Jurnal Teknik Informatika (Jutif)

Enhancing Accessibility in Local Government Data Portals via Retrieval- Augmented Generation: A Case Study on Satu Data Indonesia in Banyumas Regency

2025-08-06T03:26:03+00:00

Public access to local government data in Indonesia, such as that in the Satu Data Indonesia portal for Banyumas Regency, is severely hampered by outdated search interfaces and the technical complexity of handling heterogeneous data formats like PDF, Excel, and CSV. This research directly addresses this accessibility gap by designing, developing, and evaluating an intelligent question-answering system. We introduce a novel application of a Retrieval- Augmented Generation (RAG) architecture tailored for Indonesian local government data. The core novelty lies in our methodology for handling heterogeneous data formats (PDF, Excel, CSV) by integrating a low-code orchestrator (n8n) with a high-performance vector database (pgvector), a practical solution for a common public sector challenge. The system utilizes the text-embedding-3-large model for semantic understanding and gpt-4.1 for generating grounded, factual answers. The system's effectiveness was rigorously validated, achieving a perfect 100% score across accuracy, precision, recall, and F1-score on defined test cases. Crucially, usability testing with end-users confirmed the system is perceived as significantly more efficient and user-friendly than manual data searching. The primary impact of this work is a validated, replicable blueprint for local governments to democratize public information. By transforming complex data retrieval into an intuitive conversation, this research offers a practical AI solution to enhance governmental transparency and citizen engagement.

Multi-architectural Transfer Learning CNN for Klowong Batik Fabric Defect Classification

2025-06-13T08:09:45+00:00

Klowong is a base cloth that has been given a hot wax pattern as the initial stage in the batik making process but has not yet become a finished batik. Nowdays, written batik machine are available but still limited and production defects still occur, reducing the value of batik. Manual QC makes subjective assessments, so an accurate and efficient automated inspection system is needed for SMEs.This study proposes a defect classification approach on batik klowong fabric based on transfer learning using deep convolutional neural networks (CNN) architecture that has been verified to be reliable in image classification schemes. The basic models used include VGG16, ResNet50V2, InceptionV3, and MobileNetV2, with modifications to the fully connected layers to reduce parameter complexity. The dataset consists of 1000 klowong fabric images with a resolution of 224×224 pixels, with a ratio of 80:10:10 for training, validation, and testing. Data augmentation was applied to improve the generalization of the model. Evaluation is performed based on accuracy, precision, recall, F1-score, and inference time. The experimental results show that VGG16 has the best performance in the testing stage with 92% accuracy. The combination of VGG16 with conventional classifiers (SVM and Random Forest) significantly speeds up the inference time (up to 0.0001 seconds per image) but with a decrease in accuracy to 81-83%. Therefore, the VGG16 model with the modified final layer is recommended as the optimal solution with the best trade-off between classification performance and computational efficiency, especially for application scenarios on low-resource devices such as batik SMEs.

Multimodal Biometric Recognition Based on Fusion of Electrocardiogram and Fingerprint Using CNN, LSTM, CNN-LSTM, and DNN Models

2025-07-15T23:35:38+00:00

Biometric authentication offers a promising solution for enhancing the security of digital systems by leveraging individuals' unique physiological characteristics. This study proposes a multimodal authentication system using deep learning approaches to integrate fingerprint images and electrocardiogram (ECG) signals. The datasets employed include FVC2004 for fingerprint data and ECG-ID for ECG signals. Four deep learning architectures—Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), CNN-LSTM, and Deep Neural Network (DNN)—are evaluated to compare their effectiveness in recognizing individual identity based on fused multimodal features. Feature extraction techniques include grayscale conversion, binarization, edge detection, minutiae extraction for fingerprint images, and R-peak–based segmentation for ECG signals. The extracted features are combined using a feature-level fusion strategy to form a unified representation. Experimental results indicate that the CNN model achieves the highest classification accuracy at 96.25%, followed by LSTM and DNN at 93.75%, while CNN-LSTM performs the lowest at 11.25%. Minutiae-based features consistently yield superior results across different models, highlighting the importance of local feature descriptors in fingerprint-based identification tasks. This research advances biometric authentication by demonstrating the effectiveness of feature-level fusion and CNN architecture for accurate and robust identity recognition. The proposed system shows strong potential for secure and adaptive biometric authentication in modern digital applications.

A Hybrid LSTM–Smith Waterman Model for Personalized Semantic Search in Academic Information Systems

2025-06-04T10:36:16+00:00

The growing complexity of digital learning environments presents a critical challenge in computer science, particularly in designing intelligent academic systems capable of delivering context-aware and personalized content. Traditional academic information systems often rely on literal keyword matching, failing to interpret the semantic intent behind user queries and ignoring historical learning behavior. This study addresses these limitations by proposing a hybrid semantic search and recommendation model that integrates Long Short-Term Memory (LSTM) networks with the Smith Waterman algorithm. The LSTM component models temporal sequences of user interactions, while Smith Waterman enables local semantic alignment between user queries and learning content. Historical query logs and user-clicked topics are transformed into semantic vectors, which are further enhanced through a contextual graph and semantic relation matrix. Experimental results demonstrate the model’s effectiveness, achieving 89% accuracy, an F1-score of 0.89, and an AUROC of 0.88 by epoch 50. The hybrid architecture successfully captures the evolution of user interest and semantic relevance, outperforming baseline approaches. This research contributes to the field of computer science by bridging natural language understanding and sequential modeling to improve adaptive learning technologies. The proposed model offers a scalable foundation for developing intelligent recommendation systems in academic platforms, fostering improved learner engagement and efficiency.

Rainfall Forecasting Using SSA-Based Hybrid Models with LSSVR and LSTM for Disaster Mitigation

2025-07-09T08:27:04+00:00

Accurate rainfall forecasting is crucial for addressing the increasing risk of hydrometeorological disasters, particularly in tropical regions such as Semarang City, Indonesia. However, conventional forecasting models often struggle with inaccurate data and observations. This study proposes a novel hybrid combination of SSA-NMF with LSSVR and LSTM, offering high-resolution rainfall forecasting over multiple monitoring stations, to predict daily rainfall. As a preprocessing step, 15 years of daily rainfall data from six observation stations were denoised and decomposed using Singular Spectrum Analysis (SSA) combined with Non-Negative Matrix Factorization (NMF). This approach effectively handled data with many zero values, identified seasonal patterns or high-rainfall locations, and extracted key patterns. The prediction models were trained and validated using parameters optimized through RandomizedSearchCV for LSSVR and Keras Tuner for LSTM. Model performance was evaluated using MSE, RMSE, MAE, and Nash-Sutcliffe Efficiency (NSE). The results showed that the SSA-LSTM model consistently outperformed SSA-LSSVR model, with the highest average NSE value being 0.9 across six monitoring locations in Semarang City. Furthermore, the predicted rainfall values were spatially visualized using Inverse Distance Weighting (IDW) interpolation within a Geographic Information System (GIS) environment, producing informative rainfall distribution maps that support early warning systems and disaster mitigation efforts. In conclusion, the hybrid approach combining SSA-NMF preprocessing with LSTM-based deep learning significantly improves the accuracy and reliability of daily rainfall forecasting. This novel SSA‑NMF + LSSVR/LSTM framework delivers high‑resolution, reliable rainfall forecasts that directly empower disaster risk reduction systems and readily transfer to similar climatic regions.

