Abstract Due to the limitations of current defense methods—the neglect of spatial-temporal information and insufficient privacy guarantees, membership inference attacks (MIAs) on dynamic graph-based models remain inadequately addressed. This paper examines the vulnerability of dynamic graph-based victim models by employing a dynamic node-level MIA procedure. The attack involves three steps: training multiple shadow models, training the attack model using member and nonmember data, and querying the victim model to determine whether dynamic nodes were part of the victim training dataset. To strengthen defenses against MIAs, the paper proposes the Differential Private Spatial-Temporal Self-Attentions (DP-STSA) method, which enhances model resilience by integrating spatial-temporal self-attentions with differential privacy with adaptively allocated privacy budgets. Supplemental layers with spatial-temporal self-attentions reduce the model’s sensitivity to specific nodes, and differential privacy allocates different Laplace noise to supplemental layers for obfuscating gradients. Assuming that the attacker has a strong background knowledge of the victim models, experiments depend on four real dynamic graph datasets and simulate the attack procedure on four different models. Experimental results demonstrate that dynamic nodes are highly susceptible to MIAs. However, DP-STSA offers improved protection of dynamic node membership privacy while preserving the utility of the victim models, outperforming other mentioned defense methods.
Abstract An increasing number of wireless devices deployments are utilizing harvested ambient energy to increase system lifetime. The temporal profiles of the diverse environmental energy sources may be highly variable. So, creating an adaptive processor management mechanism enables the computing device to adjust its activity to changing environmental conditions and adopt an energy neutral operation mode. In this paper, we take into consideration any Priority Assignment (PA) rule jointly to preemptive scheduling. The paper proposes the priority assignment with energy harvesting) framework, a generalized scheduling approach that integrates a predictive look-ahead processor management mechanism into any PA rule, whether fixed or dynamic. We demonstrate that the resulting scheduler is optimal in that it allows to achieve energy neutrality with no deadline missing, no wasted energy, and no energy starvation whenever feasible with the PA rule. Our simulation results demonstrate that PA-H achieves a $100\%$ deadline success ratio as soon as the energy replenishment ratio reaches neutrality. Furthermore, we quantify the benefit of this approach, showing that it can reduce the minimum required storage capacity by $\sim 90\%$ compared with standard rate monotonic scheduling. This enables the design of smaller, more sustainable autonomous devices without compromising real-time reliability.
Advances in information technology have had a positive impact on various fields, including administrative management in the educational sector. At the Muhammadiyah Al-Furqon Islamic Boarding School in Tasikmalaya—the subject of this study—the process of issuing student leave permits was previously conducted manually, often leading to challenges such as delays in record-keeping, data errors, and inefficiency. Additionally, parents had to visit the boarding school in person to request a permit, which did not always guarantee approval. As a solution, this study designed and developed a web-based student leave permit information system that can assist administrators in automatically recording and printing permission letters, as well as make it easier for parents to submit permission requests online. This study employed the Research and Development (R&D) method using the Waterfall development model. Data collection was conducted through observations, interviews, literature reviews, and questionnaires. The feasibility study results, using the TELOS method, were as follows: Technical (96.7%), Economic (100%), Legal (100%), Operational (90.1%), and Schedule (88.5%), resulting in an overall score of 95.1%. These results fall into the "Highly Feasible" category, indicating that this website is highly effective in facilitating student permission requests for both parents and boarding school administrators.
The Urban Heat Island (UHI) effect in tropical urban settings arises from interactions among built surfaces, vegetation, water bodies, and urban energy dynamics. This study modeled Land Surface Temperature (LST) in DKI Jakarta using Random Forest and XGBoost optimized with RandomizedSearchCV and 5-fold cross-validation. The analysis used 5,821 grid points at approximately 300 m resolution and five predictors: road density, NDVI, NDBI, NDWI, and distance to green open space. XGBoost slightly outperformed Random Forest, achieving R² = 0.507 and RMSE = 1.830°C compared with R² = 0.497 and RMSE = 1.849°C, although the difference was not statistically significant (Wilcoxon, p = 0.352). The RF-XGBoost ensemble did not improve performance due to very high residual correlation (r = 0.988) and a theoretical ensemble standard deviation reduction of only ~0.3%. SHAP analysis identified NDBI as the dominant predictor (mean|SHAP| = 0.986), with the strongest interaction between NDBI and road density (0.101). Hyperparameter tuning changed model ranking, statistical significance, and the leading SHAP interaction pair.
