Sinkron : jurnal dan penelitian teknik informatika

Fraud Detection in E-Commerce Transactions Using Autoencoder Anomaly Scoring and XGBoost Classification

2026-07-09T12:56:46+00:00

The rapid expansion of e-commerce platforms has intensified the risk of digital transaction fraud, which is particularly challenging to detect due to highly imbalanced datasets where fraudulent transactions represent a small minority. This study proposes a two-stage hybrid fraud detection model that integrates an Autoencoder for unsupervised anomaly detection with XGBoost as a supervised classifier. The objective is to evaluate whether incorporating reconstruction error scores from the Autoencoder as an additional feature improves XGBoost classification performance on highly imbalanced e-commerce fraud data. The dataset used is the Credit Card Fraud Detection dataset from Kaggle (ULB), consisting of 284,807 transactions with a fraud ratio of 0.17%. The research pipeline includes stratified train-validation-test splitting, StandardScaler normalization, Autoencoder training exclusively on non-fraud data to produce anomaly scores (AE_Score), and XGBoost training with scale_pos_weight to handle class imbalance. Threshold optimization was performed using the precision-recall curve on the validation set. Results demonstrate that the two-stage model achieved a ROC-AUC of 0.9749, PR-AUC of 0.8442, precision of 0.8971, recall of 0.8243, and F1-score of 0.8592 on the test set. The AE_Score showed strong discriminative power, with a fraud mean of 4.79 compared to 0.02 for non-fraud transactions. These findings confirm that integrating Autoencoder-based anomaly scoring into gradient boosting classification effectively addresses data imbalance and improves fraud detection performance in large-scale e-commerce environments.

Multilingual-E5 Size Comparison on Indonesian Public Nutrition Regulation: A Single-Document Case Study

2026-09-15T17:24:21+00:00

Rigid structure and complex legalese frequently make it difficult to access Indonesian legal documents, such as Ministry of Health rules, and traditional Large Language Models are still prone to hallucinations. This problem is mitigated by RAG, but its accuracy is highly dependent on the semantic quality of the dense retriever, an aspect that is rarely investigated for Indonesian regulatory text, where previous studies either tested different model families on non-Indonesian corpora or used generic embeddings without efficiency trade-off analysis. In order to ascertain whether model size influences retrieval accuracy and answer quality, this study empirically benchmarks two variants of the same embedding family, multilingual-e5-large and multilingual-e5-base, within a RAG pipeline for question answering on Indonesian public nutrition law across three retrieval depths (k = 3, 5, 7). In addition to Context Relevance, Faithfulness, and Answer Relevance assessed by an LLM, 22 test questions were assessed across 132 combinations utilizing Precision, Recall, and F1-Score against ground truth. Wilcoxon signed-rank, Friedman, post-hoc, and non-inferiority tests were then used for analysis. Multilingual-e5-large and multilingual-e5-base did not differ significantly on any of the six metrics (p>0.05), although k had a significant impact on Precision, Recall, and F1-Score (p<0.05) but not the LLM-judge measures. The smaller multilingual-e5-base was non-inferior to multilingual-e5-large within a margin of 0.05 on all metrics except Recall and Faithfulness, and k can follow precision or recall priorities. Because the study covers a single regulation and computational cost was not measured, the findings should not be generalized to Indonesian legal corpora.

Fairness and QoS Comparison of Equal, Max-Min, and Demand-Proportional Allocation in 5G Network Slicing

2026-07-24T18:54:15+00:00

Network slicing is one of the main mechanisms used in 5G systems to let several service types operate over the same physical infrastructure. In this work, the shared system includes eMBB, URLLC, mMTC, and Best Effort slices, all competing for limited bandwidth and CPU capacity. Because the resource-allocation rule can change both fairness and service quality, this paper studies three policies under the same traffic conditions: Jain-based egalitarian allocation (JF), Max-Min Fairness (MMF), and Proportional Fairness (PF). A Python simulator is used to apply each policy separately to identical demand samples, after which the resulting allocations are evaluated using Jain’s Fairness Index (JFI) and the Gini coefficient. The traffic model follows a Poisson process with low, medium, and high load levels, and each case is repeated for 100 independent Monte-Carlo runs. In addition to fairness, the evaluation reports throughput, packet loss ratio, jitter, delay, bandwidth and CPU utilization, SLA satisfaction, and one-way ANOVA tests. The results show that JF gives perfect equality at all loads (JFI = 1.0, Gini = 0), although this equality reduces performance when the offered load is high. MMF is the strongest demand-aware fairness policy and provides the most balanced behavior under medium and high load. PF gives the lowest high-load jitter, but it also records the weakest SLA satisfaction (0.12). The ANOVA results indicate statistically significant differences (p < 0.05) among the policies for most metrics in the medium- and high-load scenarios.

Comment Intention Classification in TikTok Fashion Affiliate Marketing Based on Attention-Weighted Word2Vec Representation

2026-09-17T07:55:22+00:00

TikTok Fashion Affiliate generates short, informal user comments with a variety of communicative intentions, necessitating a text representation capable of preserving key information within the comments. This study aims to evaluate Norm-Based Softmax Weighting as a non-learned weighting mechanism and compare it with Mean Pooling for classifying the intent of Indonesian-language TikTok Fashion Affiliate comments. The comment dataset was independently annotated by three annotators with a Krippendorff’s Alpha of 0.7950 and classified into three classes: Question, Request, and Opinion. Word2Vec Skip-gram with a dimension of 100 was used to generate comment representations, which were then compared using Mean Pooling and Norm-Based Softmax Weighting in Support Vector Machine (SVM) and Logistic Regression models. Hyperparameters were selected via Grid Search using Macro F1 as the optimization criterion. To prevent data leakage, Word2Vec was trained only on the training data and retrained on each fold. Evaluation was performed using the test data, 5-Fold Stratified Cross-Validation, and Leave-One-Video-Out. The results showed that SVM with Norm-Based Softmax Weighting achieved a Macro F1 score of 0.7995, slightly higher than Mean Pooling’s 0.7913. In contrast, for Logistic Regression, Mean Pooling achieved a Macro F1 score of 0.7846, which was higher than the 0.7768 achieved by Norm-Based Softmax Weighting. However, ablation analysis and statistical testing did not reveal any significant differences. Leave-One-Video-Out validation revealed performance variations across videos and a decline in performance compared to random data splitting. Thus, Norm-Based Softmax Weighting is a competitive non-learned weighting alternative, but its effectiveness depends on the classifier and data characteristics.

Reinforcement Learning Evaluation for Dependent Task Offloading in Mobile Edge Computing Systems

2026-07-29T19:56:38+00:00

Dependent-task directed acyclic graphs require each task to execute locally or offload to a mobile edge computing server, coupling latency, communication cost, and device energy. This study reports a controlled evaluation of deep reinforcement learning for offloading under a common simulator, dataset, and train/validation/final-test protocol. Baselines comprise a recurrent policy with proximal policy optimization (PPO-LSTM) and a double deep Q-network; extensions comprise a Transformer policy, discrete soft actor-critic, and stabilized PPO with advantage normalization, scheduled optimization, divergence-based early stopping, and orthogonal initialization. Agents and five heuristics are evaluated on final-test graphs of ten to fifty tasks under latency-oriented and energy-aware quality of experience (QoE) objectives, with five seeds and Holm-corrected tests. Stabilized PPO attained mean QoE of 0.31–0.43 across task-size and bandwidth analyses, though not at every size or bandwidth mean, exceeding Heterogeneous Earliest Finish Time (HEFT) by 0.05 to 0.15 absolute, yet its gain over PPO-LSTM remained about 0.001 and was not Holm-significant at any bandwidth. Both PPO variants reached the validation threshold at median update five, against ten for the Transformer. Ablation identified no single driver: only removing advantage normalization produced a Holm-significant drop, and only under the energy-aware objective. The Transformer remained competitive, whereas discrete soft actor-critic reduced energy rather than latency. Matching the recurrent baseline in size and inference time, stabilization remains viable on resource-constrained devices but cannot replace retraining on much slower links.