Spice Type Recognition Based on Shape and Color Features Using K-Nearest Neighbor and Fuzzy Methods

2025-05-26T04:18:16+00:00

Spices are natural ingredients that play an important role in everyday life, especially in traditional medicine. With a variety of shapes and colors, spices are often difficult to distinguish from one another. This research aims to classify spice types based on shape and color features using K-Nearest Neighbor (K-NN) and Fuzzy methods. This research will limit the recognition of spice types to 10 specific types of spices, namely ginger, turmeric, star anise, coriander, pepper, nutmeg, galangal, cinnamon, cloves, and candlenut. Spice type recognition will be done based on shape, color and texture features extracted using 300 training data images. The application of the K-NN method and Fuzzy logic allows flexible processing of color features (HSV). Fuzzy logic classifies spice color characteristics by generating a color score (color_score), which is then used to better interpret and distinguish spice colors for the classification process between test data and training data by the K-NN method. The test results show that from a total of 100 test data, the system successfully classifies spices with an accuracy rate of 77%.

Development of a Distributed Gradient Boosting Forest Algorithm with Residual Connections in Data Classification

2025-06-18T22:47:01+00:00

The growing complexity and volume of data across various domains necessitate machine learning models that are scalable and robust for large-scale classification tasks. Ensemble methods such as Gradient Boosting Decision Trees (GBDT) demonstrate effectiveness; however, they encounter issues concerning scalability and training stability when applied to very deep architectures. This work presents a novel enhancement using residual connections derived from deep neural networks into the Distributed Gradient Boosting Forest (DGBF) algorithm. By enabling direct gradient propagation across layers, residual connections solve the vanishing gradient problem and so improve gradient flow, accelerate convergence, and stabilise the training process. The Residual DGBF model was assessed using seven distinct datasets across the domains of cybersecurity, financial fraud, phishing, and malware detection. The Residual DGBF consistently surpassed the baseline DGBF in terms of accuracy, precision, recall, and F1-score across all datasets. Particularly in datasets marked by imbalanced classes and complex feature interactions, this suggests improved generalisation and higher predictive accuracy. By proving more stable and strong gradients across the depth of the model, layer-wise gradient magnitude analysis supports these improvements and so confirms the effectiveness of residual connections in deep ensemble learning. This work improves ensemble techniques by combining the scalability and interpretability of decision tree ensembles with the residual architecture optimising benefits. The proposed Residual DGBF enables future research on enhanced deep boosting frameworks by offering a strong and scalable method to address challenging real-world classification tasks.

Comparative Study of BiLSTM and GRU for Sentiment Analysis on Indonesian E-Commerce Product Reviews Using Deep Sequential Modeling

2025-06-18T12:22:11+00:00

Sentiment analysis plays a crucial role in understanding customer perspectives, especially within Indonesian e-commerce platforms. Despite the success of deep learning in high-resource languages, its application to Indonesian sentiment data remains underexplored. Previous studies using models like BERT-CNN or fine-tuned IndoBERT achieved modest results, highlighting the need for more effective architectures for Indonesian language. This study aims to investigate the effectiveness of Bidirectional Long Short-Term Memory (BiLSTM) and Gated Recurrent Unit (GRU) models in classifying buyers’ sentiment from Indonesian product reviews on the PREDECT-ID dataset comprising 5,400 annotated product reviews. Standard NLP preprocessing techniques—including text normalization, tokenization, stopword removal, and stemming—were applied. Both models were trained using Adam and Stochastic Gradient Descent (SGD) optimizers, and their performance was evaluated using accuracy, precision, recall, and F1-score metrics. The GRU model trained with SGD achieved the highest performance, with an accuracy of 94.07%, precision of 93.84%, recall of 94.53%, and F1-score of 94.18%. Notably, the BiLSTM model combined with SGD resulted in competitive results, achieving 93.61% accuracy and 93.84% F1-score. The results confirm that GRU with SGD optimizer, are highly effective for sentiment classification in Indonesian language datasets. By leveraging deep sequential modeling for a low-resource language, this study contributes to the advancement of scalable sentiment analysis systems in underrepresented linguistic domains. The results contribute to the advancement of NLP systems for Indonesian by providing a benchmark for the future development of sentiment analysis tools in low-resource languages.

A Random Forest and SMOTE-Based Machine Learning Model for Predicting Recurrence in Papillary Thyroid Carcinoma

2025-06-18T12:32:29+00:00

PTC (Papillary Thyroid Carcinoma) is one subtype of thyroid cancer occurred most frequently in thyroid cancer cases. Although the prognosis of this cancer is typically positive, its recurrence remains a key challenge requiring early detection. This study proposes machine learning models to predict PTC recurrence, explicitly addressing the inherent class imbalance in the recurrence data. This study implemented three supervised learning algorithms, namely Random Forest (RF), Extreme Gradient Boost (XGB), and Support Vector Machine (SVM) with the Synthetic Minority Oversampling Technique (SMOTE) to balance the dataset. SMOTE was chosen for its capacity to generate synthetic minority class samples while minimizing information loss, thus effectively addressing class imbalance and improving classification outcomes. Model performance was assessed using accuracy, precision, recall (sensitivity), and F1-score. Among all approaches tested, RF with SMOTE demonstrated superior performance, achieving 0.98 accuracy, perfect precision (1.0), high recall (sensitivity) (0.95), and a strong F1-score (0.97), outperforming previous methods including SMOTEENN-based approaches. The result of this study demonstrates SMOTE specifically outperforms SMOTEENN in this clinical context, likely due to better preservation of subtle prognostic indicators with minimal information loss. This improvement suggests SMOTE's effectiveness in preserving valuable decision boundary information while addressing class imbalance in PTC recurrence prediction. These findings establish RF with SMOTE as a robust and well-balanced approach for predicting PTC recurrence, contributing significantly to the development of more precise and responsive AI-driven decision support tools for thyroid cancer.

Utility-Based Buffer Management for Enhancing DTN Emergency Alert Dissemination in Jakarta's Urban Rail Systems

2025-08-05T06:20:51+00:00

The efficiency of emergency alert dissemination in highly populated and densely urban transport networks, such as Jakarta's integrated rail system, is undermined by sporadic connectivity and limited network resources. In this environment, an initial comparison of baseline Delay-Tolerant Network (DTN) routing protocols revealed that flooding-based routers, such as Epidemic, while achieving above-average delivery rates, suffered from high overhead and poor buffer utilization. This paper fills this gap by proposing the Combined Utility Router, a novel buffer management policy that overcomes the limitations of naive strategies, such as Drop-Oldest. Our approach holistically evaluates a message's value by assigning a weighted utility function based on its Time-To-Live (TTL), estimated total replicas, message size, and a user-defined priority. The router maintains high-value messages by discarding the message deemed the lowest utility score under the buffer constraint. Utility-based simulations in The ONE simulator demonstrate that applying our approach to Epidemic routing improves delivery probability, reduces average latency in high network congestion scenarios, while maintaining overhead rates. This work confirms that, in the context of developing reliable and efficient emergency communication systems for challenging urban topographies, optimizing buffer management extends beyond simply selecting the appropriate protocol.

Performance Evaluation of Backend Frameworks for REST API: A Comparative Study of Spring Boot, Flask, Express.js, Laravel FrankenPHP, and Gin

2025-06-03T22:52:00+00:00

One major impact of this development is the shift in application development, particularly in data integration across different platforms. Web services have emerged as a solution for system integration and multi-platform application development. One implementation of Web services is Representational State Transfer. The choice of programming language and framework is also crucial in web application development, directly affecting performance and efficiency. Research on framework performance is necessary to support the development of an Academic Information System. This study will use parameters such as response time, throughput, and resource usage, employing a performance testing method modified by the author. The method includes problem identification, data collection, backend development, performance testing, and conclusion. The test results show that Spring Boot outperforms others in all parameters with stable and efficient performance. Gin is suitable for medium-scale data, Flask excels in scalability but lacks stability, Express.js is efficient CPU usage, and Laravel with FrankenPHP is Memory-efficient. These results serve as a reference for selecting frameworks according to REST API development needs. This research supports developers in selecting appropriate backend frameworks for high-performance REST API systems.

