conventional threshold-based alarms and single-condition models unreliable when setpoints change frequently. Recent studies on LSTM-Autoencoder for industrial anomaly detection have demonstrated promising results, yet most are evaluated under fixed operating conditions and do not examine how feature engineering choices affect detection performance across diverse setpoints. This study aims to determine whether physics-informed derived features improve LSTM-AE fault detection performance in a real-time multi-setpoint water level control system, and whether the improvement holds under practical deployment conditions. The proposed framework augments seven raw PLC sensor readings with three derived variables: delta flow, level error, and frequency-per-flow and applies a per-setpoint windowing strategy to prevent cross-setpoint data contamination during training. An ablation study compares the eleven-feature model against a seven-feature baseline under three labeling scenarios reflecting varying preprocessing quality. The eleven-feature model achieves an AUC of 1.0000 and F1-score of 0.9993 under onset-cut evaluation, and reduces the false positive rate from 18.48% to 15.21% under corrected labeling while maintaining perfect recall. Real-time validation across thirty fault injection experiments confirms a 100% detection rate with a mean latency of 6.37 ± 2.04 seconds, 38.2% faster than the baseline. These results confirm that derived features meaningfully improve both classification quality and temporal detection performance, though adaptive thresholding at high-variability setpoints remains an open challenge for future work.

Development Of Smartphone-Based Pornography Addiction Behavior Monitoring Application

Litafira Syahadiyanti — 2026-07-05

The rapid growth of smartphone usage has increased access to various internet content, including pornography-related websites, particularly among adolescents. Existing solutions addressing this issue generally focus on psychological screening or access restriction through parental control systems, which either rely on subjective input or emphasize blocking mechanisms rather than providing insights into actual usage behavior. This study aims to develop MindSafe, a smartphone-based application for monitoring pornography-related domain access activity using a local VPN approach. The proposed system captures domain access activity in real time by intercepting DNS queries at the device level, classifies accessed domains using a blocklist-based approach, and records the results in a structured database. Unlike conventional parental control systems that focus on access restriction, this study introduces a monitoring-oriented approach that emphasizes real-time domain logging and statistical behavior visualization. The system was developed using Agile methodology and evaluated through functional testing and usability assessment. Functional testing confirms that all core features operate as expected. Usability evaluation using the System Usability Scale (SUS) involving 20 respondents resulted in an average score of 76.0, indicating acceptable usability. The results demonstrate that the system provides an objective and privacy-aware monitoring approach, offering a data-driven alternative to existing screening and blocking-based solutions.

Comparative Evaluation of Machine Learning Algorithms for Intrusion Detection Systems

Reza Nismara — 2026-07-05

Performance estimates of intrusion detection models may vary depending on how the training and testing data are separated. Although random split is commonly used in IDS experiments, network traffic often follows time-dependent patterns that differ from one day to another. This study compares random split, single temporal split, and rolling temporal split to examine whether random evaluation produces overly optimistic performance estimates. The CIC-IDS2017 dataset was used because it contains network traffic collected across several days and includes benign as well as malicious activities. The evaluation involved five classical learning models: Decision Tree, Random Forest, Logistic Regression, K-Nearest Neighbors, and Linear Support Vector Machine. The dataset was prepared by combining daily traffic files, removing irrelevant and invalid features, converting labels into binary classes, and applying consistent preprocessing for all models. Performance was measured using accuracy, precision, recall, F1-score, ROC-AUC, PR-AUC, training time, and prediction time. The results show that random split produced very high scores, with several models reaching F1-scores close to 1.0. In contrast, temporal evaluation caused a clear performance decrease, with single temporal F1-scores ranging from approximately 0.60 to 0.71, while rolling temporal validation showed that model performance varied across different chronological testing periods. These findings indicate that random split may overestimate IDS model performance because similar traffic patterns can appear in both training and testing data. Therefore, time-aware evaluation provides a more realistic strategy for assessing IDS model generalization.

Assessing Information System Acquisition and Implementation Using COBIT 4.1 in Housing Developers

Dea Ramadhan — 2026-07-05

Housing development companies plan, construct, and market housing. As technology advances, information systems become crucial, especially in financial matters. However, no comprehensive analysis of financial information system implementation in this sector exists. This study employed a qualitative descriptive approach supported by quantitative assessment (mixed-method), combining observations and semi-structured interviews (qualitative) with a questionnaire-based maturity level calculation (quantitative). Data were collected from three respondents (Finance Manager, IT Staff, and Accounting Staff) directly involved in the system. The maturity level was determined using the COBIT 4.1 framework in the Acquire and Implement (AI) domain. Results show an average maturity of 3.08 Level 3 (Defined Process), meaning procedures are documented and structured. However, several subdomains remain below target AI5 gap 1.30 and AI7 gap 1.10 require improved implementation and control of financial information systems, while AI3 has the smallest gap (0.02). This study contributes to the academic community by (1) providing a detailed assessment of financial information system implementation in housing development companies, (2) identifying maturity gaps in each AI subdomain, and (3) offering practical recommendations for improving IT governance maturity. Routine maintenance, system monitoring, and infrastructure updates are still needed to ensure system stability. Overall, information systems management has not yet reached the targeted maturity level due to reliance on individuals, lack of formal training, and poor procedure communication. Continuous improvement is required to make information systems management more efficient and scalable.

Fish Disease Classification Using MobileNetV3Large Transfer Learning and Fine-Tuning

Dela Fifi Lusiana — 2026-07-05

Fish diseases represent a major challenge in the aquaculture industry as this phenomenon frequently leads to significant economic losses. Manual disease identification requires specialized expertise and is time-consuming in the field. Therefore, this study aims to implement the MobileNetV3Large Deep Learning architecture to automatically identify eight types of fish conditions. This research dataset utilizes 2,400 digital images distributed evenly across eight fish condition categories. Each class consists of 300 image samples, including Bacterial Red disease, Aeromoniasis, Bacterial gill disease, EUS Disease, Fungal diseases Saprolegniasis, Parasitic diseases, White tail disease, and a Healthy Fish group. The dataset was sourced from https://www.kaggle.com/datasets/irfanulhuda/fish-disease-detection-dataset. These conditions include bacterial, fungal, viral, and parasitic infections, as well as healthy fish conditions. The research methodology applies Transfer Learning techniques combined with Fine-Tuning optimization on the last 70 layers. The methodology applies a transfer learning strategy with a data split of 80% for training, 10% for validation, and 10% for testing. This step was taken to adapt the model's weights to the visual characteristics of the fish disease images. The process was evaluated using the Adam optimization function and the Categorical Cross-Entropy loss function. Experimental results demonstrate highly superior model performance on the test data. The MobileNetV3Large model successfully achieved a test accuracy of 92.92% with a loss value of 0.2099. Furthermore, evaluation through the Confusion Matrix and ROC curves yielded an average AUC value of 1.00 across the majority of classes. This figure indicates that the model possesses exceptionally high discrimination capacity and sensitivity. In conclusion, the computational efficiency of the MobileNetV3Large architecture makes this system a highly potential solution. Researchers can implement this model on mobile devices to assist fish farmers in diagnosing diseases quickly and accurately directly at the aquaculture sites

Comparative Analysis of Snort, Suricata, and Random Forest for Flood Detection

Ichdan Maulana Nur Fazri — 2026-07-05

Volumetric Denial of Service (DoS) attacks, particularly SYN Flood and ICMP Flood, remain critical threats to network availability. Signature-based NIDS tools such as Snort and Suricata are widely deployed, yet their trade-offs against machine learning approaches remain underexplored in simultaneous physical-environment studies. This study aims to quantify and compare the performance-accuracy trade-off of Snort 3, Suricata 7, and Random Forest for SYN/ICMP Flood detection on identical physical datasets. Experiments were conducted in a controlled physical laboratory using hping3-generated datasets: 28,930,364 ICMP packets (1.56 GB) and 1,532,301 SYN packets, each captured over 120 seconds. Both NIDS tools were tested in offline PCAP-replay mode. A Random Forest model was trained on 627,788 balanced samples using frame-level features, validated with 5-fold cross-validation. Results: Snort 3 achieved the highest throughput at 987,966 PPS (ICMP) and 240,908 PPS (SYN), while Suricata 7 demonstrated greater detection sensitivity with 148 alerts versus 36 matches in the ICMP scenario. The Random Forest classifier achieved Precision = Recall = F1-score = 1.00 on 125,558 test samples, confirmed by 5-fold cross-validation (99.98% ± 0.01%). Conclusion: A hybrid architecture combining signature-based NIDS as a first-line filter with Random Forest as a secondary validator represents the optimal configuration for volumetric DoS mitigation, balancing throughput and detection accuracy.