An Evolutionary Weight Optimization Framework for Adaptive PROMETHEE-Based Decision Support Systems

2026-09-24T02:51:15+00:00

Multi-Criteria Decision-Making (MCDM) methods are widely employed to support complex decision processes involving multiple alternatives and conflicting evaluation criteria. However, the performance of PROMETHEE-based decision support systems is strongly influenced by the criterion weights, which are commonly predetermined and treated as fixed parameters. Such static weighting may limit the adaptability and discriminatory capability of the resulting decision model. This study proposes an Evolutionary Weight Optimization Framework for Adaptive PROMETHEE-Based Decision Support Systems to address this limitation by integrating evolutionary optimization with the PROMETHEE ranking mechanism. The proposed framework treats criterion weights as optimization variables and iteratively searches for an improved weighting configuration before generating the final PROMETHEE ranking. The experimental evaluation was conducted using a dataset comprising 219 alternatives, with the performance of the proposed optimized PROMETHEE compared against a conventional baseline PROMETHEE model using fixed criterion weights. The evaluation focused on net flow distributions, alternative ranking changes, ranking stability, and the discriminatory behavior of the decision-support models. The results demonstrate that weight optimization substantially modifies the preference structure and produces a broader and more differentiated distribution of net flow values compared with the baseline approach. Several alternatives experienced notable ranking improvements or declines, while some alternatives maintained their positions, indicating different levels of sensitivity to the optimized weighting scheme. These findings demonstrate that evolutionary weight optimization can enhance the adaptability and discriminatory capability of PROMETHEE while preserving its interpretable ranking structure. The proposed framework provides a systematic approach for developing more adaptive, robust, and data-responsive MCDM decision support systems.

Cost-Aware Selective Text Classification with Calibrated Lightweight Models

2026-07-30T05:29:40+00:00

Reliable text classification requires calibrated confidence, abstention for uncertain cases, and explicit computational-cost reporting. This study evaluates cost-aware selective prediction for calibrated lightweight text classifiers on commodity CPU runtimes. The goal is to find a light-weight classifier that is both deployment friendly and achieves a balance between predictive accuracy, calibration, selective reliability, and computational efficiency. The pipeline trains Multinomial Naive Bayes, Logistic Regression, and LightGBM on TF-IDF features, applies post-hoc probability calibration, and accepts only predictions above a validation-selected confidence threshold. Calibrated Logistic Regression achieves the best practical accuracy-calibration-selective-risk-latency-review-budget trade-off in SMS spam, IMDb sentiment and AGNews topic classification experiments. Calibrated Logistic Regression achieves 0.9857 accuracy and 0.0070 ECE on SMS, 0.8718 accuracy and 0.0162 ECE on IMDb, and 0.8998 accuracy on AGNews. On IMDb, Logistic Regression sacrifices about one accuracy point relative to LightGBM but reduces p95 CPU latency from 2546.06 ms to 3.26 ms. Abstaining on the lowest confidence predictions moves accepted risk down from 0.0143 to 0.00457 (SMS) and 0.1282 to 0.1006 (IMDb). Thresholds are chosen on validation/calibration data, then frozen for a single test run. More generally, the results establish calibrated lightweight classifiers as strong and transparent baselines for cost-effective low-latency text classification in real-world deployments.

Implementation of a Smart Virtual Medical Assistant Using the XGBoost Method and NLP for Early Optimization of Preeclampsia Complications

2026-08-09T07:34:34+00:00

Preeclampsia is a pregnancy complication that significantly contributes to increased maternal and perinatal morbidity and mortality. Early detection remains challenging, as most prediction systems rely solely on structured clinical data and have yet to integrate unstructured information in the form of patient complaints. This study aims to develop a preeclampsia prediction model using the XGBoost algorithm by combining clinical data and text representations of patient complaints processed through Natural Language Processing (NLP). The main contribution of this study is an empirical evaluation of the effectiveness of incorporating NLP-based features in improving model performance compared to a model using clinical variables alone. The dataset consists of 100 medical records of pregnant women classified into Preeclampsia and Non-Preeclampsia classes. Complaint texts were represented using the Term Frequency–Inverse Document Frequency (TF‑IDF) method. Testing was conducted in three scenarios: Clinical Model, NLP Model, and Clinical+NLP Model. The results showed that the Clinical Model achieved an accuracy of 85% and an AUC-ROC of 0.92; the NLP Model yielded an accuracy of 75% and an AUC-ROC of 0.58; while the Clinical+NLP Model obtained an accuracy of 85% and an AUC-ROC of 0.91. However, evaluation metrics highly relevant for early detection Recall (sensitivity) and F1-score were not reported in this study. Both metrics are critically important given the high risk of undetected preeclampsia cases and should therefore be a primary focus in future development. The results clearly demonstrate that the addition of NLP features did not improve the performance of the clinical model; accuracy remained identical and AUC-ROC in fact decreased slightly. These findings confirm that the developed model is more appropriately utilized as a supportive early screening tool and is not intended to replace clinical diagnosis.

Deep Learning-Based Weather Prediction Model Using BMKG Big Data for Rice Planting Season Optimization in Indonesia

2026-08-04T11:32:08+00:00

This study develops and evaluates a deep learning-based weather prediction model built on BMKG (Indonesian Meteorological, Climatological, and Geophysical Agency) data, with the aim of informing rice planting season optimization in Deli Serdang Regency, North Sumatra, Indonesia. Two recurrent architectures, Long Short-Term Memory (LSTM) and Bidirectional Long Short-Term Memory (BiLSTM), were compared for daily rainfall prediction using 788 observations collected between January 2024 and March 2026, split chronologically into training, validation, and testing subsets. The LSTM model achieved a Root Mean Square Error (RMSE) of 24.42 mm, a Mean Absolute Error (MAE) of 15.20 mm, and an R² of 0.0420, while the BiLSTM model achieved a lower RMSE of 23.64 mm, a lower MAE of 14.89 mm, and a lower R² of 0.0233. Although BiLSTM produced marginally smaller errors, LSTM explained a slightly larger share of rainfall variability, and both R² values remained close to zero, indicating that neither model captured the full variability of the rainfall series. Because of its higher R², the LSTM-based forecasts were translated into an indicative cropping-calendar classification of wet, normal, and dry months as a preliminary basis for advancing, maintaining, or delaying the onset of the rice planting season in Deli Serdang Regency. Given the limited explanatory power of both models and the absence of baseline forecasts, repeated training, and time-series cross-validation in this study, these planting-season indications should be read as exploratory and complementary to, rather than a replacement for, the operational cropping calendar issued by BMKG and the Ministry of Agriculture. Future research should incorporate longer historical records, additional meteorological predictors, baseline models, and rigorous validation to develop more reliable, data-driven rice planting season decision support for Indonesia.

Integrated Enterprise Architecture for Coffee Shop Supply Chain and Customer Interaction Using TOGAF ADM

2026-08-17T18:04:23+00:00

Although the coffee shop sector continues to expand quickly, enterprise architecture research in this domain has generally treated the customer-facing point of sale and the upstream sourcing of coffee as two separate concerns, so that supply chain and customer interaction are rarely represented together. This study builds an integrated enterprise architecture model for a coffee shop enterprise that brings supply chain operations and customer interaction into a single design spanning three product lines: coffee beverage, non-coffee beverage, and food and snack sales. The study is conceptual and design-oriented; the enterprise modelled is a representative reference case reconstructed from operational patterns reported in earlier studies of independent coffee shops, not an empirically observed firm. The work follows the phases of The Open Group Architecture Framework Architecture Development Method, from the Preliminary Phase to Migration Planning, with ArchiMate as the modelling notation and the Business Model Canvas as the expression of the business logic. The findings show that the Architecture Vision phase connects three stakeholder groups and their strategic drivers to specific realizing components, namely integrated point-of-sale management, customer service management, and smart inventory management, while the Business Architecture phase shows that all three product lines share a recurring procurement-inventory-preparation-sales pattern and rely on a common supplier and point-of-sale application layer instead of separate systems. A preliminary analytical self-assessment of coverage and traceability indicates internal completeness, whereas expert validation remains future work. The model offers a reusable reference architecture for the digital transformation of small and medium coffee shop enterprises.

When Touch Cannot Be Digitized: ArchiMate Enterprise Architecture for Omnichannel Wellness Services

2026-08-13T09:13:39+00:00

Wellness providers such as massage studios are digitalizing their booking, membership, and home-visit channels, yet the service they sell still happens through a therapist's hands, a condition that omnichannel services and phygital services research, grown largely out of retailing, rarely translates into architectural specification. This study designs an enterprise architecture for a micro-enterprise wellness services provider in Banten, Indonesia, that plans three channels, namely regular on-site massage, monthly membership, and home service. Following design science research and TOGAF ADM phases A to D, six ArchiMate models were developed and then analysed through a digitizability lens that classifies every touchpoint as fully digitizable, digitally mediated, or irreducibly physical. Five of the nine normalized activities are fully digitizable and every one of them is information handling, two are digitally mediated, and the two irreducibly physical activities, the therapist's travel and the massage itself, appear in none of the six models. That omission is precisely what makes a distinct family of frictions architecturally invisible. Twelve anticipated frictions are therefore separated into boundary frictions, which exist because the service is physical, and generic digital frictions. The resulting target architecture surrounds the physical activity with an information envelope of identity assurance, shared availability, arrival status, and safety tracing. What cannot be digitized can still be surrounded by information that reduces uncertainty. The artifact is evaluated analytically and ex ante, no implementation exists yet, and systematic measurement is the necessary next step.