Herbal Plant Classification Using EfficientNetV2B0 Model and CRISP-DM Approach

2025-07-19T05:49:55+00:00

Herbal remedies have long been utilized by Indonesian communities as part of traditional medicine. However, identification of these natural resources is often challenging due to the morphological similarities among various species, which demand expert knowledge to differentiate. This study aims to implement the EfficientNetV2B0 model architecture for classifying medicinal leaves through an Android-based application designed to support recognition tasks. The dataset was composed of augmented images of plant foliage. The model was trained using the TensorFlow framework and evaluated to measure classification performance. Results demonstrate that EfficientNetV2B0 achieves excellent accuracy, with validation scores exceeding 97%, outperforming several other deep learning models. The resulting application allows the general public to identify local medicinal species more easily. This study contributes to the field of computer vision by providing an accurate and efficient classification framework, particularly beneficial for health-related informatics in biodiversity-rich regions.

Development of WebGIS for Street Light Mapping Using Geospatial Tools

2025-06-11T07:07:26+00:00

Padang City, as one of the cities the largest on the west coast of Sumatra Island, plays a strategic role in the economy and government. One of the vital infrastructures that supports public activities is the street lighting system. However, the monitoring and maintenance of streetlights still face obstacles, especially in North Padang District, which is the busiest area due to the presence of numerous educational facilities, government offices, and economic centers. This research aims to develop a WebGIS application that facilitates the monitoring and management of street lighting more efficiently. Our research contributes by introducing a new approach to spatial-based streetlight management strategies. This approach is based on a methodology for field data collection and spatial database development to manage all stages of streetlight infrastructure management. This application integrates geospatial technology by utilizing GeoServer, QGIS, and PostgreSQL for visualization and spatial data management. With this system, information about the location and condition of streetlights can be accessed in real-time, thereby facilitating better planning and maintenance of street lighting infrastructure. The result of this study is a WebGIS application capable of mapping and monitoring streetlight points interactively. The implementation of this system is expected to assist relevant authorities in improving the effectiveness of street lighting management in Padang City and contribute to the development of geospatial technology-based solutions for urban infrastructure.

Comparison of ANOVA and Chi-Square Feature Selection Methods to Improve Machine Learning Performance in Anemia Classification

2025-07-16T00:07:17+00:00

Anemia is a prevalent hematological condition marked by decreased hemoglobin concentration in the blood, which can lead to serious health complications if undetected. Although machine learning has shown potential in supporting early diagnosis, its effectiveness is often hindered by irrelevant or excessive features. This study investigates the impact of ANOVA and Chi-Square feature selection methods in improving the effectiveness of three distinct machine learning models algorithms, Naive Bayes, K-Nearest Neighbor (KNN), and Support Vector Machine (SVM) for anemia classification. Using a Kaggle dataset consisting of 15,300 instances and 25 features, the evaluation of each model was conducted with reference to its accuracy, precision, recall, and F1-score, both before and after applying feature selection. Experimental results show a substantial improvement in classification performance after feature selection, with the SVM + ANOVA combination achieving the highest accuracy of 94.61%. In contrast, models without feature selection performed below 90%, highlighting the need for appropriate feature reduction techniques. This study contributes a comparative analysis framework for medical data classification, emphasizing the role of statistical feature selection in optimizing model accuracy. Its novelty lies in demonstrating consistent performance improvement across algorithms using real-world anemia data and providing evidence that ANOVA and Chi-Square can significantly enhance model generalization in medical diagnostic contexts.

Evaluation IT Goverment Capabilities of the Facematch RIM Polri Recruitment System Using the COBIT 5 Framework

2025-05-24T05:33:17+00:00

One of the technology implementations used by the Government, especially the latest POLRI in the Facematch RIM POLRI system in the recruitment of POLRI new members, has been designed to prevent fraud in the Police recruitment process by recording the faces of prospective members as a valid identity. To determine the capabilities of this system, qualitative data collection was carried out through interviews with related parties and observation of overall system governance activities. The operational implementation of this system has several findings, including the recording process and image quality monitoring mechanism, the continuity of the capture process, to network and infrastructure constraints. The findings are mapped within the COBIT 5 framework domain to determine the gap for improvement in Acceptance System Governance based on Facematch RIM POLRI at POLRI. These findings can contribute to the improvement of IT Governance practices in the POLRI Admissions System in Government in line with the COBIT 5 framework domains and are expected to provide strategic recommendations to overcome the challenges faced and improve the efficiency and effectiveness of the system in supporting organizational goals.

Visual Interpretation of Machine Learning Models (Random Forest) for Lung Cancer Risk Classification Using Explainable Artificial Intelligence (SHAP & LIME)

2025-07-04T01:37:45+00:00

Lung cancer remains one of the most prevalent and burdensome cancers worldwide, with delayed diagnosis being a persistent challenge—particularly in Indonesia, where no national screening program currently exists. In this collaborative study, we aim to develop an interpretable machine learning model for classifying lung cancer risk levels using the Explainable Artificial Intelligence (XAI) approach. The CRISP-DM framework was applied, and the dataset underwent cleaning, feature selection, labeling, and transformation, resulting in 152 valid entries. Tree ensemble algorithms—XGBoost, Random Forest, and LightGBM—were used, with Random Forest achieving the best performance at 97.38% accuracy. SHAP and LIME methods were integrated to provide transparent visual interpretations. A web-based system was developed using Streamlit, incorporating these visualizations and automated narrative summaries generated by a language model to assist non-technical users. A simulated case based on a published pediatric lung cancer report was used to demonstrate its interpretability and illustrate its potential applicability in clinical workflows. The proposed system offers an interpretable and scalable solution for early lung cancer risk classification, which may enhance decision support in primary care and promote trust in AI-assisted diagnostics.

Comparative Analysis of ArUco Marker Detection Techniques Using Adaptive Thresholding, CLAHE, and Kalman Filter for Smart Cane Applications

2025-06-15T14:22:59+00:00

This study aims to analyze and compare the effectiveness of three image processing techniques Adaptive Thresholding, CLAHE, and Kalman Filter in enhancing the performance of ArUco marker detection for a smart cane system designed for visually impaired individuals at SLB Kuncup Mas Banyumas. The evaluation method includes detection accuracy, marker position precision, and computational time required by each technique under two different lighting conditions: daytime and nighttime. The results show that all three image processing techniques successfully achieved a 100% detection accuracy for ArUco markers. However, significant differences were observed in computational time, with Kalman Filter demonstrating the fastest processing speed, making it the most efficient option for real-time applications requiring quick response. CLAHE and Adaptive Thresholding performed better in uneven lighting conditions, although they required longer computational times. Kalman Filter is therefore recommended for marker-based navigation systems in environments demanding fast response times, while CLAHE and Adaptive Thresholding are better suited for settings with variable lighting intensities. The implications of these findings open opportunities for developing adaptive navigation systems capable of dynamically adjusting image preprocessing methods based on real-time environmental conditions. This study contributes practically to the advancement of assistive navigation technologies for visually impaired individuals, particularly in the development of visual marker-based detection systems. The results also provide a useful guideline for selecting appropriate image processing techniques according to environmental characteristics, thereby improving the accuracy and adaptability of navigation systems across diverse lighting conditions and operational environments.