Comparison of Naive Bayes and Support Vector Machine for Sentiment Analysis of BPJS Health Service Deactivation

Vitriayanti Payung Allo — 2026-07-05

The deactivation of BPJS Health services has become a topic of public discussion on social media, particularly on platform X, where users frequently express opinions regarding healthcare access and membership status. This study aims to analyze public sentiment toward the deactivation of BPJS Health services and compare the performance of Naive Bayes and Support Vector Machine (SVM) for sentiment classification. The dataset consisted of 8,357 tweets collected from social media X, of which 7,546 tweets were retained after preprocessing, including data cleaning, case folding, tokenizing, stopword removal, and stemming. TF-IDF and FastText were employed as text representation techniques, while model evaluation was conducted using 5-Fold Cross Validation, Grid Search Cross Validation for hyperparameter optimization, and a paired t-test for statistical significance analysis. Classification performance was measured using accuracy, precision, recall, and F1-score metrics. The results showed that negative sentiment dominated public opinion, accounting for 70.63% of the dataset, followed by neutral sentiment (26.48%) and positive sentiment (2.89%). The SVM model with TF-IDF achieved the highest performance, with an accuracy of 80.97%, precision of 80.16%, recall of 80.97%, and F1-score of 79.70%, outperforming Naive Bayes with TF-IDF (79.01%), Naive Bayes with FastText (64.26%), and SVM with FastText (80.31%). Furthermore, a paired t-test confirmed that the performance difference between Naive Bayes and SVM was statistically significant (p = 0.011). These findings indicate that SVM combined with TF-IDF is more effective for sentiment classification of high-dimensional social media text data and provide empirical evidence regarding the effectiveness of different text representation approaches for healthcare policy-related sentiment analysis.

CTRI Framework: Integrating COBIT 4.1 Maturity, Risk Priority, and Transition to COBIT 2019

Kasmala Kasmala — 2026-07-05

This study proposes the CTRI (COBIT Transition and Risk Integration) framework, a novel methodological integration for IT governance evaluation in SMEs. Unlike prior COBIT 4.1 studies that only report maturity gaps, the CTRI framework addresses three critical gaps: (1) lack of actionable recommendations, (2) absence of risk and feasibility considerations, and (3) no systematic bridge from COBIT 4.1 to COBIT 2019. The CTRI framework consists of three integrated layers. Layer 1 quantifies maturity gaps across six COBIT 4.1 domains (PO2, AI2, AI6, DS5, DS11, ME1). Layer 2 introduces the Risk–Feasibility Priority Matrix (RFPM) which calculates a Priority Action Score (PAS) = Risk Impact × Implementation Feasibility, where risk impact is high (3), medium (2), or low (1), and feasibility is easy (3), moderate (2), or difficult (1). Recommendations with PAS ≥ 6 are top priority. Layer 3 provides explicit transition mapping from each COBIT 4.1 recommendation to COBIT 2019 governance objectives and design factors. Applied to a sales application at PT Ciequ (Indonesian SME), data were collected via observation, interviews with three key informants, and documentation review. Findings reveal an average maturity level of 2.45, with largest gaps in AI2 (1.23) and DS5 (0.89). The RFPM prioritizes AI2 (PAS=9), DS5 (PAS=6), and AI6 (PAS=6) as top actions. A three-phase transition roadmap (Stabilize → Standardize → Monitor) to COBIT 2019 is provided, with APO13 (security) and APO05 (portfolio) as priority objectives. This study contributes a reusable, risk-aware, transition-forward methodology that bridges legacy and modern IT governance frameworks for resource-constrained SMEs.

PRIVA: Selective Face Blurring Video App Using YOLOv8-Face and MobileFaceNet

Muhammad Satrio — 2026-07-05

The increasing use of vlog videos on social media creates privacy risks because third-party faces are often unintentionally recorded and distributed without consent. Existing face blurring approaches generally apply uniform anonymization to all detected faces and do not provide an identity-selective mechanism that keeps the content creator visible while blurring other individuals. This study develops PRIVA, a desktop-based selective face blurring application that runs locally without an external AI server. The proposed pipeline integrates YOLOv8n-Face-960 for face detection, MobileFaceNet for face recognition using 512-dimensional embeddings, and Deep SORT for maintaining identity consistency across video frames. Face enrollment is performed through guided multi-pose webcam capture, while video evaluation is conducted on extracted YOLO analysis frames from five real vlog-like test videos. YOLOv8n-Face-960 achieved an overall detection precision of 95.02%, recall of 89.32%, and F1-score of 92.09%. The baseline comparison showed that YOLOv8n-Face-960 achieved a higher mean detection F1-score than MTCNN, while MobileFaceNet provided a smaller and faster recognition model than FaceNet for CPU-based local inference. For correctly detected face instances, PRIVA achieved a system precision of 99.45%, recall of 98.70%, F1-score of 99.08%, and accuracy of 98.50% in determining whether faces should be blurred or kept visible. Processing performance testing showed an average analysis speed of 4.83 FPS, average export speed of 70.05 FPS, and average processing ratio of approximately 2.40 times the original video duration. These results indicate that PRIVA can support practical local identity-selective face blurring for video privacy protection, although detection robustness remains important under low-light, crowded, distant, or partially occluded face conditions.

Academic Chatbot for Campus Information Services Using Retrieval-Augmented Generation

Haddad Alwi Yafie — 2026-07-05

University service centers handle many repetitive queries about academic schedules, registration, and policies stored in internal documents. Manual lookup is inefficient, and answers given by staff can be inconsistent. Rule-based chatbots only handle limited question patterns, while large language models are hard to update and may produce unsupported answers (hallucinations). This research designs an academic chatbot that combines document retrieval with answer generation so that each answer remains traceable to its source. The system extracts text from campus documents, segments it, encodes it using a multilingual embedding model, and stores it in a vector index for context retrieval. A response is generated through an instruction template that confines the output to the retrieved information and includes page references. Evaluation followed a mixed-method design: a quantitative layer measured retrieval quality (Precision@5, Recall@5) and generation quality using the four RAGAS sub-metrics (faithfulness, answer_relevancy, context_precision, context_recall) on a 100-question test set, while a qualitative layer applied thematic analysis to open-ended user comments. Statistical testing used McNemar's test for accuracy and a paired bootstrap (10,000 resamples) for retrieval metrics; 95% confidence intervals are reported. Results: the proposed RAG system achieved 84% answer accuracy (95% CI 76–90%), Precision@5 = 0.80 and Recall@5 = 0.72, with a System Usability Scale (SUS) score of 78 and a Net Promoter Score (NPS) of +32 from 30 participants. Differences in accuracy versus the lexical and LLM-only baselines were statistically significant (McNemar p < 0.05). The system offers a replicable instantiation of RAG for transparent, citation-backed campus information services in Indonesian.

Cost-Effective Big Data Orchestration via n8n Workflow Automation for Digital Health Transformation in Resource-Constrained Hospitals

Wandi Purnama — 2026-07-05

Achieving compliance with Indonesia’s national SATUSEHAT health data mandate remains a complex hurdle for underfunded regional medical centers. The primary operational challenges stem from isolated departmental information architecture combined with the exorbitant licensing expenses of commercial middleware systems. To overcome these barriers, this study introduces a budget-friendly Big Data integration framework powered by n8n, an open-source, low-code workflow engine designed to dynamically unify disparate hospital environments. The methodology employs a Hadoop-based ecosystem and Apache Kafka for robust data ingestion, while n8n automates the Extract, Transform, Load (ETL) process to map raw clinical records into standardized HL7-FHIR JSON resources. Additionally, a lightweight Linear Regression model is applied as a low-compute operational optimization for dynamic batch-size prediction to prevent network overload during data transmission. Experimental results under a 72-hour continuous simulation on a single-core legacy server using 25,000 synthetic records demonstrate that the n8n-driven framework successfully sustains a throughput of 150 to 180 records per minute with a prediction error (RMSE) of 0.042. Furthermore, by eliminating proprietary software licensing fees and utilizing existing hardware, a comparative financial model indicates an estimated 85% reduction in the Total Cost of Ownership (TCO). Ultimately, this research provides a scalable technical blueprint for automating healthcare data integration, enabling under-resourced hospitals to achieve national interoperability mandates efficiently without compromising data integrity or financial stability.

Classification of Paddy as Visual Anomaly in Rice Piles Using MobileNetV2-Based Convolutional Neural Network

Barokah Saadah — 2026-07-05

Rice is a strategic food commodity whose quality is assessed based on the visual appearance of the grains, including the presence of unhusked rice as an undesirable element in piles of milled rice. Manual inspection is subjective, time-consuming, and prone to errors, necessitating a more objective automated approach. To address this issue, this study applies a MobileNetV2-based Convolutional Neural Network with transfer learning to classify unhusked grains as visual anomalies in rice piles into two classes: normal rice and anomalous grains. In terms of methodology, the dataset consists of 1,000 self-acquired images stratified into three groups with a 70:15:15 ratio. Image preprocessing was performed via background removal using the rembg library and random background simulation with five background color variations. Training was conducted in two phases: Phase 1 (transfer learning with a frozen base model) and Phase 2 (fine-tuning by opening the last 30 layers of the base model). The evaluation results on the test data showed an accuracy of 90.67%, a macro precision of 0.9213, a macro recall of 0.9067, and a macro F1-score of 0.9058. The false positive rate across all tests was 0. Phase 1 was selected as the best model because it produced more stable performance compared to Phase 2. Grad-CAM visualizations confirmed that the model focuses its attention on the visual features of the objects, not background patterns. These findings demonstrate that a combination of preprocessing, transfer learning, and data augmentation is effective for binary image classification when dealing with limited datasets.