The Development, Trends, and Challenges of Smart Mosque Research: A Systematic Literature Review 2021–2026

2026-08-26T16:46:16+00:00

Smart Mosque research has expanded from energy efficiency and building automation toward wider digital services involving artificial intelligence, Internet of Things, data management, operational security, and inclusivity. Objective: This study maps the development, trends, and challenges of Smart Mosque research published between 2021 and 2026 using a Context-Technology-Impact framework. Methods: A PRISMA 2020-guided systematic literature review was conducted using reproducible PubMed and Crossref searches and second-reviewer validation. Records were screened against mosque/masjid or Hajj/Umrah technology relevance, assessed using an eight-item quality-appraisal instrument, and classified by evidence-access status. Results: The search produced 1,664 exported records and 870 records after deduplication. After staged screening, second-reviewer validation, and excluded-sample consensus, 72 records were retained, consisting of 10 full-text confirmed records and 62 limited-access/provisional records. The evidence shows four dominant directions: artificial intelligence for operational security and Hajj/Umrah crowd monitoring, energy and sustainability systems, Internet of Things and automation, and digital mosque-management platforms. Conclusion: Smart Mosque research is developing into a broader digital-operational ecosystem, but the evidence remains methodologically uneven because many records were available only at abstract or limited-access level and 2026 data were partial up to August 2026. Future studies should prioritize full-text-verifiable empirical research, multi-site evaluation, system reliability, user-centered assessment, privacy and data governance, and inclusive digital services.

Development of Deep Learning Based Augmented Reality Learning to Improve Student Motivation and Skills

2026-05-20T03:22:46+00:00

This research aims to develop Augmented Reality (AR) learning media based on Deep Learning to overcome the visualization barriers of abstract material in higher education. The main problem addressed is the low motivation and practical skills of students due to conventional static methods. Through the Research and Development (R&D) method with the ADDIE model, this research integrates Object Detection algorithms with YOLO into the AR system to intelligently recognize environmental objects and provide adaptive feedback. The results show that this application is highly feasible with object detection accuracy above 90%. Implementation in the classroom proves a significant increase in learning motivation (based on the ARCS model) and students' technical skills, where smart AR users can complete practicum tasks 25-35% faster than conventional methods. In conclusion, Deep Learning integration transforms AR into an effective active learning assistant that precisely enhances cognitive engagement and mastery of practical skills of students.

Enterprise Architecture for a Palm Oil Agribusiness ERP Using TOGAF and ArchiMate

2026-08-26T16:54:47+00:00

A palm oil agribusiness manages a long value chain, from estate operations and harvesting to milling, sales, and distribution, supported by workforce, finance, and maintenance functions. When these functions grow module by module without a shared architectural plan, the same master data is recorded more than once and the link between business strategy and the supporting systems becomes hard to trace. This study designs an enterprise architecture (EA) for an integrated Enterprise Resource Planning (ERP) implementation in an anonymised palm oil agribusiness company, using The Open Group Architecture Framework (TOGAF) Architecture Development Method (ADM) and ArchiMate 3.1 within a design science research approach. The design draws on the documented business process blueprints of six ERP modules, namely estate management, checkroll, production planning, plant maintenance, finance and control, and sales and distribution, from which 30 business processes and 20 principal data objects were extracted and modelled across the four ArchiMate layers, with an as-is versus to-be gap analysis. The result is an integrated ERP reference architecture, with a shared master data core and a four-stage migration roadmap, that connects the estate-to-sales value chain into one coherent model. Evaluation against four criteria, each with an explicit failure condition, found all 24 module-layer cells populated, all 30 processes traceable from a motivation element to a hosting node, and all four gaps covered by the roadmap. The contribution is a reusable EA reference model for palm oil agribusiness ERP, a domain that has received little attention in prior TOGAF-based case studies.

IoT-Based Server Room Environmental Monitoring Using ASHRAE TC 9.9 and RoC Alarm

2026-07-24T18:58:59+00:00

Server infrastructure is a vital asset for PT Hino Motors Manufacturing Indonesia (HMMI), whose performance is significantly influenced by the stability of temperature and humidity. Previously, environmental monitoring in the server room was still conducted manually and conventionally, leading to risks where hardware failures due to temperature anomalies were often detected too late. This research aims to design and implement a centralized IoT based monitoring system with the integration of ASHRAE TC 9.9 standards and the Rate of Change Alarm algorithm to overcome manual supervision constraints and minimize the potential for hardware damage. The system is built using ESP32 and ESP8266 microcontrollers as sensor nodes, the MQTT protocol for data transmission, and an interactive Next.js-based web dashboard. Frontend implementation includes real-time data visualization via WebSocket and dynamic sensor location maps. Sensor accuracy testing resulted in a temperature MAE of 0.22°C and a relative humidity MAE of 8.1 percentage points. These results indicate that the DHT11 provided adequate temperature measurements for the proposed monitoring application, although its relatively high humidity deviation remains a limitation of the current implementation. QoS testing using Wireshark proves the high reliability of the communication channel, yielding an average latency of 20.19 ms for the ESP32 and 50.85 ms for the ESP8266, packet loss ratio of 0%, and efficient bandwidth usage ranging from 200-450 bytes/s. The system has been successfully deployed on-premise at HMMI’s local server, providing a responsive monitoring solution for environmental anomalies and enhancing the IT team’s operational efficiency through systematic historical data reporting features.

Evaluating the Sensitivity–Specificity Trade-off in Risk Factor-Based Cervical Cancer Classification Using SMOTE and ACO-Optimized XGBoost

2026-08-26T17:07:10+00:00

Class imbalance is a major challenge in risk factor-based cervical cancer classification because positive cases are substantially less frequent than negative cases. This study aims to evaluate the effects of the Synthetic Minority Over-sampling Technique (SMOTE) and Ant Colony Optimization (ACO) on XGBoost performance and to investigate the trade-off between sensitivity, specificity, and overall discriminative ability. The Cervical Cancer (Risk Factors) dataset from the UCI Machine Learning Repository, containing 858 observations, was used with Biopsy as the binary classification target, comprising 803 negative and 55 positive cases. A total of 28 risk-factor features were retained after excluding diagnostic attributes with potential data leakage. Three classification scenarios were evaluated: XGBoost, SMOTE-XGBoost, and SMOTE-ACO-XGBoost. Model performance was assessed using Nested Repeated Stratified Cross-Validation, consisting of an outer 5-fold cross-validation repeated five times and an inner 3-fold cross-validation for hyperparameter and decision-threshold optimization. The results showed that SMOTE-XGBoost achieved the highest recall of 0.4182 ± 0.3340, whereas SMOTE-ACO-XGBoost obtained the highest accuracy of 0.7667 ± 0.1698 and specificity of 0.8042 ± 0.1950. Baseline XGBoost achieved the highest ROC-AUC and PR-AUC values of 0.6227 ± 0.0663 and 0.1290 ± 0.0739, respectively. The Friedman test indicated significant differences among the models across all evaluation metrics. Pairwise Wilcoxon signed-rank tests with Holm correction showed that ACO significantly improved accuracy and specificity compared with SMOTE-XGBoost but significantly reduced recall from 0.4182 to 0.2182. These findings demonstrate that class-imbalance handling and hyperparameter optimization do not produce a single model that consistently dominates all performance measures. SMOTE tends to improve sensitivity, whereas ACO optimization shifts the classifier toward higher accuracy and specificity at the cost of reduced sensitivity.

Clustering Pospay User Reviews Using Rule-Based Hybrid K-Means and TF-IDF

2026-07-19T07:25:17+00:00

Pospay is a digital financial services application created by PT Pos Indonesia, which offers a variety of financial transaction services. The huge volume of unstructured data and the exponential growth of the number of users’ assessments in Google Play Store makes it impossible to be analyzed manually. This research aims to classify Pospay customer evaluation to identify the main problems experienced by the user and provide recommendation in upgrading application services. This work adopts the Knowledge Discovery in Databases (KDD) technique. The inputs include 11000 user evaluations, of which 10943 are maintained after pre-processing. The input text has been vectorized using TF-IDF Bigram and the ideal number of clusters has been calculated using the Elbow Method and Silhouette Score. Then, a Rule-Based Hybrid K-Means technique was used, which integrated K-Means++ clustering with rule-based refinement to enhance the interpretability of the clusters. The findings produced five primary clusters that are related to verification and identity, system and error, login and account, transaction and balance, and positive reviews. Authentication and identification and transaction and balance were the most talked-about issues among users from these countries, accounting for 31.3% and 26.8% of discussions respectively. PCA visualization and word cloud analysis further supported the interpretation of each cluster. Overall, the proposed approach effectively grouped user reviews into meaningful topics and can assist developers in identifying service priorities to improve the quality and reliability of the Pospay application.