Quantitative Analysis of the Key Factors Driving Cybersecurity Awareness Among Information Systems Users

2025-06-18T12:25:25+00:00

Cybersecurity threats are increasingly complex and widespread, posing significant risks to individuals and organizations. However, many studies tend to address the technological or behavioral aspects separately. The study uses a survey-based quantitative approach using PLS-SEM to analyze key factors that influence cybersecurity awareness, including demographics, training, psychological bias, and organizational culture. The findings suggest that several constructs-such as threat awareness, perceived risk, and education-significantly predict cybersecurity awareness and behaviour. Notably, the model yields an R² value of up to 0.703 with a strong path significance (p < 0.05), which underscores the robustness of the relationship. This study offers an integrated perspective on cybersecurity by bridging the psychological, educational, and organizational dimensions. It highlights cybersecurity awareness as a mediating construct that links upstream factors to secure user behavior-a relational structure that has not been explored in previous research.

Performance Comparison of Learned Features from Autoencoder and Shape-Based Hu Moments for Batik Classification

2025-06-05T07:32:10+00:00

Batik classification depends critically on effective feature extraction to capture the unique geometric and visual characteristics of batik patterns. This study compares two distinct feature extraction methods for batik classification: learned features extracted via a convolutional autoencoder, and shape-based handcrafted features derived from Hu Moments. While autoencoders automatically learn complex latent representations that adapt to intricate pattern variations, Hu Moments provide invariant shape descriptors robust to rotation, scaling, and translation. The methodology involves extracting Hu Moment features and autoencoder latent features from the same batik image dataset, followed by evaluation with identical classifiers to ensure a fair comparison. Experimental results reveal key trade-offs: Hu Moments offer robustness and interpretability in capturing shape geometry, whereas autoencoder features better model complex, non-linear patterns. These findings highlight the complementary strengths of classical and learned feature extraction techniques, offering valuable insights for optimizing batik classification. This research advances feature extraction methodologies in cultural heritage image analysis, with broader applicability to pattern-rich domains like batik classification.

Automated Classification of Mungkus Fish Freshness Based on Eye and Gill Images Using the Naive Bayes Algorithm

2025-07-20T00:18:43+00:00

The problem of assessing the freshness of fish, especially Mungkus fish, is usually directed at several physical indicators, such as eye appearance, gill condition, meat quality, and odor. This traditional method is often considered inaccurate and requires certain expertise, therefore a more effective and objective method is needed to assess the freshness level of Mungkus fish, which in turn can provide benefits for both fishermen and the public in general. The solution to this problem by using the Naïve Bayes method in classifying the freshness level of Mungkus fish based on eye and gill images has proven to be a fairly efficient approach. The Naïve Bayes method itself is a simple but very effective algorithm in the field of machine learning, and operates based on Bayes' Theorem with the assumption that features are independent of each other. This method can be applied in the initial stage of classification by utilizing basic features taken from images of fish eyes and gills. Based on testing 30 new data sets, the clustering system demonstrated an accuracy rate of 66.67%, indicating that 20 data sets were correctly classified according to their actual conditions. On the other hand, 10 data sets, or 33.33%, could not be categorized correctly. Of the 30 old data sets tested, the system was able to correctly classify 19 (63.33%), while 11 (36.67%) still had errors in their classification predictions. Overall, the system successfully performed data clustering with 65% accuracy, with the remaining 35% still showing errors in the classification process.

Enhancing Malware Detection in IoT Networks using Ensemble Learning on IoT-23 Dataset

2025-05-28T04:12:00+00:00

The Internet of Things (IoT) has become a technological innovation that brings many benefits in various sectors, but also presents challenges, especially in terms of cybersecurity. One of the main threats is malware, which can damage devices, steal data, and disrupt system performance. With the increasing use of IoT, malware attacks on IoT devices are a serious concern. Previous research shows that malware detection models in IoT devices still have shortcomings, especially in terms of accuracy. One of the algorithms used in malware detection, Naïve Bayes, has been shown to provide low accuracy results. This study aims to improve the accuracy of malware detection on IoT networks by applying Ensemble learning techniques using traffic data from the IoT-23 dataset. The methodology used refers to the CRISP-DM (Cross Industry Standard Process for Data Mining) framework, which includes the stages of domain understanding, data understanding, data preparation, modelling, evaluation, and deployment. The results show that Ensemble learning improved the performance of individual models. Naïve Bayes as a single model produces an accuracy of 0.24, increasing to 0.35 when combined with AdaBoost, and 0.99 when combined with XGBoost. The combination of the three models also produced an accuracy of 0.99. These results demonstrate the effectiveness of ensemble learning in improving malware detection accuracy in IoT environments.

Enhancing BERTopic with Neural Network Clustering for Thematic Analysis of U.S. Presidential Speeches

2025-07-14T00:16:52+00:00

Understanding the underlying themes in presidential speeches is critical for analyzing political discourse and determining public policy direction. However, topic modeling in this context presents difficulties, particularly when clustering semantically rich topics from high-dimensional embeddings. This study seeks to improve topic modeling performance by incorporating a Neural Network Clustering (NNC) approach into the BERTopic pipeline. We analyze 2,747 speeches delivered by U.S President Joe Biden (2021-2025) and compare three clustering techniques: HDBSCAN, KMeans, and the proposed Autoencoder-based NNC. The evaluation metrics (UMass, NPMI, Topic Diversity) show that NNC produces the most coherent and diverse topic clusters (UMass = -0.4548, NPMI = 0.0234, Diversity = 0.3950, ). These findings show that NNC can overcome the limitations of density and centroid-based clustering in high-dimensional semantic spaces. The study contributes to the field of Natural Language Processing by demonstrating how neural-based clustering can improve topic modeling, particularly for complex, real-world political corpora.

Improving Diabetes Prediction Performance Using Random Forest Classifier with Hyperparameter Tuning

2025-06-18T12:51:51+00:00

Diabetes mellitus is a chronic metabolic disorder that poses a serious challenge to global healthcare systems due to its increasing prevalence and the high costs associated with treatment. Although machine learning has been widely adopted to support early diagnosis, many predictive models still underperform due to limited preprocessing strategies and inefficient hyperparameter settings. This study proposes a comprehensive machine learning pipeline to enhance diabetes prediction accuracy by utilizing a Random Forest classifier optimized through systematic hyperparameter tuning. The novelty of this method lies in its integrated approach, which includes thorough preprocessing such as removing duplicate records, handling inconsistent unique values, addressing missing data, and applying the SMOTE technique to overcome class imbalance. Additionally, hyperparameter tuning is conducted using GridSearchCV combined with 5-fold cross-validation, and only the most influential features are selected to improve model interpretability and efficiency. The proposed model achieved an accuracy of 95 percent, with a recall of 0.88 and an F1-score of 0.85, indicating its robustness in identifying diabetic cases more effectively than previous studies using standard machine learning algorithms. This model contributes to the development of a reliable and scalable early detection system for diabetes, applicable in clinical decision support environments. Further refinement can be achieved by testing on larger and more diverse datasets or by implementing more efficient tuning techniques such as Bayesian optimization.