A Decision Tree-Based Diet Recommendation System Based on Calorie and Macronutrient Constraints

Muhammad Farhansyah — 2026-07-05

Obesity and unhealthy eating patterns have become significant health concerns due to poor dietary habits and a lack of personalized nutritional guidance. Existing food recommendation systems often provide general recommendations without considering individual calorie and nutritional requirements. Therefore, this study aims to develop a web-based diet food recommendation system that integrates nutritional requirement calculations and Decision Tree-based food suitability classification. The system utilizes user information, including age, gender, weight, height, physical activity level, and diet goals, to calculate nutritional requirements through Body Mass Index (BMI), Basal Metabolic Rate (BMR) using the Mifflin-St Jeor method, and Total Daily Energy Expenditure (TDEE). A food dataset containing Indonesian foods and beverages was preprocessed and labeled using a rule-based approach based on macronutrient similarity scores. The Decision Tree algorithm was implemented to classify foods into suitable and unsuitable categories according to users’ nutritional requirements. Suitable foods were subsequently processed through a scoring mechanism and meal construction procedure to generate personalized meal plans. Experimental results showed that the Decision Tree model achieved an accuracy of 92.50%, precision of 78.26%, recall of 94.74%, and F1-score of 85.71%. System testing demonstrated that the developed features functioned properly and generated structured diet recommendations automatically. In conclusion, the proposed system can assist users in selecting foods according to their nutritional requirements and support healthier dietary planning.

A Measurement-Driven And Capacity-Aware Framework For 5G NR NSA Deployment

Afrizal Yuhanef — 2026-07-05

Conventional LTE-to-5G NR Non-Standalone (NSA) planning remains predominantly coverage-oriented and often fails to capture user-experienced throughput degradation under realistic traffic conditions. This limitation is critical in NSA architectures, where LTE acts as the anchor layer for control and mobility while 5G NR provides additional capacity. This study proposes a measurement-driven, throughput-centric spatial framework to identify LTE Capacity Bottleneck Zones (CBZ) as a basis for more realistic LTE-to-5G NR NSA deployment planning. The main novelty is the integration of a KPI-weighted RF Index with Kernel Density Estimation (KDE), DBSCAN spatial clustering, and fuzzy spatial zoning to generate throughput-aware capacity maps rather than purely coverage-based assessments. Drive-test measurements were conducted in Lubuk Alung District, Indonesia, under live LTE network conditions, yielding 8,355 radio KPI samples (RSRP, SINR) and 25,613 HTTP downlink throughput samples with geolocation. Statistical analysis using Pearson/Spearman correlation, polynomial regression, and Random Forest regression reveals consistently weak relationships between RSRP/SINR and throughput, indicating that radio-layer indicators alone provide limited explanatory power for user-experienced performance. The proposed framework classifies the study area into three spatial zones: LTE Stability Zone (28.99%), LTE Degradation Zone (63.02%), and LTE Capacity Bottleneck Zone (7.99%), where CBZs are characterized by acceptable radio conditions but localized throughput degradation. These findings enable a shift from coverage-centric evaluation toward targeted, throughput-aware capacity optimization for LTE-to-5G NR NSA deployment planning

TinyML and MFCC Feature Extraction for Energy Efficient Automatic Air Purifier Control

Mangasa Manullang — 2026-07-09

Classroom environments are highly susceptible to airborne disease transmission due to high occupant density and prolonged interaction times. Conventional mitigation strategies often rely on continuously operating air purification systems throughout building operational hours. This always-on approach guarantees continuous air circulation but results in massive and unnecessary electrical energy consumption, especially during idle periods or when biological contamination is absent. This research aims to design and implement an energy-efficient smart classroom system that automatically controls air purifiers based on real-time acoustic detection of sneeze events. The system utilizes Tiny Machine Learning embedded on an edge microcontroller with an onboard microphone. Audio datasets comprising sneeze, cough, and speech classes were processed using Mel-Frequency Cepstral Coefficients feature extraction at a 16 kHz sampling rate to optimize memory usage, followed by a neural network classifier training. The hardware prototype controls two air purifiers positioned for cross-ventilation, activating them for 15 minutes exclusively upon sneeze detection. The trained model achieved an overall accuracy of 97.5%, with a perfect precision rate in recognizing sneeze events. Field testing during an active class period demonstrated that the event-driven system consumed only 92.8 Watt-hours. Compared to the conventional continuous operation method, the automated system successfully reduced electrical power consumption by 71.4%. Implementing edge-based artificial intelligence for acoustic environmental monitoring provides a highly reliable approach to automated facility management, balancing health risk mitigation through optimal cross-ventilation with significant electrical energy conservation in smart classrooms. Future integration with low-power wireless modules is highly recommended to transmit event logs to a central dashboard, completing the sustainable facility management ecosystem.

Classifying Student Academic Achievement from Limited Categorical Institutional Records: A Comparative Study of Naive Bayes, K-Nearest Neighbor, and Decision Tree

Relita Buaton — 2026-07-05

Student academic achievement prediction is an important application in Educational Data Mining (EDM) that supports proactive academic decision-making. This study investigates a specific and underexplored condition in the literature the classification of student academic achievement when the only available predictors are categorical institutional background attributes without behavioral, attendance, or course-level data. This condition reflects data infrastructure limitations commonly found in Indonesian private higher education institutions. Three widely used classification algorithms Naive Bayes (BernoulliNB), K-Nearest Neighbor (KNN), and Decision Tree (CART) are compared against a majority class baseline through a five-stage preprocessing pipeline encompassing label normalization, cohort feature extraction, KNN k-value sensitivity analysis, and reporting of balanced accuracy and macro F1-score for fair evaluation under mild class imbalance. Results show that Decision Tree (depth=5) achieved the highest balanced accuracy (57.77%) and macro F1-score (57.51%), while Naive Bayes demonstrated the best generalization stability based on 10-fold cross-validation (60.07% ± 6.02%). All three models substantially outperformed the majority class baseline on balanced accuracy (+5–8 percentage points) and macro F1-score (+19–21 percentage points). Feature importance analysis identified IPS prior major background (15.6%) and the 2020 cohort (14.4%) as the most discriminative features. These findings provide evidence based algorithm selection guidance for data-constrained institutions and establish a reproducible performance benchmark for the categorical attributes only classification condition.

Adaptive Hybrid Model for Academic Performance Prediction and Learning Strategy Recommendation

Nenna Irsa Syahputri — 2026-07-10

Academic performance prediction is important for helping lecturers identify student learning needs before academic problems become difficult to address. However, students differ in learning preferences, engagement, prior knowledge, and academic achievement, making uniform learning strategies less effective. This study proposes an adaptive hybrid model for academic performance prediction and learning strategy recommendation by integrating K-Means Clustering, Simple Additive Weighting, and an Artificial Neural Network. The dataset consists of 74 student samples containing VARK learning preferences, engagement scores, pretest scores, and GPA-like academic indicators. After data cleaning, median imputation, and standard scaling, K-Means was applied to segment students into five learning profiles. Cluster centroids were then transformed into three decision criteria, namely Engage, Retention, and Effort. Simple Additive Weighting was used to rank three learning strategies: Micro-video Learning, Quiz Drill Practice, and Peer Discussion. The resulting recommendation labels were used together with the academic features to train an Artificial Neural Network for performance prediction and strategy classification. The evaluation showed that both models achieved an unrounded accuracy of 99.63%, while the rounded classification report displayed nearly perfect precision, recall, and F1-score. These findings indicate that the proposed integration can support data-driven adaptive learning decisions. Nevertheless, the high performance should be interpreted carefully because the dataset is limited and comes from a single institutional context. Further validation with larger, more diverse datasets is required to confirm generalizability.

A Comprehensive Analysis of Heap Sort Algorithm for Efficient Sorting Using C++ Programming Language

Rakhmat Purnomo — 2026-07-07

Sorting is a need in the computational system including in big data analysis, database management systems, and real time applications. Heap sort is an efficient, comparison-based sorting algorithm that visualizes an array as a binary tree and transforms it into a heap data structure (usually a max heap for ascending sort). The algorithm repeatedly takes the largest element from the root of the heap, swaps it with the last element, and thus reduces the heap size until the heap is sorted The algorithm repeatedly takes the largest element from the root of the heap, swaps it with the last element, and thus reduces the heap size until the heap is sorted. The steps of this algorithm: a. Create Max Heap: Convert the input array into a Max Heap. b. Sort: Swap the root element (the largest element) with the last element, decrease the heap size by 1, and then convert the new root element into a heap. c. Repeat step 2 until the heap is empty. C++ is a known programming language. In this journal we use C++ programming to sort unsorted array. The code is presented in the details. One thorough step by step simulation is given in real data with heap sort and the program is run. The analysis is given by 7 data, namely: [13, 10, 30, 2, 6, 7, 9]. The result is a presented with sorted heap sort. With 7 datasets to be analysed, it is concluded that 6 swaps happened.