OpenStack Private Cloud Design and Cost Model for Real Estate Data Management

2026-09-17T19:12:11+00:00

Large real estate enterprises hold dense concentrations of regulated data while running enterprise applications at steady, predictable load, a combination poorly matched to public cloud pricing and to cross-border processing under Indonesia's Personal Data Protection Law. This study aims to propose an on-premises private cloud design for such an organisation and to model its cost position with all assumptions explicitly stated. The method comprises four parts: a real estate workload profile that maps application classes to criticality labels; a four-layer reference architecture built on OpenStack with a Ceph storage backend and an orchestration layer adapted from the Private Cloud Bespoke Orchestrator framework, in which a single data-criticality label propagates through storage placement, network isolation, snapshot cadence, and recovery ordering; an analytical estimate and a stochastic simulation of disaster recovery duration under three load conditions; and a parametric three-year total cost of ownership model that includes transition costs, evaluated through three scenarios and a Monte Carlo analysis. In the reference scenario, the private cloud becomes cost-favourable once annual public cloud expenditure exceeds approximately IDR 393 million, with a 90 per cent interval of IDR 272 to 600 million. Post-migration staffing and node acquisition cost are the dominant model assumptions. In simulation, recovering 1.0 terabyte of critical data meets the thirty-minute target in 63 per cent of runs at moderate load. The design has not been deployed or audited, so simulated recovery times rest on assumed parameters and the compliance posture is unverified.

Implementation of a Hybrid CNN-BiLSTM-Attention Model for Detecting Respiratory Sound Abnormalities

2026-07-24T16:23:03+00:00

Early detection of respiratory disorders generally relies on digital stethoscope devices which are expensive and have limited availability. Although the use of smartphones offers a highly practical alternative for independent screening purposes in the community, respiratory sounds recorded directly through built-in microphones are highly susceptible to environmental noise and acoustic characteristic differences (domain shift). Consequently, artificial intelligence (AI) models trained exclusively on public medical datasets often fail to adapt in real-world clinical scenarios, yielding an initial accuracy rate of only 57.78%. To address this critical limitation, this study proposes the application of a Transfer Learning technique with a Full Fine-Tuning approach on a hybrid Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (Bi-LSTM) architecture. Furthermore, an Attention Mechanism is specifically integrated so the model can focus its computational weight on very short-duration spectral anomalies that are often masked by background noise. Following the pre-training phase, the model was comprehensively calibrated using 45 primary respiratory sound samples from smartphones. The testing results demonstrated a substantial performance leap, with the overall accuracy rapidly increasing to 93.33%. The proposed system successfully identified all abnormal pulmonary cases without a single miss, achieving 100% Sensitivity. This finding Indicates that the integration of CNN-BiLSTM and Attention Mechanism is promising in handling acoustic distortion. Thus, the model is highly viable to be implemented as an early clinical decision support system with a showing high sensitivity on the evaluated localized cohort predictions.

Explainable Tree-Based Machine Learning for Event-Level Lateral Movement Detection in Active Directory

2026-09-17T18:59:24+00:00

Windows Active Directory governs authentication across enterprise networks, making it a primary target for lateral movement techniques such as Pass-the-Hash and Golden Ticket, which disguise themselves as legitimate administrative activity within high-volume Sysmon logs. Prior work reported near-perfect accuracy for tree-based machine learning on this task but relied on global feature-importance methods that cannot explain individual predictions, leaving the source of that accuracy unexamined. This study aims to compare four tree-based algorithms and to determine which technical indicators actually drive their decisions. The method comprises leakage-aware feature selection, stratified splitting, partial SMOTE oversampling applied to training data only, evaluation through Macro F1-Score, Matthews Correlation Coefficient, ROC AUC and per-event inference time, and instance-level interpretation using SHAP and LIME, supported by an ablation study across four feature sets. On 35 behavioral features Extra Trees achieved the highest Macro F1 of 0.9370, and every algorithm exceeded the throughput required for event-level stream processing. The ablation study produced the central result: on the full feature set the model reached a near-perfect Macro F1 of 0.9998, yet this stemmed largely from memorizing attack-tool identities rather than learning behavior, because precision on the credential-theft class collapsed to 0.7343 in the ablation setting once binary-identity features were removed, corresponding to 0.7345 for the same feature set in the main 200-estimator run. It is concluded that Explainable AI serves not only as a transparency mechanism but as a diagnostic instrument capable of exposing shortcut learning behind seemingly excellent detection performance.

IoT-Based Empirical Validation of the Open-Meteo Global Weather Model

2026-07-29T20:22:23+00:00

The deployment of affordable Personal Weather Stations (PWS) may expose a set of Citizen IoT Vulnerabilities, including hardware siting bias, sensor faults, and temporal-resolution mismatches, that remain under-documented. This preliminary single-day, Single-Site Case Study explores the cataloguing of those gaps in a tropical Microclimate setting. This paper evaluates a custom ESP32-based IoT station deployed for a single day (4 May 2026) in Curug Sangereng, Tangerang, Indonesia, sampling at 30-second intervals. Using a Fault-Aware Quality Control protocol and an autocorrelation-adjusted effective sample size () to correct for Temporal Pseudo-Replication, the IoT record was benchmarked against Open-Meteo’s 15-minute forecast, BMKG (3-hourly forecast), and Meteostat (hourly archive) datasets. Variance decomposition showed that sub-15-minute fluctuations account for up to 53.9% of total-day wind-speed variance (and up to 67.1% for sub-3-hour fluctuations), compared to under 1.5% and 38% for temperature, pressure, and relative humidity, demonstrating that high-frequency sampling value is variable-dependent rather than universally necessary. Cross-dataset comparisons revealed severe hardware siting biases: Open-Meteo ran cooler than the IoT station, while BMKG ran warmer than Meteostat () but cooler than IoT (). This sign reversal reflects unshielded rooftop heat bias at the IoT site. Poor wind-speed agreement across all pairs further indicated local anemometer obstruction. These single-day, single-site results are offered as a catalogue of PWS setup pitfalls and a worked diagnostic protocol. They motivate four falsifiable hypotheses for a multi-day confirmatory study rather than establishing general conclusions.

Algorithms and Web Architecture for Lightweight Multiplayer Archipelago Monopoly Game Engine

2026-09-24T02:55:09+00:00

Browser-based turn-based simulations frequently encounter client-side rendering bottlenecks and transactional race conditions when operating on low-bandwidth networks without dedicated server daemons. This study introduces an optimized web game architecture and programmatic pipeline implementing canonical economic board game rules contextualized within Indonesian geography. The proposed framework integrates four core algorithmic solutions: an atomic advisory file-locking transaction system utilizing flock() for room-level state concurrency control without relational database engines; a category-interleaved deck shuffling algorithm establishing mathematical guarantees against prolonged penalty sequences; an autonomous heuristic decision model that dynamically balances property acquisition costs against risk-adjusted liquidity reserves; and real-time audio synthesis executed entirely via native browser oscillators. Rigorous experimental evaluation spanning computational simulations, ApacheBench load tests, and user testing verified system capabilities. The card distribution mechanism completely bounds consecutive adverse event occurrences to two draws across two thousand execution cycles, mitigating punitive card clustering. In two hundred simulated four-player matches, the autonomous agent demonstrated a thirty-five percent victory frequency, significantly outperforming greedy over-leveraging and naive random baseline strategies while sustaining sub-millisecond calculation speed. Concurrency benchmarking confirmed server throughput of four hundred fifty requests per second under peak load with zero data corruption, and thirty participants reported a composite system usability score of eighty-four. These outcomes validate that disciplined algorithm and programming methodologies coupled with lightweight state isolation deliver high-performing multiplayer gaming experiences over standard web protocols without requiring complex server daemons.

An Integrated AHP–PROMETHEE II Framework for Refrigerated Transport Service Provider Selection

2026-08-02T06:20:33+00:00

Refrigerated transport is essential for preserving the quality of temperature-sensitive products and maintaining continuity across cold chain logistics. Selecting a refrigerated transport service provider is therefore a multidimensional decision involving technical reliability, operational capability, commercial conditions, sustainability, and cost. This study develops an integrated framework combining the Analytic Hierarchy Process (AHP) and the Preference Ranking Organization Method for Enrichment Evaluations II (PROMETHEE II) to evaluate six providers in Ho Chi Minh City, Vietnam. Assessments from nine experts were used to determine the weights of four criteria groups and fifteen subcriteria through AHP, while PROMETHEE II was applied to establish the provider ranking. The Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) and group-weight sensitivity analysis were used to examine the robustness of the results. Service Quality and Cost were the most influential criteria groups, with weights of 0.3981 and 0.2802, respectively. Service cost, temperature compliance, and quotation transparency received the highest subcriterion weights. Provider S2 ranked first under PROMETHEE II, with a net outranking flow of 0.3144, and retained this position under TOPSIS and across most variations in group weights. The proposed framework enables firms to compare providers using a consistent set of criteria, identify the preferred alternative, and assess whether the decision remains stable under different ranking methods and weighting assumptions.