DEGREE: Development and Validation of a User Experience Model for Digital Educational Games Using Cronbach’s Alpha and Fuzzy Logic

2025-06-24T09:13:11+00:00

The rapid growth of digital educational games demands an evaluation model that accurately captures user experience and adopts a human-centred approach. This study introduces DEGREE (Digital Educational Game Review and Evaluation Engine), an enhanced model extending MEEGA+ by incorporating two previously underrepresented dimensions: Control and Feedback. Using a quantitative approach, questionnaires were distributed to high school students who actively use Minecraft and Duolingo, yielding 4800 responses.
Reliability analysis via Cronbach’s Alpha revealed that the Player Experience + Control combination achieved the highest score (α = 0.914), while the inclusion of Feedback reduced reliability (α = 0.864), leading to its exclusion in the final model. The DEGREE model consists of two core domains: Usability (Aesthetics, Learnability, Operability, Accessibility) and Player Experience (Focused Attention, Fun, Challenge, Social Interaction, Confidence, Relevance, Satisfaction, Perceived Learning, User Error Protection, Control). Evaluation scores were calculated using the Fuzzy Weighted Average (FWA) method and Mean of Maximum (MoM) defuzzification. The Control dimension emerged as the most influential (0.2735), followed by Fun (0.2664) and Satisfaction (0.2516), highlighting the significance of user agency in digital learning environments. The DEGREE model offers a statistically robust and user-oriented framework for evaluating educational games, delivering actionable insights for developers and educators to design more effective and engaging digital learning experiences. This study contributes a new validated and generalizable evaluation framework that strengthens the theoretical foundation of user experience assessment in educational game design.

A Study Concentration Selection With a C4.5 Algorithm, KNN, and Naive Beyes

2025-03-03T04:56:54+00:00

The course of concentration is a crucial aspect for students at the university amikom purwokerto.This decision doesn't just affect their academic journey., but also determine their readiness in the face of the working world.Various factors that affect the concentration selection, the challenges that students face, as well as solutions to help them choose concentrations that fit their interests and career goals.There are still many students who have been confused in deciding which courses best fit their interests and career goals..This confusion is often caused by a lack of adequate information and proper guidance. This study attempts to analyze the lecture amikom purwokerto concentration of students in the universities of the use of the method c4.5 algorithm 3, k-neareset naighbors and naïve beyes. Academic student data used as the basis analysis to determine the dominance in the lecture concentration.Of the result of the research uses phon 60,24 % decision is, there are using k-neareset naighbors 75.36 % and use naïve beyes 100,00 % there are, the prediction could be the basis for deciding the lecture the concentration by mainstream student.The result is expected to help the university in recommended it to students study concentration related to the election.

A Comprehensive Benchmarking Pipeline for Transformer-Based Sentiment Analysis using Cross-Validated Metrics

2025-06-19T07:22:58+00:00

Transformer-based models have significantly advanced sentiment analysis in natural language processing. However, many existing studies still lack robust, cross-validated evaluations and comprehensive performance reporting. This study proposes an integrated benchmarking pipeline for sentiment classification on the IMDb dataset using BERT, RoBERTa, and DistilBERT. The methodology includes systematic preprocessing, stratified 5-fold cross-validation, and aggregate evaluation through confusion matrices, ROC and precision-recall (PR) curves, and multi-metric classification reports. Experimental results demonstrate that all models achieve high accuracy, precision, recall, and F1-score, with RoBERTa leading overall (94.1% mean accuracy and F1), followed by BERT (92.8%) and DistilBERT (92.1%). All models exceed 0.97 in ROC-AUC and PR-AUC, confirming strong discriminative capability. Compared to prior approaches, this pipeline enhances result robustness, interpretability, and reproducibility. The provided results and open-source code offer a reliable reference for future research and practical deployment. This study is limited to the IMDb dataset in English, suggesting future work on multilingual, cross-domain, and explainable AI integration.

Sentiment Analysis of Fizzo Novel Application Using Support Vector Machine and Naïve Bayes Algorithm with SEMMA Framework

2025-07-20T04:32:49+00:00

The increasing popularity of digital reading platforms in Indonesia, such as Fizzo Novel, has generated many user reviews that can be analyzed to understand their satisfaction. This study analyzes user sentiment toward Fizzo Novel using the SEMMA (Sample, Explore, Modify, Model, Assess) framework, and compares the performance of the Support Vector Machine (SVM) and Naïve Bayes algorithms. A total of 139,759 reviews were collected from the Google Play Store through web scraping. The data was then processed through normalization, tokenization, lexicon-based sentiment labeling, and feature extraction using TF-IDF. To address class imbalance, the SMOTE technique was applied. The results showed that SVM achieved the highest accuracy, exceeding 96%, with a consistent F1-score across all sentiment classes. In contrast, Naïve Bayes recorded lower accuracy (75.82% before SMOTE and 73.63% after SMOTE), along with a decline in performance for the neutral class. SVM proved more reliable in handling large and imbalanced text data. Practically, the results of this study can help application developers such as Fizzo Novel in automatically understanding user opinions. With an accurate sentiment classification model, developers can monitor reviews in real-time, identify issues such as excessive advertising or an unpopular chapter division system, and design feature improvements based on real user needs. This research also provides a foundation for algorithm selection in future large-scale sentiment analysis projects and recommends SVM as the more appropriate choice in this context.

Prediction of Turbidity Removal Time in Electrocoagulation Wastewater Using Random Forest, XGBoost, and Others: A Data-Driven Information System Approach

2025-06-15T14:25:41+00:00

Electrocoagulation is an effective and environmentally friendly technology for treating wastewater by removing contaminants such as turbidity, heavy metals, and organic compounds. Accurately predicting turbidity removal time is essential for optimizing treatment performance and operational efficiency. However, this is challenging due to complex, nonlinear relationships between multiple parameters including current, voltage, electrode configuration, conductivity, and turbidity removal rate. This study aims to develop a predictive framework by comparing six supervised regression models, namely Linear Regression, Polynomial Regression, Random Forest, Support Vector Regression (SVR), XGBoost, and Long Short-Term Memory (LSTM), using key electrocoagulation parameters. After extensive data preprocessing, a dataset of 281 samples was used for training and validation. Among them, Random Forest achieved the best performance (R² = 0.876, RMSE = 601.15). A data-driven information system is proposed to integrate these predictive capabilities for real-time monitoring and control. By improving turbidity prediction accuracy, the system enables the sustainable utilization of water as a valuable asset, even in its wastewater form. The approach enhances decision-making by providing intelligent feedback for process optimization. This research contributes to the advancement of intelligent, sustainable wastewater treatment systems by integrating machine learning prediction models with practical process control applications in informatics.

From Monoliths to Microservices: Designing a Scalable Super App Architecture for Academic Services at Universitas Jenderal Soedirman

2025-08-05T21:57:20+00:00

Jenderal Soedirman (Unsoed) currently operates more than 30 monolithic information systems built with heterogeneous technology stacks, resulting in duplicate functionality, inconsistent user experience, and high maintenance costs. This study designs a modular, microservices‑based Super App architecture that integrates core academic services (KRS/KHS, transcript, student & lecturer attendance, lecturer activity log) and a parent/guardian monitoring feature. Using the Design Science Research (DSR) method, we (1) identified problems via a technology audit and problem–objective matrix; (2) designed the artifact with Domain‑Driven Design, C4 modelling, and API‑first contracts; (3) demonstrated a working prototype with API Gateway, SSO, and event‑driven notifications; (4) evaluated performance (<300 ms latency for 500–1000 concurrent users) and stakeholder impact; and (5) communicated results through this paper. The proposed architecture reduces integration complexity, supports zero‑downtime deployment, and enhances transparency for parents without violating consent and privacy. The validated blueprint provides a roadmap for transforming legacy campus systems into a scalable, observable, and governable Super App.