Evaluating Kubernetes Progressive Delivery in Constrained Environments Flagger vs. Argo Rollouts

Gagah Syuja Saka Abdullah — 2026-07-09

Cloud‑native progressive delivery orchestrators reduce deployment risk by automating canary deployment rollback procedures, dramatically reducing mean time to recover from deployment failures. However, existing research predominantly evaluates these tools in hyperscale environments, masking the transient computational overhead they introduce in resource‑constrained edge deployments. This study empirically evaluates and compares the automated incident mitigation latency and computational resource volatility of Flagger and Argo Rollouts within a strictly resource‑constrained Kubernetes environment. A low virtual central processing unit Kubernetes testbed was provisioned using Talos Linux with strict hypervisor‑level central processing unit pinning, simulating edge computing conditions. Deterministic fault injection spanning four fault classes, two workload runtimes, and two network topology configurations was executed across thirty trials. A Shapiro-Wilk normality assessment, Welch t-test, Mann-Whitney U test, Cohen's d, and 95% confidence intervals were applied to compare temporal and computational metrics. Memory utilization remained statically bounded, averaging 24.01 megabytes for Flagger and 35.47 megabytes for Argo Rollouts. Under standard fault conditions, neither orchestrator demonstrated a consistent temporal advantage. However, under memory exhaustion progressing to CrashLoopBackOff, Argo Rollouts recovered in a mean of 29.67 seconds against Flagger's 166.79 seconds, a statistically significant 5.6-fold degradation with a large effect size. Argo Rollouts sustained transient central processing unit surges of 159 to 168 millicpu against Flagger's bounded ceiling of 17 to 18 millicpu. Progressive delivery automation introduces non‑negligible and fault-type-dependent computational overhead in resource‑constrained environments. Flagger is recommended for strict resource predictability in threshold-breach environments, while Argo Rollouts is recommended where broader fault-type resilience is operationally critical.

E-Book Design Analysis: The Effectiveness of Interactive Media in Understanding Characters and Ethnoscience Concepts

Dede Latipah — 2026-07-05

This research aims to develop a contextual-based interactive e-book integrating West Javanese folktales and ethnoscience concepts to enhance students’ understanding of natural phenomena in second-grade science at SDIT Abdurrahman Bin Arif, Curug, Bogor City. The study employed the ADDIE development model which includes Analysis, Design, Development, Implementation, and Evaluation stages. Data were obtained through expert validation, individual and small group testing, and effectiveness testing using pre-test and post-test instruments. Validation was conducted by media, language, and ethnoscience experts. The trial was performed on 29 second-grade students. The results showed excellent feasibility with expert assessments reaching over 90% in each category. The t-test analysis revealed a significant improvement in students’ learning outcomes after using the e-book (sig. 0.000 < 0.05). This indicates that the e-book was effective in increasing students' conceptual understanding of environmental changes through the integration of cultural narratives. The use of contextual e-books not only supports meaningful learning but also promotes character values such as responsibility and environmental awareness. Thus, the developed e-book is considered valid, practical, and effective for science learning in primary education

Explainable Hybrid XGBoost Fuzzy Logic Model for Accurate Anemia Risk Classification

Jepri Banjarnahor — 2026-07-05

Anemia remains a major global health concern that impairs oxygen transport and contributes to fatigue, cognitive decline, reduced productivity, and severe clinical complications. Although machine learning has shown promise for automated anemia detection, multiclass classification remains challenging due to class imbalance, overlapping hematological characteristics, and limited model interpretability. This study proposes an explainable hybrid framework integrating Extreme Gradient Boosting (XGBoost), SHapley Additive exPlanations (SHAP), and Fuzzy Logic to improve anemia risk classification and clinical decision support. The publicly available SKILICARSLAN dataset containing 15,300 anonymized patient records across five anemia-related classes was utilized. Seven hematological parameters, namely hemoglobin (HGB), mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (MCHC), red blood cell count (RBC), hematocrit (HCT), and ferritin, were employed as predictive features. The workflow comprised data auditing, stratified train–test splitting, Synthetic Minority Oversampling Technique (SMOTE), hyperparameter optimization, multiclass XGBoost modeling, SHAP-based explainability analysis, and fuzzy risk interpretation. Experimental results demonstrated 82.94% accuracy, 87.27% weighted precision, 82.94% weighted recall, and 84.78% weighted F1-score, with a mean cross-validation F1-score of 87.00%. The model further achieved a macro-average ROC–AUC of 0.81 and a weighted-average ROC–AUC of 0.90, indicating robust discriminative performance despite class imbalance. SHAP analysis identified HGB, ferritin, and RBC-related variables as the most influential predictors. Moreover, the fuzzy logic layer enhanced interpretability by translating model outputs into clinically meaningful risk levels. These findings demonstrate the potential of explainable hybrid intelligence for transparent and reliable anemia screening and decision-support applications.

Integration of Invisible Watermarking Based on a Hybrid DWT-SVD Approach in AI-Based Image Generators for Content Authentication

Herlina Harahap — 2026-07-05

The rapid advancement of artificial intelligence (AI) in the field of image generation has raised new challenges for digital content authentication and validity. AI-generated images are often indistinguishable from real photographs, creating potential risks of misuse in disinformation, visual manipulation, and copyright infringement. This study proposes the integration of an invisible watermarking method based on a hybrid of Discrete Wavelet Transform (DWT) and Singular Value Decomposition (SVD) directly into the AI-image generation pipeline. The system is developed end-to-end with three main components: a Translator API to support Indonesian text inputs, an AI-image generator to create images from descriptive text, and a watermarking module to embed and extract hidden watermarks automatically.

Experimental results confirm that the visual quality of watermarked images was preserved, with PSNR values consistently above 35 dB and SSIM ≥ 0.95, indicating that the watermark is imperceptible to human vision. Watermark extraction evaluation achieved a position accuracy of 59.43% after normalization and a subsequence accuracy of 80.20%, demonstrating reliable recognition of the embedded watermark sequence. Robustness tests under common manipulations such as JPEG compression, rotation, cropping, and noise addition showed that the watermark remained detectable, although accuracy decreased under extreme cropping. These findings demonstrate that the hybrid DWT–SVD method is effective for ensuring the authenticity of AI-generated content without compromising visual quality, while offering novelty through its integration into the generative pipeline and its support for local language inputs.

IndoBERT-Based Sentiment Analysis of Indonesian Social Media Discourse on AI-Generated Images

Halvino Iqbal Nataprawira — 2026-07-05

The rapid emergence of generative artificial intelligence has disrupted creative ecosystems, prompting widespread discourse across Indonesian social media. However, the exact sentiment structure of this public reaction remains empirically unmapped due to the contextual complexities of informal language. The objective of this research is to evaluate the efficacy of contextual language models by fine-tuning IndoBERT and benchmarking it against classical machine learning classifiers—including Complement Naive Bayes, Logistic Regression, and Support Vector Machine—for classifying social media sentiment. A multi-platform dataset comprising 2,981 Indonesian-language posts from X, Reddit, and YouTube was collected and manually annotated into positive, neutral, and negative classes. To address inherent class imbalance, Synthetic Minority Oversampling Technique was applied to classical models, while class-weighted loss and Masked Language Modeling augmentation were utilized for IndoBERT. Performance was evaluated using macro-averaged F1-score across five repeated stratified random splits. IndoBERT achieved a mean macro-F1 of 0.7131 ± 0.0180, outperforming the best classical baseline by approximately 0.12, demonstrating a pronounced advantage in resolving ambiguous neutral discourse. Negative sentiment heavily dominated the corpus at 61.8%, reflecting a prevailing critical stance toward AI-generated imagery concerning ethical and copyright issues. Furthermore, evaluation variance across random seeds exceeded variance from augmentation strategies, indicating test set composition is a major performance determinant. In conclusion, this study establishes a robust empirical baseline for Indonesian sentiment analysis, proving transformer architectures superior for nuanced public opinion mining.

A Multi-Model Time Series Framework for Forecasting Vietnam’s Tourism Revenue in the Post-COVID Recovery Era

Ho Nhat Hiep — 2026-07-05

This study forecasts Vietnam’s tourism revenue in the post-COVID-19 recovery period using a multi-model time series framework. The dataset includes three groups (ID 292–294) covering tourism business performance, economic sectors, and regional revenue. Six forecasting models are applied and evaluated using MAPE, MAD, and MSD. Results show that decomposition and Holt–Winters achieve the best accuracy (e.g., MAPE as low as 18%), while moving average performs well in specific cases (MAPE ≈ 28%). Forecasts indicate that tourism revenue may nearly double by 2030, driven mainly by domestic demand and the non-state sector, although international tourism recovers more slowly.