Evaluating Kubernetes Autoscaling and Load Balancing to Improve Student Feedback Application Performance

2026-08-12T09:24:01+00:00

The increasing use of digital platforms in higher education requires student feedback applications to maintain stable performance under growing workloads. This study evaluates Kubernetes autoscaling and load balancing through an integrated deployment in which multiple application pods, Kubernetes Service-based traffic distribution, Traefik routing, and Horizontal Pod Autoscaler configuration were enabled together. A conventional single-container deployment and the Kubernetes-based deployment ran on the same virtual machine and used the same intended duration-based workload pattern with a maximum of 294 virtual users. Each deployment scenario was evaluated through one independent workload execution. The results are interpreted as descriptive observations of the tested deployment configurations rather than statistically replicated experiments. Interval-level observations within each run were treated as temporally correlated measurements rather than independent experimental replications. Across the recorded runs, the Kubernetes-based deployment recorded 75.36% lower mean virtual-machine-level processor utilization, 95.87% lower mean interval-level p99 request duration, and 139.13% higher mean request rate than the single-container deployment.. For the retained workload-band time series above 200 virtual users, mean processor utilization was 99.22% for the single-container deployment and 39.53% for the Kubernetes-based deployment. The corresponding mean interval-level p99 values were 23.605 and 0.188 seconds, while mean request rates were 45.32 and 255.10 requests per second. The single-container deployment recorded failed-request rates of up to 3.14 requests per second in the retained time series, whereas no failed requests were recorded for the Kubernetes deployment. These results support the performance comparison of the complete configuration in which autoscaling and load-balancing mechanisms were enabled. However, replica-count, HPA-event, and per-pod traffic-distribution data were not collected, so the individual causal contributions of autoscaling and load balancing cannot be isolated.

Explainable AI-Based Hybrid LightGBM–LSTM for Banking Fraud Detection

2026-08-15T17:07:25+00:00

Banking fraud detection faces changing transaction behaviors, class imbalance, and the need to interpret the outputs of complex models. This study develops a Fraud Detection and Analysis System (FDAS) as a Decision Support System in the form of a prototype that integrates tabular and temporal modeling, imbalance handling, operational validation, and post-prediction interpretation. A quantitative experimental approach was employed using a synthetic transaction dataset constructed based on the characteristics of observed fraud and normal patterns, comprising 50 accounts, consisting of 25 fraud accounts and 25 normal accounts, covering the period from January to December 2025. Account-level splitting was used to prevent historical information leakage across accounts, while TPH-SMOTE and KHOI-SMOTE were applied only to the training set. LightGBM and LSTM were trained independently to represent tabular and temporal data, respectively, and their probabilities were subsequently combined through weighted decision-level late-fusion, with SOP serving as the operational validation layer. SHAP was employed as a post-hoc explanation to support interpretability, traceability, and human oversight. Across 1,000 paired evaluation units, the final configuration achieved an Accuracy of 97.4%, Precision of 96.8%, Recall of 96.2%, F1-score of 96.5%, and ROC-AUC of 0.984. Comparison across configurations indicated that the highest performance was achieved by the integrated configuration within the experimental environment used. FDAS demonstrates end-to-end integration to support fraud investigation through recommendations and interpretive information, while the final decision remains with the analyst. The use of synthetic data, 50 accounts, and a prototype status limits direct generalization to operational banking populations and production readiness.

Enterprise Architecture for Multi-Sided E-Commerce Marketplaces: A Completeness-Driven TOGAF-Zachman Approach

2026-08-26T17:15:57+00:00

Indonesian micro, small, and medium enterprises (MSMEs) are moving into digital commerce, but the move remains scattered. National statistics show that 94.76% of e-commerce businesses sell through instant-messaging applications and only 17.23% through a marketplace platform, while 48.03% still arrange delivery themselves and cash remains the dominant settlement method. Multi-sided marketplace platforms can consolidate this activity, but designing them is architecturally demanding. Existing enterprise architecture (EA) studies also seldom verify whether every fundamental descriptive dimension of the enterprise has been addressed. This study proposes a completeness-driven EA for a multi-sided e-commerce marketplace. It integrates the TOGAF Architecture Development Method (ADM) as the design process, ArchiMate 3 as the modeling language, and the Zachman Framework’s 5W1H ontology (What, How, Where, Who, When, Why) as an explicit completeness lens. Within a Design Science Research approach, the artifact is developed across TOGAF ADM Phases A to D and evaluated through architectural consistency checking, gap analysis, and a structured completeness assessment that maps every model element onto a perspective-by-aspect matrix under a stated population rule. Of the 24 cells assessed, 20 are fully populated, 3 are only partially populated, and 1 is weak. The What, How, Where, Who, and Why aspects are answered at all four perspectives, while the temporal When aspect accounts for every cell that falls short. The study contributes a reusable blueprint for marketplaces that advance SME digitalization, an escrow mechanism modeled end-to-end including its dispute path and regulatory driver, and a replicable procedure for assessing architectural completeness.

Implementation of Soft Voting Ensemble Learning for Early Detection of Type 2 Diabetes Mellitus

2026-08-14T01:12:55+00:00

Type 2 Diabetes Mellitus is a chronic disease that requires early detection to minimize the risk of complications. This study aims to develop a machine learning model for the early detection of Type 2 Diabetes Mellitus using an ensemble Soft Voting approach. The proposed model combines three algorithms, namely LightGBM, Random Forest, and Support Vector Machine. The dataset used consists of patient health attributes, including pregnancies, glucose level, blood pressure, skin thickness, insulin level, body mass index, diabetes pedigree function, and age. The preprocessing stage was conducted to prepare the data before the training process, including data balancing so that each class contained the same number of records, namely 500 diabetes data and 500 non-diabetes data. The evaluation results show that the model achieved an accuracy of 80%, with precision, recall, and f1-score of 0.80 for both classes. In addition, the Area Under the Curve value of 0.89 indicates that the model has a very good capability in distinguishing between diabetes and non-diabetes classes. The system was also implemented in the form of a Graphical User Interface to facilitate patient clinical data input and display risk detection results. Based on these results, the developed Soft Voting model can be used as a practical and sufficiently accurate tool for the early detection of Type 2 Diabetes Mellitus.

Predicting E-Commerce Greenwashing Using Random Forest Machine Learning: Legal Analysis and Marketing Management Strategies

2026-09-10T02:53:49+00:00

The rapid growth of e-commerce and the increasing use of environmental claims in product marketing have created challenges in identifying potentially misleading green marketing communication. This study aims to develop a machine learning framework for predicting greenwashing indications in e-commerce products and interpreting the findings from consumer protection and sustainable marketing perspectives. A dataset of 610 products from Shopee and Tokopedia was independently annotated and validated by two annotators. Product descriptions were processed using Natural Language Processing (NLP) and represented using Term Frequency–Inverse Document Frequency (TF-IDF) with unigram and bigram features. Random Forest was employed as the primary classifier, while Support Vector Machine (SVM) and Multinomial Naive Bayes were used as benchmark models. A leakage-safe feature engineering configuration was additionally evaluated. Model performance was assessed using Accuracy, Precision, Recall, F1-score, confusion matrices, and 5-fold stratified cross-validation. On the independent test set, Random Forest with TF-IDF achieved the highest Accuracy (92.62%) and F1-score (95.81%), while 5-fold cross-validation showed that Random Forest with TF-IDF and Strict Feature Engineering achieved the highest mean Accuracy (95.49%) and F1-score (97.42%). Feature importance analysis identified textual structure and specific lexical patterns as predictive signals associated with model decisions. The findings indicate that the proposed framework can support early screening of potentially misleading environmental claims, consumer protection assessment, and more transparent and verifiable green marketing practices. However, model predictions represent predictive signals rather than causal evidence or legal determinations and require further factual and contextual verification

Comparative Evaluation of Random Forest and XGBoost for Bearing Capacity Prediction On Slopes

2026-07-14T16:17:09+00:00

Predicting the ultimate bearing capacity of shallow foundations is crucial in geotechnical planning because it is directly related to the safety and stability of the structure. This is especially true for foundations located at the edge of a slope. However, conventional analytical methods often fail to capture the complex and nonlinear relationships between soil and foundation parameters. This study evaluates the performance of Random Forest (RF) and Extreme Gradient Boosting (XGBoost) ensemble machine learning models for predicting the ultimate bearing capacity (q_ult) of shallow foundations in non-cohesive soils on slopes, with emphasis on methodological rigor and physical model interpretability. The dataset consists of 393 secondary data points obtained from numerical modeling based on the Finite Element Method (FEM) in a previous study. Input variables include slope angle (sin β), foundation width (B), foundation depth (Df), and soil angle of internal friction (tan φ), while ultimate bearing capacity (q_ult) is used as the output variable. Model performance was evaluated using the coefficient of determination (R²), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and the A20 index. Both models achieved a test R² of 0.97. XGBoost demonstrated superior generalization with a test RMSE of 7.130 kN/m² and MAPE of 8.833%, while Random Forest achieved a higher A20 index of 88.136% compared to 86.44% for XGBoost.