Interpretable Machine Learning for Employee Recruitment Prediction Using Boruta, CatBoost, Lasso, Logistic Regression, NLP, and RFE Feature Selection

2025-06-04T06:00:34+00:00

Employee recruitment is one of the crucial processes in human resource management that has a direct impact on the performance and success of the company. In the digital era, the use of Machine Learning (ML) in candidate selection processes is increasingly prevalent due to its ability to enhance efficiency, accuracy, and transparency. This research is important because conventional recruitment methods often face issues such as subjective bias, slow processing times, and limitations in assessing a candidate’s true potential. ML offers a more objective, data-driven, and faster approach, enabling companies to identify the best candidates more effectively. This study aims to identify the main features that influence recruitment decisions, as well as evaluate the effectiveness and interpretability of several ML models, namely Boruta, CatBoost, Lasso Regression, Logistic Regression, Natural Language Processing (NLP), and Recursive Feature Elimination (RFE). This study uses a dataset consisting of 1,501 samples with 10 features and one class variable (0 = Not Hired, 1 = Hired). The evaluation is carried out based on the ability of each model to identify the features that make the most significant contribution to the classification results. This study has several limitations, particularly the potential bias in the data, such as demographic bias that may be reflected in historical recruitment decisions. This could lead the ML models to replicate or even reinforce such biases. Additionally, the limited dataset size may affect the models' ability to generalize to new data. In the context of this study, the main parameter used to assess the superiority of the model is the most dominant feature or the highest feature produced by each method. The test results show that the Boruta model identifies Gender as the most influential feature, while the CatBoost, Lasso Regression, Logistic Regression, and NLP models consistently place Recruitment Strategy as the most significant feature in predicting candidate eligibility. Meanwhile, the RFE model produces Distance from the Company as the highest feature that influences recruitment decisions. The uniqueness of this study lies in its approach that integrates feature interpretability models within the real-world context of recruitment decision-making. This approach not only emphasizes prediction accuracy but also promotes transparency and a clear understanding of the rationale behind each decision. It supports the development of a fairer and more accountable selection process, particularly by minimizing unconscious bias in data-driven recruitment systems. From a practical standpoint, the findings are highly relevant for human resource professionals, as the identified key features can be used to design more objective selection strategies and enhance the efficiency of candidate evaluations. Therefore, this study makes a tangible contribution to the advancement of modern, technology-based recruitment systems that prioritize fairness and decision-making efficiency. Additionally, the selection of evaluation metrics could be further elaborated to strengthen the analysis, for example by presenting the overall accuracy of each model or comparing them with alternative approaches to provide a more comprehensive view of the models' performance.

Classification Of Sea Wave Heights On The North Coast Of Central Java Using Random Forest

2025-07-17T07:01:23+00:00

Global climate change has triggered an increase in the occurrence of significant wave heights (SWH) and sea level rise (SLR) in coastal areas, including the northern coast of Central Java, Indonesia (Pantura). These phenomena directly impact maritime activities, coastal erosion, and tidal flooding. This study aims to classify and predict significant wave height (SWH) and sea level rise (SLR) trends using a machine learning approach based on the Random Forest (RF) algorithm. Daily meteorological and oceanographic observation data from 2019 to 2024, provided by BMKG, serve as the main dataset. The dataset includes wind speed, ocean current velocity, air pressure, and wave direction. SWH is categorized into three classes: Calm, Low, and Moderate. The classification model achieved excellent performance with an accuracy of 98.54%, a macro F1-score of 0.942, and maintained strong accuracy even for the minority class (Moderate) despite data imbalance. The RF Regressor for SWH prediction yielded an R² of 0.864, MAE of 0.067, and RMSE of 0.109 m. Visualizations such as scatter plots, boxplots, and heatmaps supported the conclusion that ocean current speed and wave period are key factors influencing SWH. The study concludes that Random Forest is effective for classifying and predicting sea conditions in tropical regions like Pantura, and it is feasible for implementation in data-driven early warning systems to mitigate coastal risks. This contributes to marine safety and coastal risk mitigation planning.

Comparison of Time Series Algorithms Using SARIMA and Prophet in Predicting Short-Term Bitcoin Prices

2025-05-26T02:19:09+00:00

Digital finance, particularly Bitcoin, has become a global phenomenon with high volatility, posing great challenges for traders in predicting short-term prices. This study compares the performance of the SARIMA and Prophet algorithms in predicting short-term Bitcoin prices using daily closing price data from October 1, 2014, to October 1, 2024. The study utilizes two different data timeframes, a 10-year dataset (2014-2024) and the last 5 years (2019-2024) for comparative analysis. The SEMMA methodology is used to analyze and compare the two algorithms, which consist of the stages Sample, Explore, Modify, Model, and Assess. The experimental results show that SARIMA provides more stable and consistent results with an MAPE value of 1.24% and RMSE of 896.15 in Scenario 1 and an MAPE value of 1.27% and RMSE of 920.24 in Scenario 2. In contrast, Prophet shows different performance in each scenario. In Scenario 1, Prophet shows optimal results but not so good with an average MAPE of 1.74% and an RMSE value of 1214.86. On the other hand, Prophet showed good performance in Scenario 2 with a lower average MAPE of 0.71% and a smaller RMSE of 489.94, indicating Prophet's ability to handle newer and more dynamic datasets. Both models show their respective advantages; SARIMA is better for long and stable historical data, while Prophet is more effective for shorter and dynamic data. This research provides practical insights for traders and investors in choosing the right prediction model, with results for further study in predicting crypto asset prices.

Comparison of LightGBM With XGBoost Algorithms in Determining Arrhythmia Classification in Students

2025-07-15T14:58:34+00:00

Arrhythmia is a heart rhythm disorder that may occur unpredictably with life-threatening risk if it were not treated immediately. This heart disorder generally affects the elderly, but symptoms of this disorder can also arise in children and adolescents, especially for those with heart problems or are often under stress. The implementation of this research is aimed at analyzing the symptoms of early arrhythmia in adolescent children using electrocardiogram signals. In order to obtain the best possible results in determining the higher performing algorithm, two machine learning methods were used to predict the classification of arrhythmia which will be compared for their accuracy. The subjects of this study included 106 students from SMK Swasta Teladan Sumatera Utara 2 located in the city of Medan, of which 72 final subject data were used to train the capability of both models used to predict arrhythmia classification categorized into four categories, namely normal, abnormal, potential of arrhythmia, and high potential of arrhythmia. The LightGBM model outperformed the XGBoost model, with 95.11% accuracy and 95.03% F1 Score, and although the loss value of the LightGBM model is higher than the loss value of the XGBoost model, the difference between these two values is negligible and the loss value of LightGBM can be considered as excellent with a value of 0.1503. This research contributes to the advancement of digital health by demonstrating the potential of machine learning-based ECG analysis for highly accurate early arrhythmia detection in adolescent, non-clinical populations.

Analyzing ChatGPT’s Impact on Graduates’ Communication, Collaboration, and Logical Thinking Skills Using an Extended Technology Acceptance Model

2025-06-19T10:05:41+00:00

The rapid rise of ChatGPT in Indonesia—now the third-highest user base worldwide—raises questions about its impact on essential soft skills for new graduates. Recent evidence warns that while ChatGPT supports academic and professional tasks, it may also reduce critical thinking, collaboration, and communication if not properly guided. This study aims to evaluate how ChatGPT usage affects communication, collaboration, and logical thinking skills among recent graduates in Jabodetabek. A cross-sectional survey of 384 respondents was conducted, and data were analyzed using Partial Least Squares Structural Equation Modeling (PLS-SEM). The modified Technology Acceptance Model (TAM) demonstrated strong explanatory power, with R² values of 0.830 for Behavioral Intention, 0.699 for Actual Use, and 0.651 for Attitude Toward Use. Hypothesis testing confirmed significant effects, including Perceived Ease of Use on Perceived Usefulness (β = 0.946; t = 172.023; p < 0.001) and Behavioral Intention on Actual Use (β = 0.836; t = 50.416; p < 0.001). Positive attitudes toward ChatGPT were strongly associated with enhanced teamwork, communication, and logical reasoning. This study contributes to the discourse on digital literacy and educational technology in Southeast Asia, demonstrating that ChatGPT can strengthen graduate employability when integrated with proper guidance and ethical use. The findings provide practical implications for computer science and education fields, offering a framework for balancing AI adoption with the preservation of critical human skills.