Herbal Leaf Identification for Balinese Lontar Usada Knowledge Preservation Using YOLOv8 Object Detection

I Nyoman Hary Kurniawan — 2026-07-05

Indonesia is a megabiodiversity country with more than 30,000 documented medicinal plant species. Much of this ethnobotanical knowledge is preserved in the Lontar Usada Bali, a traditional Balinese manuscript that records the medicinal uses of plants. However, preserving this knowledge is challenging due to the declining number of traditional practitioners and the difficulty of identifying medicinal plants in natural habitats. This study proposes a deep learning-based medicinal plant detection system using the YOLOv8 architecture to identify 12 classes of medicinal plant leaves in Taman Usada Bali. A total of 1,344 images containing 3,230 annotated leaf objects were collected under diverse lighting and background conditions. To improve model generalization, horizontal flipping, vertical flipping, rotation, Mosaic, and MixUp augmentations were applied. Five YOLOv8 variants (YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, and YOLOv8x) were evaluated using Precision, Recall, F1-score, mAP50, and mAP50–95 metrics. Experimental results showed that all models achieved high detection performance, with YOLOv8m obtaining the highest mAP50–95 score of 0.8676. However, Cost Benefit Analysis (CBA) using the Weighted Sum Model (WSM) identified YOLOv8n as the optimal model. Although YOLOv8m achieved the highest accuracy, YOLOv8n obtained the highest WSM score (2.6400) by balancing detection performance (mAP50–95 of 0.8464 ) and computational efficiency. With a 6 MB model size, 2.7 ms inference time, and 1.559 hours of training, YOLOv8n is suitable for real-time mobile and edge-computing applications. The novelty of this study lies in integrating Lontar Usada Bali taxonomy into a structured dataset, applying WSM for model selection, and enhancing detection robustness through Mosaic and MixUp augmentation.

Comparative Evaluation of Feature-Driven Development and Scrum in Public Facility Booking Systems

Wahdini Wahdini — 2026-07-05

The Parittiga Sub-District Office still relies on manual procedures for multipurpose building borrowing, resulting in limited schedule transparency, double booking vulnerabilities, and delayed confirmations. However, a critical research gap exists: no quantitative comparison between Feature-Driven Development (FDD) and Scrum has been conducted for public facility booking systems. To address this, a web-based booking system was developed using FDD and empirically compared with Scrum through a comparative experimental design. Using identical two-developer teams, both methods were evaluated across five metrics: development time per feature, functional error rate (black-box testing), developer understanding (1-5 scale), usability (SUS), and end-to-end borrowing efficiency. Findings demonstrate that FDD outperforms Scrum with 50% fewer functional errors (2 vs 4), higher developer understanding (4.3 vs 3.8), and better usability (84.2 - Excellent vs 71.6 - Good), despite requiring 2 additional hours per feature (12 vs 10). Moreover, the FDD-based system accelerates the complete borrowing process by 63% (4 vs 11 minutes). This research contributes empirical evidence to the agile methodology literature and offers a structured, replicable evaluation framework for practitioners selecting development methods for stable, documentation-intensive public service environments.

Comparative Scalability Analysis of Python Multiprocessing and OpenMP on Windows and WSL2

Achmad Fauzan — 2026-07-05

The advent of multicore processors has made efficient parallel programming models increasingly important for CPU-bound workloads. Process-based and thread-based parallelism in native Windows and WSL2 has not been comprehensively compared, despite extensive research on Python multiprocessing, OpenMP, and Windows Subsystem for Linux 2 (WSL2). This paper presents a unified multicore scalability assessment of Python multiprocessing and OpenMP, based on a CPU-intensive floating-point benchmark. Experiments were performed on an Intel Core i9-9900 processor with 8 physical cores and 16 logical threads in different process and thread configurations. Performance was evaluated in terms of execution time, speedup, parallel efficiency, CPU utilization, and statistical reliability metrics from ten repeated executions. Results demonstrate OpenMP performance advantage over Python multiprocessing for all tested configurations. OpenMP with 16 execution units provided maximum speedups of 6.482 on Windows and 7.065 on WSL2, compared to 5.624 and 5.790 for Python multiprocessing. The highest CPU utilization was achieved by OpenMP on WSL2 (97.31%). The reliability analysis confirmed experimental consistency, with coefficient-of-variation values below 10% for all the considered platforms. In general, WSL2 also had slightly better scalability and processor utilization than native Windows. The results show that WSL2 is a suitable environment for multicore computing and that thread-based parallelism provides better scalability for CPU-bound workloads. This study provides a comprehensive perspective on multicore scalability across different parallel programming models and execution environments by integrating multiple performance and reliability metrics into a single benchmark framework.

Lean Software Development for Reducing Bureaucratic Waste in Internship Administration

Nasir Zumali — 2026-07-05

Government internship administration in Indonesia remains heavily reliant on manual procedures, causing substantial operational waste. While prior studies have applied agile methods to digitize similar workflows, none have quantitatively measured bureaucratic waste reduction using inferential statistical validation. This study contributes a novel integration of Lean Software Development with Value Stream Mapping and inferential statistics to empirically validate waste elimination in public sector internship administration. A mixed-method explanatory sequential design was employed. First, Value Stream Mapping quantified lead time and cycle time on 30 paired internship applications. A web-based system was then developed using Lean Software Development principles. Functional validity was verified through Black Box testing with 50 test cases, and user satisfaction was assessed using the System Usability Scale on 30 respondents. The Shapiro-Wilk normality test confirmed non-normal data distribution, justifying the Wilcoxon Signed-Rank Test for hypothesis testing. Results demonstrated complete functional validity, a 97.5% reduction in lead time from 72.5 to 1.8 hours (Z=-4.782, p<0.001, r=0.87), a 90.9% reduction in cycle time from 25.4 to 2.3 minutes (Z=-4.801, p<0.001, r=0.88), and an excellent System Usability Scale score of 82.5. These findings provide empirical evidence that integrating Lean principles with robust statistical validation offers a replicable framework for quantitatively measuring bureaucratic efficiency gains, advancing beyond descriptive case studies in digital governance research.

Evaluation of Green Marketing Strategy Using Fuzzy AHP-Based Decision Support System

Marsono Marsono — 2026-07-05

Growing environmental awareness among consumers has encouraged companies to integrate ecological considerations into their marketing activities. However, many firms still find it difficult to determine which green marketing strategy should be prioritised, because the decision involves multiple conflicting criteria and a high degree of subjective human judgment. This study designs and applies a decision support system based on the fuzzy analytic hierarchy process (FAHP) to evaluate and rank green marketing strategy alternatives for a consumer goods company. The decision problem was structured into a three-level hierarchy consisting of the main goal, five evaluation criteria, namely green product, green price, green promotion, green distribution, and green corporate image, and four strategy alternatives. Expert judgments were gathered through pairwise comparison questionnaires using linguistic variables that were converted into triangular fuzzy numbers. Chang’s extent analysis was applied to compute the fuzzy synthetic extent and the degree of possibility, and the priority weights were normalised and verified through a consistency check (CR = 0.095, below the 0.10 threshold). To strengthen the validity of the recommendation, the alternative ranking obtained from FAHP was cross-validated against the Simple Additive Weighting (SAW) and TOPSIS methods. The results indicate that green product is the most important criterion (0.327), followed by green promotion (0.277) and green corporate image (0.210), while sustainable packaging is identified as the most preferred strategy alternative (0.294). All three methods produced an identical ranking, and a sensitivity analysis confirmed that the ranking remained stable under reasonable variations in the criteria weights. The proposed decision support system, whose architecture and interface are also presented, offers a transparent, consistent, and reproducible tool that helps managers allocate resources to the most effective green marketing strategy.