Usability Evaluation of the PSP Mobile Application Based on User Experience

2026-08-26T16:57:15+00:00

PSP Mobile is an application developed to support digital payment services and the delivery of billing information. This study aims to evaluate the usability of the PSP Mobile interface based on user experience, specifically focusing on ease of use, task completion effectiveness, completion time, usage errors, and information clarity. The study employed a usability evaluation approach involving eight users who performed several task scenarios, ranging from accessing the app and logging in to locating the payment menu, completing the payment process, and obtaining proof of payment. Data were collected through usability testing, direct observation, and post-test interviews. The evaluation results indicate that, overall, users were able to successfully complete the tuition payment process. The average time required for users to locate and access the payment menu was 38.5 seconds. However, several usability issues were identified, such as an excessive number of icons on the home screen, relatively small text size, visually similar symbols, limited button visibility, and a lack of information regarding where proof of payment is stored. These findings demonstrate that while the application supports key user activities, improvements are still needed regarding navigation, information readability, and feedback presentation. The evaluation results serve as the basis for design improvement recommendations aimed at making PSP Mobile easier to understand, more efficient to use, and better aligned with user needs. These improvements are expected to enhance usability and the user experience for individuals with varying levels of digital literacy, enabling more optimal use of the application.

Robust Multi-Sensor Machine Learning Models for Goat Barn Microclimate Automation

2026-07-14T02:47:41+00:00

Conventional goat barn management faces critical productivity drawbacks due to unmonitored microclimatic fluctuations, where environmental heat stress and toxic gas accumulations suppress livestock welfare. This study proposes "CapraFarm," a proactive thermal control architecture that integrates ambient temperature, relative humidity, and ammonia gas monitoring into a multi-criteria predictive automation pipeline. To stress-test algorithmic resilience against physical hardware degradation, a 5,000-entry dataset derived from sensor calibration error margins was injected with 5% to 10% Additive Gaussian White Noise. Six machine learning algorithms—including Random Forest, Extra Trees, LightGBM, XGBoost, Support Vector Machines, and Gaussian Naive Bayes—alongside conventional and hysteresis-enhanced rule-based baselines were evaluated using Stratified 10-Fold Cross-Validation. Experimental results revealed that Random Forest achieved the highest predictive performance, maintaining a classification accuracy of 97.12% ± 0.89%, a macro F1-score of 0.9354, and an Area Under Curve of 0.9941. This performance significantly outperformed the hysteresis-enhanced rule-based baseline (94.58% accuracy) under McNemar’s statistical test (). Feature importance attribution using SHapley Additive exPlanations confirmed that ammonia gas spikes exert critical non-linear overrides for instant hazard mitigation. Hardware-in-the-Loop profiling established that while Random Forest is optimal for centralized cloud optimization, Gaussian Naive Bayes excels in edge microcontroller environments with a minimal memory footprint of 45 Kilobytes and an execution latency of 1.94 milliseconds. This study solidifies a dual-layer Hybrid Edge-Cloud Paradigm for continuous climate control, laying a validated computational foundation for future long-term in-vivo physical field deployments in smart livestock facilities.

Implementation of YOLOv11 on a Rotating Mechanical System for Smart Waste Sorting

2026-09-15T17:22:54+00:00

Waste management still faces limitations in efficiency and consistency in manual sorting. Hence, a computer vision-based smart waste-sorting system integrated with mechanical actuation is required. This study applies YOLOv11 for object detection and waste classification in a smart waste-sorting system integrated with a rotational mechanism that directs objects into organic and inorganic categories. The dataset comprised 17.231 waste images collected from multiple publicly available datasets on Kaggle, resized to 512 × 512 pixels, normalized, and augmented. The YOLOv11n model was trained for 100 epochs with a batch size of 16, an automatic optimizer configuration, and an initial learning rate of 0.01 to identify five classes: Organic, Paper, Glass, Plastic, and Metal. The system transmitted detection results from a laptop or mini PC to the ESP32 via UART to control the sorting mechanism using two servos. The five detection classes were mapped to two operational categories: organic and inorganic. Model evaluation resulted in an overall precision of 77.3%, recall of 67.9%, F1-score of 72.29%, mAP@50 of 78.2%, and mAP@50–95 of 59.1%. Physical testing yielded 178 successful trials out of 180 (98.89%). The average system response time was 1.78 ± 0.16 seconds, rotation movement time was 1.77 ± 0.03 seconds, and total processing time was 3.54 ± 0.17 seconds. The mechanism operated smoothly for objects weighing 2–159 grams during testing. The results indicate that integrating YOLOv11 with a rotational mechanism can support automated smart waste sorting under the tested conditions.

Implementation of Integrity Zone Document Classification Using IndoBERT Model and Logistic Regression

2026-07-24T18:55:01+00:00

The Indonesian government's Integrity Zone program mandates systematic classification of administrative documents into hierarchical compliance codes, yet manual categorization remains labor-intensive, inconsistent, and unscalable for higher education institutions. This study aims to develop and evaluate an automated multi-label classification pipeline that maps Indonesian bureaucratic documents to hierarchical compliance codes while maintaining computational efficiency for institutional deployment. A curated dataset of 330 Integrity Zone documents from the Faculty of Mathematics and Natural Sciences, Universitas Negeri Medan, annotated across 82 hierarchical codes, was processed using a frozen IndoBERT encoder to extract 768-dimensional contextual embeddings. These features were classified using a balanced One-vs-Rest Logistic Regression model, with decision thresholds optimized via grid search on a held-out validation set to balance precision and recall. The complete pipeline was deployed as a REST microservice integrated into an existing PHP-based document management system. On a held-out test set of 66 documents not used for training, validation, or threshold selection, the pipeline achieved a Macro F1-score of 0.872, Micro F1-score of 0.894, Hamming Loss of 0.082, and a samples-averaged accuracy of 0.917 (pooled label-wise accuracy 0.918; subset accuracy 0.412). The optimized decision threshold of 0.38 favored recall over the conventional 0.50 cutoff, consistent with the higher institutional cost of missing a relevant compliance code. End-to-end inference latency averaged 2.36 seconds on the deployment server. The proposed pipeline shows practical viability as a decision-support tool for resource-constrained public institutions operating under mandatory human verification; the present evaluation is nonetheless limited by a small held-out test set, and several methodological details are reported in full in the Method section to support reproducibility and to rule out validation leakage.

Multi-Objective Chili Pepper Production Optimization Using HSA Under Price and Rainfall Variability

2026-09-12T18:57:08+00:00

Chili pepper production faces uncertainty associated with price dynamics and agroclimatic conditions, particularly rainfall variability. This study develops a multi-objective optimization model based on the Harmony Search Algorithm (HSA) to determine chili pepper production levels while considering gross revenue, exposure to rainfall deviations, and total production. The analysis uses 36 monthly periods from January 2022 to December 2024. Three objective functions are formulated: maximizing gross revenue, minimizing exposure to rainfall deviations, and maximizing total production. The reported HSA run, using 1,000 iterations, produced a fitness value of 0.6537, with gross revenue of Rp462.463.225.492.496,44, rainfall-deviation exposure of 804.788.341,60, and total production of 9.158.801,66 tons. The corresponding normalized objective values were 0.8165, 0.3476, and 0.7969. A post-hoc sensitivity analysis of the objective weights shows that the aggregated fitness changes when greater emphasis is placed on revenue, rainfall exposure, or production, indicating that the equal-weight solution should be interpreted as one compromise specification rather than a unique optimum. Because annual FAOSTAT production was evenly disaggregated into monthly baseline values, the production series has limited temporal variation (CV 0.56%); consequently, correlation and regression results should be interpreted as exploratory rather than as evidence of actual monthly causal relationships. The HSA was executed once in the original analysis; therefore, no statistical claim of algorithmic superiority or solution stability is made. The proposed model provides a framework for integrating economic and agroclimatic considerations into chili production planning, subject to these data and computational limitations.