This study was motivated by the manual transaction recording system used at Lapak Pak Iyan, an MSME engaged in scrap goods collection in Rawalumbu, East Bekasi. This condition makes it difficult for the owner to identify frequently traded types of scrap goods, so purchasing decisions are not yet supported by structured data. The research aims to apply data mining using the Naive Bayes algorithm to classify scrap goods based on sales frequency into three categories: High-Selling, Moderate-Selling, and Low-Selling. The research methods include data collection, preprocessing, calculating item frequencies, category labeling, and classification using RapidMiner Studio. The dataset consists of 63 transaction records from January to June 2026 covering seven types of scrap goods. Model testing is conducted using Leave-One-Out Cross Validation (LOOCV) because the available dataset is relatively small. The classification results are expected to provide practical guidance for the owner in prioritizing purchasing decisions based on historical sales patterns. Naive Bayes successfully classified the seven types of scrap goods into three categories with an accuracy of 100%. Therefore, the results can be used as a reference to improve stock management efficiency.
PT. Sumatera Ekspres still relies on manual recording of advertisement bookings and customer data. Therefore, this study aims to design and develop a web-based information system for advertisement booking and customer management that integrates an analytics dashboard with an automatic WhatsApp notification feature. The study employs the Research and Development (R&D) method combined with the Rapid Application Development (RAD) approach, comprising the Requirements Planning, User Design, Construction, and Implementation stages. The system was developed using the Laravel framework, MySQL database, and Fonnte API for WhatsApp integration, and was evaluated using Black Box Testing. The results show that the system successfully digitizes and integrates advertisement booking and customer data management, provides real-time data visualization through a Chart.js-based analytics dashboard, and automatically sends WhatsApp notifications to customers. Black Box Testing of 10 scenarios showed that all scenarios were valid, indicating that the tested features functioned according to their specifications. The integrated platform provides a more structured and centralized approach to managing advertisement and customer data at PT. Sumatera Ekspres.
The microclimate inside a greenhouse, especially temperature and humidity, must be monitored continuously to maintain conditions that support optimal plant growth. However, manual monitoring is often time-consuming, susceptible to human error, and can cause delays in decision-making. To address these issues, this study developed an Internet of Things (IoT)-based monitoring system using an ESP32 microcontroller and a DHT22 sensor. The DHT22 measures temperature and relative humidity, and the ESP32 processes the readings and sends them through a Wi‑Fi connection. The data are then visualized in real time on a Blynk dashboard accessible via a smartphone, enabling remote observation of environmental changes without requiring physical presence in the greenhouse. System testing was performed in a simulated greenhouse environment to evaluate sensor reading behavior, communication reliability, system stability, and the continuity of data updates. The results indicated that the dashboard refreshed automatically with an average delay of about 1 to 3 seconds. Overall, the proposed ESP32-based solution is simple, low-cost, and practical for greenhouse microclimate monitoring. It can serve as a foundation for further development toward automated monitoring and control by adding more sensors, improving data processing, and integrating control actuators.
Delayed study completion among undergraduate students is a complex issue associated with the interplay of internal and external factors. This study aims to identify and examine the dimensions of risk factors associated with delayed study completion among Information Systems students using Principal Component Analysis (PCA) with Varimax rotation. Primary data were collected through a structured questionnaire administered to 299 undergraduate students of the Information Systems Study Program, semesters 3 to 8, at UNPAM Viktor Campus, covering 17 operational indicators. The analysis reduced the 17 indicators into 4 major risk factor dimensions, explaining 50.15% of the total variance: (1) Mobilization & Work Factor (14.02%), (2) Lecturer Mentorship Factor (14.28%), (3) Academic Competence Factor (13.07%), and (4) Assistance & Integrity Factor (8.78%). Factor 4 (Assistance & Integrity) was characterized by indicators related to dependence on external resources, including the use of third-party/joki services (0.717), peer assistance (0.598), and AI tools (0.592). Factor loading evaluation also identified two weak indicators (trap items) with loadings below 0.40: P1 (Employment Status; 0.363) and P4 (Primary Transportation Mode; 0.194). These findings highlight the importance of hybrid thesis mentoring, assignment load management, and academic integrity oversight to support timely study completion.