DTUF-PS: A Design Thinking-Based Usability Framework for Android Harvest Monitoring in Oil Palm Plantations

Novita Novita — 2026-07-05

One of Indonesia’s major plantation commodities is oil palm, but the still-manual documentation process hinders operational efficiency and real-time monitoring. Although the use of Design Thinking in mobile app development is growing, critical gaps remain, such as the lack of research that systematically integrates plantation-specific parameters—for example, the Harvest Density Index, Average Bunch Weight, and rotation cycle—into a usability evaluation framework tailored to the cognitive load of field workers. This study aims to develop and evaluate a prototype of an Android-based harvest monitoring system using a UI/UX approach and the Design Thinking method. The study involved 30 respondents, consisting of 5 division assistants and 25 field workers, and utilized the System Usability Scale (SUS), User Experience Questionnaire (UEQ), and time efficiency measurements. Empirical findings show an average SUS score of 80 (SD=6.1), categorized as “Good” (70th percentile), UEQ Perspicuity of 1.79 (85th percentile), “Very Good,” and a 69.9% increase in time efficiency from 600.6 to 180.7 seconds, paired t-test p<0.001, Cohen’s d=4.03, with instrument reliability confirmed by Cronbach’s α ranging from 0.73 to 0.87). The main contribution of this study is the Design Thinking-Based Usability Framework for Plantation Systems (DTUF-PS), which introduces three transferable constructs: Domain Artifact-Driven Empathize (DADE), Parameter-Embedded Ideation (PEI), and the Triadic Usability Evaluation Model (TUEM). This framework is proposed as an initial transferable model that requires further validation in other plantation contexts, such as rubber, cocoa, and coffee. The research results indicate that the Design Thinking method is effective in meeting user needs in agricultural monitoring.

Evaluation of YOLOv8n Performance for Real-time Human Detection on Autonomous Mobile Robots

Alif Daffa Dziqy Riyansah — 2026-07-09

This study presents the implementation and evaluation of the You Only Look Once version 8 nano (YOLOv8n) algorithm for real-time human detection on an autonomous mobile robot. The proposed system is designed as an edge-computing-based surveillance solution for monitoring restricted or difficult-to-access areas. The hardware platform integrates a Raspberry Pi 4B for visual processing and an Arduino Mega 2560 for navigation control through serial communication. Human detection is performed using a night-vision camera, while obstacle avoidance is supported by three ultrasonic sensors. A custom dataset was collected under various human postures, object distances ranging from 1 to 10 meters, and different lighting conditions. The YOLOv8n model was trained using 300 epochs with an image resolution of 640 × 640 pixels. Experimental results demonstrate that the proposed system achieves reliable real-time performance under varying environmental conditions. Under lighting variation tests, the model achieved 100% precision, 93.5% recall, 96.6% F1-score, and 93.55% accuracy with an average processing speed of 24.30 frames per second. Distance-based testing produced 100% precision, 92.42% recall, 96.06% F1-score, and 92.42% accuracy at 23.2 frames per second. Furthermore, autonomous navigation experiments confirmed that the robot was capable of simultaneously detecting humans and avoiding obstacles with response times ranging from 2.4 to 3.2 seconds. These findings indicate that You Only Look Once version 8 nano (YOLOv8n) provides an effective balance between detection accuracy, processing speed, and computational efficiency, making it suitable for deployment on edge-computing-based autonomous mobile robots.

Multi-Metric Evaluation of Machine Learning Algorithms for Diabetes Prediction Using Feature Importance and ROC Analysis

Fendi Setiawan — 2026-07-05

Diabetes mellitus has become a major global health threat, and many undiagnosed cases remain undetected due to some limitations of the conventional diagnostic methods. Despite the promising results of machine learning (ML) for early diabetes diagnosis, the majority of the current research assessing algorithms either uses insufficient metrics or does not follow a consistent assessment approach. This paper addresses that gap by utilising an integrated evaluation framework. The framework includes feature importance analysis, Pearson correlation assessment, confusion matrix decomposition, and ROC-AUC comparison. It applies this framework to the Pima Indians Diabetes Dataset (mde) and four popular ML classification algorithms: Naive Bayes, Decision Tree, Random Forest, and Logistic Regression. The most significant predictors, according to our feature analysis, were glucose (27.6%), body mass index (16.0%), age (12.7%), and diabetes pedigree function (12.7%). Among the classifiers, Random Forest exhibited the greatest accuracy (76.0%) and precision (68.1%), Naive Bayes the best recall (64.8%), and Logistic Regression the highest AUC-ROC (82.3%). For patients at high risk, the models' virtual projections across all three risk profiles were in agreement. Model selection should be determined by the unique clinical screening aim, since these findings suggest that there is no one better universal method. Random Forest and Logistic Regression are the most promising for assisting in preliminary diabetes prediction, although further validation on diversity datasets is needed prior to clinical deployment.

Amplitude-Based Statistical Filtering Method for Resonance Localization in Low-Cost Acoustic Sensing Systems

Feri Iskandar — 2026-07-05

Acoustic resonance experiments using closed organ pipes are widely used to study the relationship between sound frequency, wavelength, and air-column length. However, low-cost sensor-based systems often produce unstable acoustic signals contaminated by environmental noise, amplitude fluctuations, and sensor-response variability, making resonance identification difficult. This study aims to develop and evaluate an amplitude-based statistical filtering method to improve signal stability and determine the fundamental resonance position more accurately. The proposed method was evaluated using a closed organ pipe experiment integrated with an acoustic sensor and microcontroller-based data acquisition system. Acoustic signals were processed using amplitude-based statistical filtering to extract dominant resonance responses and improve resonance localization. Statistical evaluation was conducted to analyze signal stability and measurement accuracy. The results showed that the filtering process reduced the standard deviation from 9.24 cm in the raw dataset to 6.72 cm in the final resonance candidates, indicating improved resonance localization stability. The experimental resonance length obtained after filtering was 16.12 cm, while the theoretical resonance length was 16.75 cm, resulting in a relative error of 3.76%. These findings demonstrate that the proposed filtering method can improve resonance detection accuracy using a simple, practical, and computationally efficient approach suitable for low-cost educational laboratory systems.

A WebGIS-Based Location Analysis System for Disaster Mitigation in Bitung City Using Ray Casting and Haversine Formula

Romeo Fernando Mikhael Dendeng — 2026-07-05

Indonesia is highly vulnerable to multi-hazard natural disasters, and Kota Bitung specifically faces high risks due to its geographical contours. Currently, the dissemination of disaster information by the Regional Disaster Management Agency (BPBD) relies heavily on conventional static maps. Most existing disaster WebGIS platforms focus merely on static visualization and lack an integrated system capable of instantly analyzing a user's coordinate status against multi-hazard spatial polygons while simultaneously providing location-based evacuation routing. To address this gap, this research aims to design and develop a responsive WebGIS that allows users to independently detect their risk status and logically find the nearest evacuation route. The system development utilizes the Rapid Application Development (RAD) method. The core engine integrates the Ray Casting algorithm to solve the Point-in-Polygon problem against disaster zone boundaries, and the Haversine Formula to calculate the nearest available evacuation point. Based on comprehensive evaluations, including accuracy testing, spatial distance validation, and 11 distinct black-box testing scenarios, the system successfully processed GPS-based coordinate inputs, handled polygon boundary edge-cases, and generated evacuation routes using the OpenSource Routing Machine (OSRM). Ultimately, the proposed system provides a functional prototype for location-based disaster risk analysis and evacuation point recommendation, serving as a foundational interactive instrument to support emergency preparedness in Kota Bitung.

Deep Learning-Based Classification of Cikadu Batik Motifs Using ResNet50 and MobileNetV2

Rizki Ripai — 2026-07-07

Batik motif recognition is essential for cultural heritage preservation and the digitization of traditional Indonesian textile knowledge. This study proposes a deep learning-based framework for the automatic classification of Cikadu Batik motifs from Tanjung Lesung, Banten — a regionally distinct batik pattern that has not been systematically studied in prior computational literature. Two convolutional neural network (CNN) architectures were implemented and comparatively evaluated under identical experimental conditions: ResNet50, a high-capacity model employing residual skip connections, and MobileNetV2, a lightweight model utilizing depthwise separable convolutions and inverted residual blocks. A curated dataset of 2,500 images spanning five motif classes was constructed through collaboration with local batik artisans, preprocessed via resizing (224×224), pixel normalization, and augmentation (rotation, zoom, horizontal flip, brightness adjustment), and partitioned using a stratified 70:15:15 split. Both models were trained with transfer learning from ImageNet weights, using the Adam optimizer (lr=0.0001), categorical cross-entropy loss, batch size of 32, and early stopping over 30 epochs. Model evaluation employed accuracy, precision, recall, F1-score, AUC-ROC, inference time, and parameter count. ResNet50 achieved 95.51% accuracy, 95.67% precision, 95.34% recall, 95.50% F1-score, and 99.56% AUC-ROC, with an inference time of 18.2 ms and 25.64 million parameters. MobileNetV2 achieved 92.13% accuracy, 92.28% precision, 91.98% recall, 92.13% F1-score, and 98.89% AUC-ROC, with an inference time of 8.7 ms and only 3.54 million parameters — approximately 7× lighter and 2× faster. These results empirically establish a clear accuracy-efficiency trade-off, with ResNet50 favored for accuracy-critical server-based systems and MobileNetV2 better suited for real-time mobile deployment. This study constitutes the first published benchmark for deep learning-based Cikadu Batik classification and provides a principled basis for architecture selection in regional batik recognition applications

Image-Based Food Classification for Nutritional Information Estimation Using Deep Learning

Sahrial Ihsani Ishak — 2026-07-05

This study aims to develop an image-based food classification application integrated with nutritional information retrieval using a deep learning approach. The proposed system is designed to recognize food types from images and provide nutritional information based on an Indonesian food nutrition database. The method involves collecting a dataset of 8,248 images representing 38 categories of Indonesian traditional foods, performing image preprocessing and data augmentation, and developing a Convolutional Neural Network (CNN) model based on the MobileNetV2 architecture through transfer learning. Model performance was evaluated using a 3-fold stratified cross-validation strategy and measured using accuracy, precision, recall, and F1-score metrics. Experimental results showed that the proposed model achieved average accuracy, precision, recall, and F1-score values of 98.85%, 98.88%, 98.85%, and 98.85%, respectively, demonstrating robust and consistent classification performance across the validation folds. The trained model was subsequently deployed into a mobile application using TensorFlow Lite to support real-time food classification and nutritional information presentation. The main contribution of this study is the development of an end-to-end mobile system that integrates deep learning-based food classification with an Indonesian food nutrition database, enabling users to obtain calorie, protein, fat, and carbohydrate information quickly and conveniently for dietary monitoring and health awareness.