Random Forest Implementation for Indonesian Coffee Identification Based on Digital Image Feature Extraction

2026-07-24T16:24:08+00:00

Accurate identification of coffee types is a critical challenge in Indonesia's coffee industry, which relies heavily on subjective human visual inspection. This study implements the Random Forest ensemble learning algorithm to classify three commercially important coffee species Arabica (Coffea arabica), Liberica (Coffea liberica), and Robusta (Coffea canephora) using digital image features. The dataset consists of 1,913 real coffee bean images from Roboflow Universe (Coffee Bean Type v1, CC BY 4.0), split into train (1,530), valid (287), and test (96). Feature extraction produces an 11-dimensional vector combining five Gray Level Co-occurrence Matrix (GLCM) texture features (Contrast, Energy, Homogeneity, Correlation, Dissimilarity) computed at four angles and six RGB color statistical features (mean and standard deviation per channel). The model is trained on 1,817 combined train-valid samples with optimal hyperparameters (n_estimators=200, max_depth=None, criterion=gini) selected via Grid Search with 5-fold stratified cross-validation. Evaluation on 96 held-out test samples yields accuracy 98.96%, macro precision 99.05%, macro recall 98.67%, macro F1-score 98.84%, and ROC-AUC 0.9996, with only one misclassification. Cross-validation confirms model stability at 98.07%±0.60%. Feature importance analysis identifies GLCM Energy (21.98%) and Homogeneity (14.57%) as the most discriminative features. The Robusta class achieves perfect classification (100% all metrics). Manual verification confirms that the reported metrics are consistent with the confusion matrix. These results demonstrate that Random Forest with GLCM and RGB color statistical features is effective for image-based coffee classification, achieving higher accuracy than several prior KNN and neural-network-based studies, although those studies differ in task, dataset, and class composition and are therefore not directly comparable.

Forensic-by-Design Framework Using Multimodal Behavioral Biometrics for Mobile Pharmacy Security

2026-09-17T18:57:31+00:00

In the mobile pharmacy sector, digital transformation increases the exposure of sensitive data and opens up insider threats that are not easily addressed by traditional security mechanisms such as login authentication and encryption, whose traces are only collected after an incident and often do not meet evidentiary standards. In this study we develop a multimodal behavioral biometric framework based on Forensic-by-Design principles to passively and continuously detect user interaction anomalies in mobile pharmacy applications. We develop three modules with hybrid architecture (Long Short-Term Memory network for keystroke dynamics, eXtreme Gradient Boosting for scrolling behavior and time threshold heuristic for detecting input that comes from paste or auto-injection). The keystroke module was tested on the Carnegie Mellon University keystroke benchmark (51 subjects) session disjoint split over 5 training runs and the framework was piloted in a mobile pharmacy application with 12 users and 19 sessions. For the keystroke module using key release intervals equal error rate of 11.1% was obtained (identification accuracy 74.4%) when key hold times and press latencies were added. For copy-paste heuristic requiring three consecutive intervals below 50 ms false positives on 20,400 genuine typing samples dropped from 16.3% to 0%. These results show that the proposed multimodal approach can support forensic readiness of mobile pharmacy applications by generating forensic-relevant interaction traces but more testing on a larger user population is needed to validate the scroll module, fusion layer and external validity of the findings.

API Call-Based Windows Malware Detection Using CNN-LSTM-Attention and Explainable SHAP

2026-07-29T20:13:24+00:00

The Windows operating system remains a primary target for cyberattacks, with modern malicious software increasingly relying on obfuscation, packing, and polymorphism to evade signature-based detection. This research addresses this challenge by developing a hybrid Convolutional Neural Network, Long Short-Term Memory, and Multi-Head Attention model to classify Windows malicious software from sequences of the first 100 dynamic Application Programming Interface calls. Drawing on the Kaggle "Malware Analysis Datasets: API Call Sequences" refined to 13,206 unique samples with a severe class imbalance of 95.36% malware versus 4.64% benign, the model employs a class-weighting strategy to ensure robust learning. Rigorous evaluation through five-fold cross-validation demonstrates that the proposed architecture achieves a mean Area Under the Receiver Operating Characteristic Curve of 0.9562 (± 0.0186) and an accuracy of 95.78% (± 1.05%). Ablation studies confirm that the Multi-Head Attention mechanism is statistically indispensable (p < 0.05), as its removal degrades performance to an Area Under the Receiver Operating Characteristic Curve of 0.9073. While competitive with ensemble baselines such as Random Forest and XGBoost, the proposed approach provides superior interpretability by integrating SHapley Additive Explanations. This enables local and global forensic accountability by pinpointing critical temporal execution windows specifically at steps t₁, t₅₀, and t₇₅ that drive malicious predictions. These results indicate that combining hybrid deep learning with post-hoc explainability produces malware detection that is both robust and forensically transparent.

Clinical Validation of an Offline Android Certainty Factor System for Pharyngitis Screening

2026-09-30T02:09:29+00:00

Certainty Factor (CF) expert systems for disease screening are widely developed in Indonesia, yet they are rarely validated prospectively against specialist diagnosis in patients who reflect the full spectrum of real-world presentations. A previously published offline Android expert system for acute and chronic pharyngitis reported 93.33% agreement with a single general practitioner on 30 retrospective cases that included no patients with other causes of sore throat. This study prospectively validates that system against the diagnosis of otolaryngologists as the reference standard, following the STARD 2015 reporting guideline. Consecutive patients presenting with sore throat entered their symptoms into the application before being examined by an otolaryngologist blinded to the application output. A confidence threshold was introduced so that outputs were classified as acute pharyngitis, chronic pharyngitis, or other, and performance was evaluated using a 3×3 confusion matrix, sensitivity, specificity, predictive values, F1-score, Cohen’s kappa, and ROC analysis. The robustness of expert-elicited weights was examined through Monte Carlo perturbation, and usability was measured with the System Usability Scale (SUS). Among 110 participants, overall accuracy was 71.8% (95% CI 62.8–79.4) and kappa was 0.55. Sensitivity was 89.1% for acute and 88.9% for chronic pharyngitis, with AUCs of 0.81 and 0.93. Most misclassifications involved patients with other diagnoses who were labelled as acute or chronic pharyngitis (22 of 31 errors). The mean SUS score was 68.2. The findings show that the system detects acute and chronic pharyngitis with high sensitivity but rarely recognises patients who have neither condition (sensitivity for “other” 21.4%), and identify examination-dependent signs as the principal constraint on unassisted self-screening.

Analysis of SIEM and NIDS Integration for Network Monitoring

2026-07-31T07:09:07+00:00

Network security monitoring is generally still carried out manually (dashboard-gazing), creating a time gap between the moment an incident is detected and the moment it is recognised and responded to by the administrator. This condition is aggravated by a Network Intrusion Detection System (NIDS) that works on its own, because its detection logs remain raw and unstructured, making them difficult to manage and analyse. This study proposes the integration of a Suricata-based NIDS with a Wazuh-based Security Information and Event Management (SIEM), on the grounds that Wazuh is able to manage and organise the raw Suricata logs into structured data while correlating them centrally. The integration was carried out by connecting the Suricata sensor to the Wazuh Server, so that the detection logs (eve.json), which were originally in raw JSON format, could be parsed into alerts with rule IDs and levels that are easy to analyse. Testing was conducted through the simulation of five attack scenarios (brute force, LFI, directory brute force, SQL injection, XSS) over 12 days, producing a correlation of 25,414 security logs covering all scenarios. The security readiness measurement using the KAMI Index v5.0 in the Technology Area shows that 5 of 35 criteria (14.3%) are directly fulfilled and 7 criteria (20%) are partially fulfilled. As an added value, all alerts resulting from the integration were forwarded through webhook-based Discord notifications without any delivery failure, shortening the time between a threat being detected and that threat becoming known to the administrator.

Bayesian-Optimized Word2Vec-BiLSTM for Malicious Prompt Detection in LLMs

2026-08-07T18:07:42+00:00

Large Language Models (LLMs) are increasingly utilized across a wide range of artificial intelligence applications, but their growing adoption also introduces security concerns related to malicious prompts that may manipulate model behavior. This study investigates an optimized Word2Vec-BiLSTM approach for detecting binary malicious prompts using the Malicious Prompt Detection Dataset (MPDD). Four deep learning configurations, namely FastText-BiLSTM, FastText-BiGRU, Word2Vec-BiLSTM, and Word2Vec-BiGRU, were evaluated alongside a TF-IDF combined with Logistic Regression (LogReg) baseline under consistent experimental conditions and across three random seeds. Bayesian Optimization was applied to identify effective hyperparameter configurations and quantify performance differences between manually configured and optimized models. The models were evaluated using precision, recall, F1-score, accuracy, false-positive rate (FPR), false-negative rate (FNR), precision-recall area under the curve (PR-AUC), and computational cost. The optimized Word2Vec-BiLSTM with seed 128 achieved the highest overall performance across the primary classification metrics: 97.03% precision, 96.86% recall, 96.89% F1-score, and 96.89% accuracy, along with an FPR of 0.0070, an FNR of 0.0557, and a PR-AUC of 99.39%. The optimized deep learning models generally outperformed the TF-IDF + LogReg baseline and exhibited relatively similar computational characteristics across the evaluated deep learning configurations. These findings demonstrate that Bayesian hyperparameter optimization can improve the performance of the Word2Vec-BiLSTM model for binary malicious prompt detection within the MPDD-based experimental setting.