AI shopping assistants increasingly employ agent-based retrieval, combining lexical search, structured filtering, and LLM-mediated selection. However, effective retrieval often requires knowledge beyond textual matching. This exploratory single-case study of an Indonesian grocery e-commerce assistant triangulates 279 observations, 52 failure traces, 74 practitioner-reported defects, a schema audit, and four practitioner interviews. The analysis identifies six retrieval-critical product knowledge dimensions and shows that knowledge externalization fails at two distinct representational layers. At the product data-model layer, essential fields—such as allergens, dietary constraints, and age suitability—were absent and remained missing despite architectural changes. At the retrieval-schema layer, available knowledge could not reach candidate sets: structured filters were present in only 29% of calls, and even correctly invoked filters frequently returned empty sets from non-empty pools. Addressing these layer-specific failures, the study proposes a structured knowledge framework based on the knowledge management process cycle. It positions GraphRAG as a future direction to enable hybrid structured retrieval for candidate formation and improve eligibility, substitution, and context-aware ranking.
Hotel booking cancellations are a critical problem in hotel revenue management because they can cause operational inefficiencies and financial losses. This study develops an explainable cancellation prediction model using CatBoost integrated with SHAP, based on real-world Property Management System (PMS) operational data from a budget hotel in Central Java, Indonesia. The dataset spans 67 months (October 2020–April 2026) and, after preprocessing and data cleaning, yields 74,826 independent reservation records from 80,110 raw entries. The cancellation rate is extremely low (1.56%), creating a severe class imbalance challenge. Instead of synthetic oversampling, the proposed method applies cost-sensitive learning via CatBoost’s scale_pos_weight, computed from the natural class ratio. Model performance is evaluated using hold-out validation (80% training, 20% testing). The proposed CatBoost model achieves an F1-score of 72.04%, with precision of 87.20% for the cancellation class and an AUC-ROC of 0.86, outperforming Random Forest and XGBoost baselines. SHAP analysis indicates that lead time, deposit type, and arrival month are the most influential features driving cancellation predictions. These findings support early warning decision-making for risk mitigation in hotel operations.
Ganoderma disease is one of the most destructive diseases affecting oil palm plants, causing basal stem rot, reduced productivity, plant mortality, and substantial financial losses. This study proposes an image enhancement pipeline for Ganoderma disease identification by combining Gamma Correction and CNN-Based Enhancement. Gamma Correction is applied to improve illumination and pixel intensity, while CNN-Based Enhancement is used to enhance structural and textural details through a deep learning approach. After enhancement, the processed images are classified into three categories: Healthy, Infected, and Initial Infection. The model achieved a test accuracy of 0.7714 with a test loss of 0.4165. For each class, the F1-scores of Healthy, Infected, and Initial Infection were 0.8571, 0.7500, and 0.7200, respectively. The results indicate that image transformation techniques can effectively support Ganoderma disease identification in oil palm plants. By improving image quality prior to classification, the model becomes more stable in emphasizing disease-relevant visual features under diverse lighting conditions. Overall, the proposed enhancement process, which integrates intensity correction with CNN-based reconstruction, offers a promising direction for developing automated detection systems that are more accurate, adaptive, and suitable for field deployment.
The increasing administrative workload of teachers makes manual preparation of assessment questions time-consuming and may lead to the reuse of questions from previous years. This study examines the application of the transformer-based Generative Pre-trained Transformer 2 (GPT-2) model for automatic question-and-answer generation in elementary school Natural and Social Sciences (IPAS). The model was fine-tuned using 239 Indonesian context-question-answer triples, with 90% used for training and 10% for testing. Fine-tuning was conducted using four dataset sizes: 50, 100, 150, and 239 samples, to examine the relationship between training data volume and generation quality. Model performance was evaluated by examining parameter changes between the pre-trained and fine-tuned models and using the ROUGE metric to measure textual similarity between generated and reference questions and answers. The 239-sample dataset produced the most coherent and contextually appropriate questions, with ROUGE-1, ROUGE-2, and ROUGE-L scores of 1.0 for questions and 0.41, 0.24, and 0.36 for answers. These findings suggest that GPT-2 fine-tuning can support automatic question-generation tools when sufficient training data are available.