Mosquito Species Classification Using Wingbeat Acoustic Signals Based on Bidirectional Long Short-Term Memory

2025-06-24T08:15:02+00:00

The increasing prevalence of mosquito-borne diseases such as Dengue, chikungunya, and malaria underscores the urgent need for effective mosquito vector monitoring. This study proposes a non-invasive classification system of mosquito species based on wingbeat acoustic signals using a deep learning approach with Bidirectional Long Short-Term Memory (BiLSTM). The audio dataset was collected from the Wingbeats repository, consisting of six major mosquito species. Preprocessing was performed using Discrete Wavelet Transform (DWT) to reduce noise. Feature extraction combined Linear Predictive Coding (LPC) and Mel-Spectrogram to represent spectral and temporal signal characteristics. Each binary model was trained in a one-vs-rest scheme to recognize a target species against others, and a BaggingClassifier was used to fuse predictions from six BiLSTM models. Evaluation showed that the proposed system achieved a final accuracy of 96.85% and F1-score of 95.03%, with confusion matrices showing near-diagonal performance. The results indicate that the hybrid LPC-Mel features and ensemble BiLSTM architecture are effective for mosquito species classification using acoustic signals.

Palm Oil Seed Origin Classification Based on Thermal Images and Agricultural Data Using Convolutional Neural Network

2025-07-04T01:27:33+00:00

The traceability of palm oil seed origins plays a vital role in ensuring transparency, legality, and sustainability across the palm oil supply chain. Recent advances in deep learning have created new opportunities to improve classification systems by leveraging both visual and contextual data. This study proposes a deep learning-based model for classifying the origin of palm oil seeds by integrating thermal imagery with agricultural data. Two convolutional neural network (CNN) architectures, ResNet50 and MobileNet, were evaluated under three experimental setups: using only thermal images, combining thermal images with agricultural features (socio-economic, soil, and spectral fruit characteristics), and applying hyperparameter tuning to the best-performing model. The results show that ResNet50 consistently outperformed MobileNet, particularly in multimodal configurations. The highest performance was achieved using ResNet50 with the Adam optimizer, a learning rate of 0.001, and a batch size of 16, resulting in training accuracy of 99.75%, validation accuracy of 99.92%, and test accuracy of 100.00%. Evaluation metrics confirmed the model’s robustness with precision, recall, and F1-score all reaching 100.00%. This research highlights the significant potential of combining thermal imagery and agricultural data in CNN-based models for accurate and reliable classification of palm oil seed origins. The approach can support traceability systems in the palm oil industry, offering a scalable and data-driven solution for ensuring supply chain integrity and sustainability.

Evaluating Software Quality in a Point of Sales System in a Fast-Food Restaurant Using the McCall Model

2025-06-16T02:05:18+00:00

Software quality is an critical aspect in ensuring system performance and user satisfaction. This study evaluates the quality of the system called Sampos. is a system used by internal employees in managing fast food business operations for record transactions. manage raw material stocks and help track daily reports. The evaluation was conducted using the McCall model, which focuses on five primary quality factors: correctness, reliability, efficiency, integrity, and usability. Each factor is assessed through indicators that reflect the system's performance in that aspect. The measurement stage begins by assigning weights to each indicator based on its level of importance. Then. The quality value of each factor is calculated to get a comprehensive picture of system performance. The results of the evaluation showed that the correctness value was 56.2%, reliability 56%, Integrity 47.8%, and usability 46%, which are generally classified as "Pretty Good.". Meanwhile, the value of the efficiency factor is only 38.2%, so it is categorized as "not good." Overall, the Sampos system obtained an average score of 41% - 60%. This indicates that the system requires improvement, especially in the aspect of efficiency. This study contributes to proving that McCall's method can be used to evaluate applications built without documentation and by a single developer. Therefore, this study contributes a practical case study on the application of McCall’s Model as an effective method for identifying and quantifying quality weakness in small-scale operational systems.

Implementation of K-Means on Packaged Coffee Sales Data for Simulating Goods Entry in Sole Proprietorship Businesses

2025-08-06T21:42:55+00:00

In retail businesses operating under the sole proprietorship structure, decision-making regarding partnerships with beverage distributors—especially those offering packaged coffee—remains a challenge. Store owners often face uncertainty about the profitability of accepting product offerings, which can lead to suboptimal inventory decisions. This study addresses that issue by simulating goods entry scenarios and applying clustering techniques to historical packaged coffee sales data, enabling data-driven insights into product performance and distributor value. Studies focusing on clustering within retail include segmenting customer behaviour and stock management strategies, yet many lacked specific application to single owner businesses and product-centric simulations. This research is novel in its contextual focus on packaged coffee distribution within sole proprietorship environments, integrating real sales metrics and clustering algorithms to empower store owners with actionable evaluation tools. Results demonstrate that clustering reveals patterns of profitable product categories and distributor consistency, offering scalable insights for micro-retail optimization. The findings provide a framework that differs from prior studies by emphasizing the intersection between small business dynamics and algorithmic decision support. Ultimately, this research contributes to the advancement of informatics by demonstrating how clustering-based simulations can enhance decision-making in micro-retail environments through practical, data-driven methodologies.

Security and Performance Evaluation of PPTP-Based VPN with AES Encryption in Enterprise Network Environments

2025-06-04T22:39:24+00:00

In the context of the current digital era, Virtual Private Networks (VPNs) serve a critical function in ensuring the confidentiality and integrity of data transmitted across public networks, particularly within corporate environments. This study presents a comprehensive analysis of VPN security and performance, with a specific focus on the Point-to-Point Tunneling Protocol (PPTP) and the implementation of encryption algorithms such as AES-128 and AES-256. Despite the widespread adoption of PPTP due to its simplicity and broad compatibility, it exhibits significant security vulnerabilities, primarily stemming from its reliance on the outdated RC4-based Microsoft Point-to-Point Encryption (MPPE) and the susceptible MS-CHAP authentication protocol, which is highly vulnerable to brute-force and dictionary attacks. Empirical findings indicate that, although AES-128 and AES-256 introduce minor performance trade-offs compared to unencrypted configurations, AES-256 demonstrates markedly enhanced security, achieving a 98.9% authentication success rate and a threat detection time of 122 milliseconds. Nevertheless, increased user load adversely impacts network performance, with throughput declining from 95 Mbps to 40 Mbps as the user count rises from 5 to 50, accompanied by elevated latency and packet loss. Comparative analysis across three encryption scenarios AES-128, AES-256, and MPPE-PPTP reveals a consistent degradation in network performance as user load increases, with AES-256 offering the strongest security at the cost of slightly reduced throughput and increased latency under high-load conditions. MPPE-PPTP, while providing better throughput, lacks adequate security, making it unsuitable for high-risk environments. Based on these observations, this study recommends the implementation of AES-256 encryption in enterprise networks requiring high security, supported by continuous performance monitoring and strategic capacity planning. Furthermore, the adoption of a secure site-to-site VPN architecture is proposed to facilitate reliable and secure communication between geographically distributed office locations.