Sentiment Analysis of Flip Application Reviews Using DBSCAN Outlier Removal and SVM

2026-08-17T18:17:20+00:00

Flip is an Indonesian financial technology application offering free interbank transfers, whose Google Play Store review column has become a large unstructured record of user experience that a development team cannot read manually. Such reviews are noisy, containing slang, repeated characters, and very short comments that behave as outliers. This study measures whether removing DBSCAN-detected outliers from the training data improves a support vector machine that assigns Flip reviews to three rating-derived categories. From 20,000 collected reviews, 15,241 remained after duplicate removal and preprocessing. The corpus was split once into stratified training and test partitions before any fitting, so that TF-IDF, the truncated singular value decomposition and DBSCAN were fitted on the training partition alone and both scenarios were evaluated on one identical held-out test set of 3,049 documents, repeated over five random seeds. Because the k-distance curve shows no pronounced knee, eps was fixed at the ninetieth percentile of that distribution, and eps and minPts were varied over a grid of nine combinations. Accuracy was 0.8023 without filtering and 0.8033 with it, a paired difference of +0.0010 with a 95 percent confidence interval of [-0.0017, +0.0038] and p = 0.347; macro F1-score moved from 0.6390 to 0.6402. No combination in the grid produced a distinguishable improvement. Of forty flagged documents inspected manually, none was genuine noise, and 91 percent of the flagged documents belonged to the majority positive class. The results give no evidence that density-based filtering benefits short-text classification.

Hierarchical Coverage Optimization of 5G New Radio Non-Standalone in Tourism Corridors

2026-08-22T05:26:21+00:00

A hierarchical coverage-optimization strategy is proposed for 5G New Radio Non-Standalone (NR NSA) networks in hilly tourism corridors with transient, high-mobility users. A drive-test campaign (7,522 measurement points, G-NetTrack Pro) was conducted on the Telkomsel 5G NR NSA network in the Jam Gadang corridor, Bukittinggi, West Sumatra, Indonesia, and benchmarked against three radio-planning scenarios simulated in Atoll. Three interventions were evaluated in order of increasing capital cost: antenna tilt and azimuth tuning, small-cell densification, and the addition of new macro-sites. Drive-test results show 97.55% of RSRP samples in the very-good band (≥ −85 dBm), confirming throughput — not coverage — as the dominant constraint. Against the simulation baseline (77.30% very-good RSRP, 70.55% very-good SINR), antenna tuning alone raises RSRP to 84.50% very-good and reduces the poor-band SINR from 2.03% to 0.06% at zero hardware cost. Small-cell densification achieves the strongest throughput gain (99.00% very-good TPUT-DL), while new macro-site addition achieves the strongest coverage gain (92.20% very-good RSRP, 0.00% poor RSRP). These results substantiate a cost-ordered hierarchical planning sequence transferable to similar tourism heritage corridors in emerging markets.

Comparative Study of TF-IDF and SBERT Feature Representations for Random Forest-Based Resume Classification

2026-08-18T16:52:38+00:00

Manual resume screening in high-volume recruitment is time-consuming and prone to inconsistent evaluation, motivating automated screening using Machine Learning (ML) and Natural Language Processing (NLP). Prior comparisons of TF-IDF and Sentence-BERT (SBERT) often differ in classifiers or task formulations, limiting direct comparison of feature representations. This study compares TF-IDF and SBERT using an identical Random Forest classifier for resume-to-role classification across 32 role categories. The dataset contains 10,102 resume-job description pairs from a public Hugging Face corpus. Three sources of data leakage were identified and mitigated, followed by five-fold StratifiedGroupKFold cross-validation and Bonferroni-corrected paired statistical testing. TF-IDF Standard achieved a mean macro F1 of 0.9953, TF-IDF Enriched 0.9964, and SBERT 0.9816. The difference between the two TF-IDF configurations was not significant (p = 0.1367), whereas both significantly outperformed SBERT (p = 0.0021 and p = 0.0015). Component-wise ablation showed that resume-only input retained high performance (macro F1 up to 0.9977), while job-description-only input achieved at most 0.2339, indicating that resume content provides the primary predictive signal. Equalizing the input budget to 256 tokens also preserved the TF-IDF advantage. Role-mismatch results further showed no performance degradation when job descriptions were replaced with descriptions from unrelated roles. These findings indicate that TF-IDF provides strong lexical representations for resume-to-role classification under the evaluated protocol, while highlighting the importance of leakage auditing and controlled input conditions.

Self-Sovereign Identity Using WalletConnect for Issuing, Verifying, and Revoking Academic Credentials

2026-09-09T16:58:01+00:00

Student data at vocational high schools is managed solely by the schools, so external parties must contact the school to confirm a record, while account access still relies on passwords vulnerable to credential theft. This study implements a Self-Sovereign Identity system for student academic data at SMK Bina Sriwijaya Indonesia Palembang that combines passwordless WalletConnect authentication with credential issuance, verification, and revocation anchored on a public blockchain. Following the Design Science Research Methodology, the artefact was built as a web prototype using React.js, Express.js, and PostgreSQL, connected to MetaMask and to a CredentialRegistry smart contract on Ethereum Sepolia. Users authenticate through an EIP-191 challenge-response signature, student identity is expressed as a did:ethr Decentralized Identifier, and academic data is issued as W3C Verifiable Credentials whose canonicalised Keccak-256 hash and status are recorded on chain, while the complete record remains off chain. Evaluation covered four levels: eighteen black-box functional cases, thirty smart contract unit tests, seventeen backend tests, and fourteen security cases at the authentication and verification boundary, plus gas measurement of every state-changing function on Sepolia. All seventy-nine cases produced the expected output, which demonstrates functional conformance and correct access control rather than a general guarantee of security, since independent penetration testing and a formal contract audit were not performed. Within these limits, a verifier can check authenticity, issuer, and validity independently, without confirmation from the school and without exposing personal data on a public blockchain.

LLM-Assisted Annotation in Comparative Sentiment and Trend Analysis of Indonesian K-12 Education on X

2026-06-21T19:21:43+00:00

Indonesia was ranked 69th out of 81 participating nations in the 2022 PISA test results, reflecting persistent challenges in K-12 education quality. Public discourse on these challenges has grown substantially on social media, yet automated sentiment analysis of Indonesian education-related text remains difficult due to informal language and implicit sentiment expression. Prior work on Indonesian education sentiment has focused on short observation windows or a single policy, without comparing transformer-based models directly against traditional classifiers or scaling annotation with large language models. This study aims to compare traditional machine learning classifiers and transformer-based models for sentiment analysis of Indonesian K-12 education discourse on X, as well as the trend of sentiments longitudinally and key terms associated with different sentiments. The dataset comprises 38,957 tweets in the Indonesian language collected between January 2022 and October 2025 that were annotated via a multi-stage, large language model-assisted few-shot classification pipeline with self-consistency sampling tiebreaker, reaching inter-annotator agreement of 82.01% (Cohen’s kappa = 0.672). This study evaluated the accuracy and macro F1 score of four traditional classifiers, including Logistic Regression, Support Vector Machine, Naïve Bayes, and Random Forest, alongside three transformer-based models, namely IndoBERTweet, IndoBERT-base-p1, and IndoBERT-large-p1. Transformer-based models performed better than traditional classifiers based on all metrics. IndoBERTweet reached the best performance with 0.884 accuracy and 0.859 macro F1, compared to Logistic Regression as the best traditional classifier achieved 0.768 and 0.714, respectively. Negative sentiment dominated public discourse, with “kurikulum merdeka” as the most frequent term across all sentiment categories.

Comparison of Classification Methods for Email Spam Detection

2026-08-26T17:01:57+00:00

Email spam remains a persistent cybersecurity problem, since unsolicited messages waste user time and often deliver phishing or malicious content. Prior comparative studies of conventional spam classifiers rarely state clearly whether their preprocessing avoids leakage between training and test data, or whether reported metrics come from a held-out test set or from cross-validation. This study addresses that gap by comparing five conventional classifiers — Naïve Bayes, Logistic Regression, Decision Tree, Random Forest, and Support Vector Machine (SVM) — under a pipeline in which the 80:20 train-test split is performed first, TF-IDF is fit only on the training partition, and SMOTE oversampling is applied only to the training data, so the test set stays unseen during model development. A public Kaggle email dataset of 5,157 messages (4,516 legitimate, 641 spam) was used, with spam explicitly defined as the positive class. On the untouched test set, SVM achieved the best overall performance (98.55% accuracy, 98.40% precision, 90.44% recall, 94.25% F1-score), followed by Random Forest (98.16% accuracy, 97.56% precision), while Logistic Regression obtained the highest recall (95.59%). Comparing these results with five-fold cross-validation on the already-oversampled training data revealed a large optimistic gap — up to about 30 points in precision for Naïve Bayes — demonstrating why resampling must be repeated inside each fold rather than applied once beforehand. The findings show that reported performance depends strongly on how resampling interacts with data splitting, and provide a transparent, leakage-aware baseline for future spam-detection research.