Item-loan administration in one subdivision still relies on physical handover forms and non-centralized records, making loan status and history difficult to trace. This study developed and evaluated an item-loan information system based on Google Apps Script. The study applied the Waterfall model, covering requirements analysis, design, implementation, testing, and maintenance. Evaluation involved six actual users through total sampling, comprising three officers and three administrators. Functional suitability was assessed using User Acceptance Testing (UAT), while usability was measured using the System Usability Scale (SUS). The UAT score reached 90.56%, with 88.89% for officers and 92.22% for administrators. The mean SUS score was 74.58, categorized as acceptable and good, although three respondents scored below the reference value of 68. Implementation reduced physical-document searches, improved record completeness and loan traceability, and centralized data. The system met functional needs but still requires improvements in initial guidance, activity notifications, stock checks, and handover reporting. The results are limited to six users in one subdivision.
This study analyzes sentiment in InDrive user reviews from the Google Play Store using IndoBERT, Genetic Algorithm (GA), and K-Nearest Neighbor (KNN). A total of 2,000 reviews were assigned to three sentiment classes: 1,695 negative, 235 positive, and 70 neutral reviews. The pretrained indobenchmark/indobert-base-p1 model was used as a feature extractor by taking the [CLS] representation to produce 768-dimensional embeddings. The dataset was divided using a stratified 80:20 split into 1,600 training and 400 testing samples. The optimal K value was determined through stratified five-fold cross-validation on the training data. GA was applied only to the training set using a population of 30 individuals, 25 generations, a crossover rate of 0.8, a mutation rate of 0.005, and a feature penalty of 0.002. GA selected 250 features, reducing the dimensionality by 67.45%. IndoBERT + KNN correctly classified 365 of 400 test samples, achieving 91.25% accuracy (95% CI: 88.07–93.64%) and a macro F1-score of 66.88%. IndoBERT + GA + KNN correctly classified 362 samples, achieving 90.50% accuracy (95% CI: 87.23–93.00%) and a macro F1-score of 62.71%. Both models exceeded the 84.75% majority-class baseline. However, only three and two of the 14 neutral samples were correctly classified, respectively. GA substantially reduced feature dimensionality but did not improve predictive performance, indicating a trade-off between representation compactness and minority-class classification performance.
Abstract Federated learning is a technology that is used to protect data privacy in machine learning. Nonetheless, in federated learning, updating the global model requires the use of gradient descent algorithm, which involves multiple rounds of interaction between entities to complete the iterative updates, inevitably incurring massive computational and communication overhead. In 2020, Wang et al. first proposed a non-interactive federated regression scheme, which effectively improves the training efficiency of regression models while protecting the privacy of local training data. However, like most current federated regressions, it involves a third authority (TA) to generate keys for each entity, which poses a significant privacy risk and results in considerable communication overhead. From the view of security and practicality, this paper first proposes a multi-party homomorphic encryption algorithm named MPaillier. Furthermore, we have designed PNFR, a privacy-preserving federated learning scheme for regressions training built on the MPaillier algorithm. The participating entities of PNFR are the data owners and a cloud server, eliminating the need for a TA, thus enhancing the practicality and efficiency of the scheme. Experimental results demonstrate that our scheme is $\sim 10^{3}$ times faster than interactive federated regressions PrivFL and about 80% faster than non-interactive federated regressions VANE.
BKB Paud Ceria is an early childhood education institution that has traditionally disseminated school information through manual channels, requiring prospective students’ parents to visit the school. This study aims to develop a web-based company profile using WordPress as a medium for information and promotion. The system was developed using the Waterfall method over four months (May–August 2026), with system design represented through use case, activity, and class diagrams. Functional testing used Black Box Testing with fourteen scenarios, while user acceptance was evaluated using a ten-item Likert-scale questionnaire covering Usability, Reliability, and Usefulness. The questionnaire was distributed through Google Forms to 27 respondents selected using purposive sampling. The results show that all fourteen test scenarios were valid (100%), while the User Acceptance Test obtained a score of 1,234 out of 1,350, equivalent to 91.4%, which falls into the Very Good category. The results indicate that the website was well accepted as a medium for providing school information and supporting communication with parents through its integrated WhatsApp feature.