SC-Literature Intelligence: A Retrieval-Augmented Generation Framework for Multi-Category AI Literature Synthesis in Supply Chain

Setio Basuki — 2026-07-06

This paper develops SC-Literature Intelligence, a retrieval-augmented generation (RAG) framework for research synthesis of scientific literature on artificial intelligence (AI) in the supply chain domain. The study addresses the fragmentation of scientific findings, which makes cross-document understanding difficult by supporting four categories of literature-analysis queries: trend analysis, gap detection, comparative synthesis, and evidence-based question answering (QA). The primary novelty lies in introducing a category-aware research synthesis framework capable of evaluating RAG performance across multiple literature-analysis tasks rather than conventional question answering. The framework is built from Scopus-indexed abstracts through pre-processing, chunk-based embedding using BGE-M3 and LaBSE, vector storage, semantic retrieval, and prompt-guided generation evaluated using the RAGAS framework across 640 experimental runs. The results show that BGE-M3 consistently outperforms LaBSE on all RAGAS indicators with the best configuration (chunk size 64, Top-K 5) achieving scores between 0.722 and 0.856 across faithfulness, answer relevancy, context precision, and context recall. Gap detection emerges as the best-supported query category, whereas comparative synthesis remains the most challenging. Failure analysis further reveals that retrieval-stage issues dominate over generation-stage issues, identifying embedding quality as the primary bottleneck. These findings demonstrate that category-aware RAG-based synthesis can support structured, evidence-grounded literature analysis in the supply chain AI domain.

Two-Stage Framework Using IndoBERT for Sentiment Analysis of Tokopedia Reviews under Extreme Class Imbalance

Ades Tikaningsih — 2026-07-05

The rapid growth of the Indonesian e-commerce industry has generated a large volume of customer reviews for sentiment analysis, but the data distribution often suffers from extreme class imbalance. The review dataset exhibits a 97.6% dominance of the positive class, causing the single-stage transformer model to produce high accuracy that does not fully represent classification capability. The baseline model achieves a macro-averaged F1-score of 0.599, with a neutral-class recall of 26.3%. Approaches based on loss function adjustment, such as class-balanced loss, focal loss, weighted cross-entropy, and decision-threshold adjustment, are unable to fundamentally address this issue, yielding only limited performance improvements. This study proposes a two-stage classification approach that decomposes the multi-class classification task into two sequential binary classification stages using a BERT-based Indonesian-language transformer model (IndoBERT). The first stage separates the positive class from the non-positive class, while the second stage distinguishes between the neutral and negative classes in a more balanced decision space. The proposed approach achieves a macro-averaged F1-score of 0.761, representing a 16.2% improvement over the baseline and outperforming all loss-function-based methods. These findings suggest that, under conditions of extreme class imbalance, simplifying the decision space through gradual task decomposition is more effective than intervention at the loss-function level. Furthermore, error propagation analysis and qualitative evaluations demonstrate that this approach improves sensitivity to minority classes, although challenges remain in cases involving ambiguous expressions.

Semantic Embedding and Profile-Based Ranking for Automated Reviewer Recommendation

Azisya Luthfi Bintang — 2026-07-05

Manual reviewer assignment in peer review is difficult to scale because submission volumes grow faster than editors can inspect reviewer expertise, and reviewer profiles shift across topics and time. Existing automated approaches often rely on keyword or lexical matching, which cannot capture semantic similarity, and few combine dense retrieval with interpretable reviewer evidence. This study develops and evaluates an explainable reviewer recommendation system using a BERT-first Reciprocal Rank Fusion semantic-profile backend. The system retrieves candidate evidence using BERT and SPECTER2 semantic representations, extracts candidate reviewers from retrieved paper authors, and ranks them using fused retrieval evidence supported by frequency, h-index, and recency signals. The expertise-scoring component was evaluated using the Stelmakh/OpenReview benchmark, while end-to-end recommendation was evaluated on an OpenAlex citation-based proxy dataset using a validation split for configuration selection and a held-out test split for final reporting. SPECTER2 max pooling achieved a weighted Kendall tau loss of 0.22 on the Stelmakh/OpenReview benchmark, consistent with the public SPECTER2 baseline. On the held-out test split, the selected BERT-first RRF semantic-profile backend achieved the highest NDCG@10 of 0.2621, significantly outperforming BERT, SPECTER2-only, BM25, TF-IDF, and the previous profile-heavy backend. These findings indicate that rank-level fusion of complementary dense retrieval signals can improve reviewer candidate ranking while retaining interpretable profile evidence for editorial workflows. The local evaluation uses citation-based proxy relevance rather than true editorial assignments, so further validation using human-annotated reviewer data is needed.

Multi-Device IoT Integration Using an API-Based Modular Architecture for Environmental Monitoring Systems

Andi Marwan Elhanafi — 2026-07-05

The Internet of Things (IoT) has increasingly played a significant role in the development of adaptive and real-time environmental monitoring systems. However, integrating multiple IoT devices remains challenging due to variations in data transmission intervals, communication protocols, and processing capabilities across devices. These differences often complicate system interoperability and data management within a unified monitoring platform. To address this issue, this study proposes an API-based modular architecture as a solution for integrating heterogeneous IoT devices in environmental monitoring systems. The proposed architecture separates core system functions into independent modules, including data acquisition, device management, and data visualization. The proposed architecture is evaluated through a multi-device environmental monitoring implementation configured with different logging intervals in order to assess communication performance and data consistency. The novelty of this study lies in its architectural approach to handling heterogeneous data transmission intervals in multi-device IoT environments using a modular API-based design. The experimental results indicate that the average communication latency is approximately 200ms, while the average daily data logging volume exceeds 3,500 entries per device. Furthermore, analysis of logging interval variations shows a time deviation of less than 3 seconds, which remains within the acceptable range for real-time environmental monitoring applications. The results demonstrate that the proposed architecture achieves success rate of over 97%, confirming the reliability of the proposed API-based modular architecture. Overall, the findings suggest that the modular API-driven architecture not only improves the flexibility and scalability of multi-device IoT integration but also maintains reliable data consistency and efficient communication performance.

Application of Sentence-BERT Embeddings for Semantic Deduplication of Industrial Material Records

Seno Hardijanto Purnomo — 2026-07-05

Industrial material master data in Enterprise Resource Planning (ERP) and Enterprise Asset Management (EAM) systems accumulates duplicate records that distort inventory, procurement, and analytics. Traditional deduplication relies on string-similarity measures such as Levenshtein, Jaro–Winkler, and TF-IDF cosine, which can struggle on catalogs mixing Indonesian and English terminology—e.g. Valve versus Keran—and on paraphrastic variants with different word order or abbreviation style. This study formally specifies a semantic deduplication pipeline that encodes material descriptions as sentence embeddings using Sentence-BERT (SBERT) and compares them via cosine similarity, then diagnostically evaluates the extent to which SBERT improves over those baselines. Following Design Science Research, the pipeline specifies normalisation, encoding with a multilingual paraphrase-tuned SBERT variant, and pairwise comparison within candidate sets produced by hybrid blocking; the diagnostic evaluation reports scores on the raw descriptions to expose baseline behaviour before domain-specific harmonisation. A sample of 291,000 records from two Indonesian industrial power plants motivates the design. On a diagnostic set of 100 record pairs derived from existing engineer-annotated duplicate markers, Jaro–Winkler achieves F₁ = 0.925 (precision 1.000, recall 0.860) and SBERT achieves F₁ = 0.875 (precision 0.913, recall 0.840) at threshold τ = 0.65; qualitative analysis of twelve representative pairs further reveals that SBERT excels on structural paraphrase (cosine 0.73–0.88 where character-level methods score below 0.50), while Jaro–Winkler remains competitive on abbreviation, unit-standard, and cross-language pairs—particularly those involving Indonesian technical vocabulary under-represented in the model’s training distribution. The central finding is that Sentence-BERT complements rather than replaces string baselines, which motivates future work on multi-channel architectures combining textual semantics with structural context.