The development of Internet of Things (IoT) technology supports the implementation of automated student attendance systems to address limitations in attendance recording, data recapitulation, and real-time monitoring. This study aims to implement a Radio Frequency Identification (RFID)-based student attendance system integrated with IoT. The Prototype method was used, consisting of requirement analysis, system design, implementation, and testing. The system utilizes an RFID Reader RC522, NodeMCU ESP8266, HTTP Request communication, PHP API, and MySQL database. RFID data is transmitted through a Wi-Fi network and displayed on a web-based dashboard. Testing with 10 trials resulted in a 100% success rate for UID reading, data transmission, student validation, attendance data storage, and dashboard visualization. The average response time from card detection to data display on the dashboard was 1.82 seconds. These results indicate that the system can support real-time IoT-based student attendance monitoring.
Hospitals are essential healthcare institutions, making an efficient drug logistics system highly important. This study aimed to identify, analyze, and minimize activities that generated waste or did not provide added value in the pharmacy warehouse of RS. XYZ. Value Stream Mapping was employed to map the actual operational flow and produce a Current Value Stream Map, while Process Activity Mapping was used to classify activities into Value-Added, Non-Value-Added, and Necessary Non-Value-Added categories. Direct observation identified two dominant types of waste, namely Defect and Delay/Waiting, whose root causes were determined through Root Cause Analysis using the 5 Why method and Fishbone Diagram. Several improvement proposals were then developed, including routine Standard Operating Procedure training and additional staff, and illustrated in a Future Value Stream Map representing the improved process flow. Process Cycle Efficiency increased from 2.2% in the Current Value Stream Map to 57.5% in the Future Value Stream Map, an improvement of 55.3%, indicating that the implementation of Value Stream Mapping and Root Cause Analysis effectively reduced waste and improved the efficiency of the drug logistics system in the pharmacy warehouse.
Abstract Transportation revitalization index (TRI) forecasting is crucial for monitoring post-pandemic urban recovery and supporting adaptive traffic management. Existing methods mainly use graph-based spatiotemporal models to encode pairwise intercity relations and directly fuse epidemic observations with TRI features. However, TRI recovery under pandemic shocks often involves group-level intercity co-movements, while epidemic variables contribute unevenly to forecasting, limiting the modeling of heterogeneous higher-order dependencies and selective auxiliary effects. To address these issues, we propose PGSTHFM, a prototype-guided spatiotemporal hypergraph fusion model for regional multi-city TRI forecasting. Specifically, PGSTHFM employs a prototype-guided relation-decoupled hypergraph convolution module to encode hyperedge semantics and disentangle heterogeneous higher-order dependencies into multiple latent relation channels. We further introduce a two-stage epidemic-aware data fusion module that summarizes epidemic signals conditioned on TRI representations and then refines TRI features through epidemic-guided node-wise interactions. Finally, a residual gated temporal convolution module is employed to capture both short-term fluctuations and long-range temporal dynamics. Experiments on a real-world TRI dataset covering 29 Chinese cities demonstrate that PGSTHFM consistently outperforms representative baselines across multiple forecasting horizons and evaluation metrics. Compared with the strongest hypergraph baseline, PGSTHFM reduces the average MAE, RMSE, and MAPE by 7.69%, 8.43%, and 11.95%, respectively.
Abstract Due to too much ambiguity and lexical overlap in today’ social media texts, cyberbullying detection has been difficult. To detect cyberbullying contents and address gaps, a hybrid decoding-enhanced bidirectional encoder representations from transformers (BERT) with disentangled attention-bidirectional gated recurrent unit (DeBERTa-BiGRU)-Attention model (explainable context-aware cyberbullying BERT-based hybrid model) is proposed for multiclass classification, which outperformed baseline approaches across different metrics. DeBERTa model is used to model content and positional embeddings and enable fine-grained contextual representation; it utilizes disentangled attention with improved relative positional encoding. BiGRU is utilized resulting embeddings from DeBERTa, for sequential dependency and capturing bidirectional token relations. A token-level attention layer is used on top of these models for prioritizing both enhancing interpretability and predicting performance. The sequential integration of all three techniques together complements recurrent modeling with the transformer model, yielding a +4.2% F1-score over standalone DeBERTa. In this study, explainability evaluation demonstrates a high alignment between attention weights and gradient-based importance scores at the token level to preserve faithfulness. An efficient and comprehensible framework for fine-grained multi-class cyberbullying detection is provided by combining DeBERTa with BiGRU and an Attention layer, as demonstrated by experiments.