Implementation of CMM Method to Measure Maturity Level of SPBE in Palembang City Government

Nyimas Hamidah Purnama Agustriani — 2026-07-05

The Palembang City Government faces challenges in synchronizing internal policies with the operational implementation of the Electronic-Based Government System (SPBE). This research aims to measure the maturity level of the Internal Policy Domain and the SPBE Governance Domain in the Palembang City Government using the Capability Maturity Model (CMM) method. Additionally, the study aims to identify gaps through gap analysis and formulate strategic recommendations to achieve the "Optimum" maturity level. The research focuses on evaluating 20 indicators covering the Internal Policy and SPBE Governance domains within the Palembang City Government. The maturity level analysis refers to the framework of Permenpan-RB No. 59 of 2020 and Menpan-RB Guideline No. 3 of 2024. The indicator assessment results show that the Internal Policy Domain reached an index of 4.1 (Very Good), while the Governance Domain obtained an index of 3.4 (Good). Overall, the Palembang City Government possesses a very strong regulatory foundation; however, governance effectiveness remains sectoral and is not yet fully aligned with technical implementation across all regional apparatus. It is recommended that the Palembang City Government strengthen cross-sector coordination through integrated SPBE budget synchronization and conduct periodic policy reviews accompanied by formal documentation to ensure the sustainable improvement of SPBE quality.

Graph-Based Hybrid GNN-Transformer for Imbalanced Credit Card Fraud Detection

Muhammad Bayu Wijaya Putra — 2026-07-05

Credit card fraud detection faces two major challenges: severe class imbalance and the limited ability of conventional feature-based models to capture relational patterns among transactions. This study proposes a graph-based Hybrid GNN-Transformer architecture for imbalanced credit card fraud detection by integrating transaction-level relational learning through k-nearest neighbor graph construction and feature-interaction learning through multi-head self-attention. The novelty of this study lies in combining graph-based transaction modeling and Transformer-based feature interaction within a unified architecture. Using the selected graph configuration and validation-based threshold tuning, the proposed model achieved 79.71% precision, 74.32% recall, 76.92% F1-score, 96.06% ROC-AUC, and 68.65% PR-AUC. Compared with Logistic Regression, Random Forest, and Gradient Boosting baselines, the hybrid model showed competitive fraud detection sensitivity, although the baseline classifiers still achieved stronger overall F1-score and PR-AUC. Ablation results show that the hybrid architecture improves minority-class detection compared with single-branch variants by combining relational transaction information from the GNN branch and feature-interaction information from the Transformer branch. These findings indicate that graph-based hybrid representation learning is a promising direction for imbalanced fraud detection, while further optimization is still required to improve precision-recall balance and competitiveness against strong feature-based baselines.

CLASSIFICATION OF COFFE FRUIT DRYING USING VGG16

Annisa Diyan Novitasari — 2026-07-05

The drying process is a crucial stage in coffee post-harvest handling that directly affects the final product quality, especially in the specialty coffee segment. Assessment of the coffee fruit drying level in the field is still largely carried out visually and subjectively, which can potentially lead to inconsistent quality. This study aims to develop an automatic classification system for coffee fruit drying levels based on digital images using a deep learning method with the Convolutional Neural Network (CNN) VGG16 architecture. The dataset used consists of 561 coffee fruit images classified into three classes: Wet, Medium, and Dry. The preprocesssing stages include background removal, auto-cropping, and image standardization. Two models were developed: a baseline model without data augmentation and a model with data augmentation and selective fine-tuning on the final layers of VGG16. The evaluation results show that the baseline model achieved a validation accuracy of 83%, while the model with augmentation and fine-tuning improved the accuracy to 94%, accompanied by significant increases in precision, recall, and F1-score values. The proposed model also demonstrates a high and stable level of prediction confidence. These results prove that the VGG16 approach is effective for classifying coffee fruit drying levels and has the potential to be applied as an objective post-harvest quality control support system.

Comparison of Chronos and Conventional Models: Predicting Machine Downtime using Time Series

Hendri Hendri — 2026-07-05

This study analyzes the comparison between a pretrained transformer model (Chronos) and conventional models in predicting industrial machine downtime using time series data to achieve greater accuracy and efficiency for companies. More specifically, this research focuses on early detection before downtime occurs to reduce company losses in terms of both costs and product quality, and to ensure that Key Performance Indicator targets are met. Design/methodology/approach: The research methodology includes primary data collection, data preprocessing, and sequential data splitting (80% training, 10% validation, 10% testing) to prevent potential data leakage. Model evaluation is measured using the Mean Absolute Error loss function, focusing on the “handling machine” category, which yields 4,069 to 4,101 data rows after the preprocessing stage. Research showed that the conventional XGBoost model with tuning performed best, with the lowest Mean Absolute Error among the other models. XGBoost proved to be highly effective and was capable of outperforming advanced transformer-based models (such as Chronos), particularly when applied to a limited dataset of 4,069 data points. Conversely, transformer architectures like Chronos performed poorly on small datasets because they were designed for massive datasets. This study focuses on the application and evaluation of modern artificial intelligence technologies, specifically transformer architectures such as the Chronos model. Although previous similar studies have successfully predicted downtime accurately using conventional models (such as ARIMA, Random Forest, Support Vector Machine, and autoencoders), those earlier studies have not tested the effectiveness of transformer architectures in detecting machine downtime.

A Comparative MCDM Framework Integrating AHP, SAW and TOPSIS for Robust Public Sector Selection

velisia kartika — 2026-07-05

The selection of top employees at DINPMP2KUKM in Bangka Regency faces challenges including subjective evaluation, a lack of clear criteria, and difficulties in assessing employee performance across different fields, which may lead to perceptions of unfairness and decreased work motivation. This study aims to apply a comparative multi-criteria decision-making framework combining the Analytic Hierarchy Process (AHP), Simple Additive Weighting (SAW), and the Technique of Order Preference by Similarity to Ideal Solution (TOPSIS). A mixed-methods approach was employed, involving five respondents (four division heads and one human resources sub-division head) and three employee alternatives. Data were collected through structured interviews and paired comparison questionnaires. AHP was used to determine criterion weights, while SAW and TOPSIS were applied to rank alternatives, followed by sensitivity analysis to test ranking robustness. Unlike prior studies that generally used only a single MCDM method without robustness testing, this study validates ranking consistency across three different MCDM methods with sensitivity-based robustness testing. The results indicate that cooperation (29.2%) is the dominant criterion, followed by performance (22.6%), discipline and innovation (16.8% each), and integrity (14.6%), with a consistency ratio of 0.05 indicating consistent evaluations. All three methods produced identical rankings, with Employee 3 selected as the top employee (43%), excelling in four out of five criteria. Sensitivity analysis confirmed that Employee 3 remained at the top when the weight of cooperation was altered by ±20%, demonstrating decision robustness. Given the limitations in the number of respondents and alternatives, these findings suggest that the combination of AHP, SAW, and TOPSIS with sensitivity analysis can produce more consistent employee selection outcomes compared to single-method approaches. Further research is recommended to develop a web-based decision support system with a larger number of respondents and alternatives.

An End-to-End Balinese Lontar OCR Framework Using Bayesian-Optimized Multiscale Retinex and MobileNetV3

Putu Ayu Febyanti — 2026-07-05

The preservation of Balinese lontar manuscripts has become increasingly important due to their cultural, historical, and religious significance, while physical degradation such as uneven illumination, faded ink, texture interference, and manuscript aging continues to reduce readability and complicate digital preservation efforts. This study proposes an end-to-end Optical Character Recognition (OCR) framework for degraded Balinese lontar manuscripts by integrating Bayesian-optimized image enhancement, adaptive preprocessing, morphology-based segmentation, domain-specific augmentation, and lightweight deep learning recognition using MobileNetV3. The proposed enhancement pipeline combines Multiscale Retinex, adaptive gamma correction, edge-preserving filtering, and hybrid binarization to improve character visibility under degraded manuscript conditions. Bayesian Optimization with Optuna and Tree-structured Parzen Estimator (TPE) was employed to automatically optimize enhancement parameters according to manuscript quality characteristics. Experimental results demonstrated substantial improvements in manuscript image quality, where Laplacian Variance increased from 306.7596 to 6685.7641, RMS Contrast improved from 28.976 to 83.9085, Michelson Contrast increased from 0.8238 to 1.0, and Ink Ratio Score improved from 0.6096 to 0.9847. The MobileNetV3-based OCR recognition model achieved a test accuracy of 80.52% and a best validation accuracy of 83.78% across 102 Balinese script classes. The proposed framework demonstrates that adaptive enhancement optimization combined with lightweight OCR recognition can provide robust and computationally efficient recognition performance for degraded historical manuscripts while supporting scalable digital preservation and mobile-oriented cultural heritage applications.