A User-Driven E-Audit System for Improving Transparency and Efficiency in Regional Government Supervision

2025-07-21T23:07:46+00:00

Internal audit processes in regional government institutions often face challenges such as time inefficiency, low transparency, and poorly digitized documentation. This study aims to develop an E-Audit system to enhance the effectiveness of internal supervision in a regional inspectorate environment. Employing a user-centered design approach and a structured system development methodology, this research involved key roles—auditors, technical controllers, and follow-up teams—throughout the design and testing stages. The developed system integrates three core phases of the audit process—planning, reporting, and follow-up—into a single, modular, and interactive digital platform. Implementation results indicate a significant improvement in audit efficiency, with a reduction of more than 50% in process duration compared to manual methods. The system also enhances documentation consistency through digital audit trails, role-based dashboards, and automatic reporting features. User acceptance testing revealed a high level of satisfaction, with users highlighting the system’s ease of use, increased accuracy, and alignment with daily audit tasks. Additionally, user feedback emphasized the need for integrated notification features and inter-unit communication tools, indicating readiness for more advanced digital transformation. Overall, this study provides practical value as a model for digital audit implementation at the regional government level while contributing to the advancement of Computer Science through the application of software engineering principles and information systems to support digital government oversight. The developed E-Audit model can serve as a reference for designing real-time collaborative public auditing systems relevant to the development of information systems engineering and computational governance.

Garbage Image Classification Using Deep Learning: A Performance Comparison of InceptionResNetV2 vs ResNet50

2025-05-26T12:19:42+00:00

Garbage problem is a worldwide problem. Efforts to address garbage problem have been performed in several aspect, including automatic garbage classification to support automatic garbage sortation in small scale. In the field of garbage classification, deep learning has been widely used because of its ability to learn feature and also to classify with high accuracy. Several promising architectures in deep learning such as ResNet50 and InceptionNet have been used for this classification task. InceptionResNet is introduced to combine the strength of both architectures. This research aims to classify Garbage Classification data set which consist of 15150 images from 12 classes by using InceptionResNetV2 architecture. In addition, experiment by using ResNet-50 is also performed to provide comparison of its performance. During experiment, Hyperparamater tuning was performed, namely the learning rate, dropout rate, and the number of neuron in the dense layer. The results show that InceptionResNetV2 outperform ResNet50 in all scenarios. This architecture is able to achieve highest accuracy of 97.54%. Even though the classification time is longer for InceptionResNetV2, this finding is able to prove the outstanding performance of InceptionResNetV2 in garbage classification. This study contributes to the field of garbage classification by introducing robust and better model for better classification.

Stacking Ensemble RNN-LSTM Models for Forecasting the IDR/USD Exchange Rate with Nonlinear Volatility

2025-07-15T23:44:21+00:00

Abstract - Predicting exchange rates with high volatility and nonlinear patterns presents a critical challenge in financial analysis. Deep learning models such as RNN and LSTM are widely used for their ability to capture temporal dependencies, yet each has limitations when applied individually. This study aims to enhance the prediction accuracy of the Indonesian Rupiah (IDR) to US Dollar (USD) exchange rate by implementing a stacking ensemble approach that combines RNN and LSTM models. The dataset consists of 522 weekly observations from January 2015 to December 2024, sourced from the official website of Bank Indonesia (bi.go.id). In the proposed framework, RNN and LSTM serve as base learners, while linear regression acts as the meta-learner. Model performance is evaluated using RMSE, MAPE, and MSE. The results indicate that the stacking ensemble consistently outperforms the individual models, achieving an RMSE of 117.91, a MAPE of 0.01, and an MSE of 13,901.67. The model effectively captures historical patterns and delivers stable and accurate predictions. In conclusion, the stacking ensemble approach developed in this study contributes to the advancement of ensemble learning techniques in computer science and offers practical value for financial decision-makers, particularly in managing complex and dynamic exchange rate scenarios.

Improving Extreme Gradient Boosting Model for Heart Disease Prediction Using SMOTE for Class Imbalance

2025-06-29T00:10:11+00:00

The goal of this study is to come up with an intelligent predictive model that can classify the severity of heart disease. The model will employ both XGBoost and oversampling to resolve the problem of data imbalance. In addition, the model will be implemented for real-world application using an interactive interface. The study uses the UCI Heart Disease dataset, which includes many clinical features. Preprocessing involves handling missing values, removal of features with a substantial fraction of missing values, and the use of SMOTE resampling for learning from class-balanced instances. The main classifier that was used for the research purposes was the XGBoost classifier, while the dataset was split 80:20 for training and testing purposes. For ease of individual-level real-time testing of the predictions, the model is implemented through Streamlit. The XGBoost model worked extraordinarily well, with the accuracy standing at 92%, as did precision along with recall, as well as the F1-score, being 92%. These findings clearly outperform other current studies of the same sort that have made use of alternative classifiers. In addition, its deployment using Streamlit makes it even more clinically applicable. Innovation The novelty of the research lies in the combined application of SMOTE with XGBoost, enabling effective classification under imbalanced conditions, along with the real-time implementation using Streamlit for user-level predictions. The model is of high value for early identification and stratification of the severity of heart disease in clinical decision support settings.

Development of an AI Governance Model for Higher Education Using the Capability Maturity Model Integration (CMMI)

2025-07-15T10:19:39+00:00

The increasing adoption of Artificial Intelligence (AI) in higher education presents strategic opportunities for institutional transformation, while introducing complex challenges related to ethics, accountability, transparency, and regulatory compliance. Responding to the growing complexity of AI implementation in academic environments , this study proposes a governance model for AI named GOVAIHEI (Governance of Artificial Intelligence for Higher Education Institutions), conceptualized using the Capability Maturity Model Integration (CMMI) framework. The model was developed using the Design Research Methodology (DRM), which consists of four stages: Research Clarification, Descriptive Study I, Prescriptive Study, and Descriptive Study II. GOVAIHEI encompasses five primary domains: Data and Information, Technology and Infrastructure, Ethics and Social Responsibility, Regulation and Compliance, and Monitoring and Evaluation. Each domain is articulated into capability areas and measurable practices, assessed using the tiered NPLF scale (Not, Partial, Largely, Fully Achieved) to determine institutional capability and maturity levels. The model was validated through expert judgment by three domain specialists, confirming its relevance, methodological soundness, and alignment with CMMI principles. A web-based evaluation system was also developed using Laravel, PostgreSQL, Redis, and Nginx, enabling structured, efficient, and automated assessments. Implementation in a case study at Institute XYZ revealed an initial maturity level (Level 1) with development goals toward Level 3 (Defined). The findings demonstrate a practical foundation for navigating the multifaceted nature of AI adoption in higher education through a structured and adaptable governance approach, which aligns with the increasing demand for robust digital governance frameworks in technology-driven environments.

A Hybrid SOAR-BSC-AHP Framework for Strategy Selection in Digital Cultural Tourism

2025-06-23T14:37:26+00:00

Digital transformation in cultural tourism presents significant challenges, particularly in heritage villages like Kotagede, Yogyakarta. Problems such as limited infrastructure, low digital literacy, and the absence of a structured planning framework hinder progress toward community-based sustainable tourism development. This study addresses these challenges by proposing an integrated decision-making framework that combines SOAR analysis, the Balanced Scorecard (BSC), and the Analytic Hierarchy Process (AHP). The SOAR-BSC framework captures strategic objectives from qualitative data through focus group discussions and stakeholder interviews, while the AHP quantitatively prioritizes eight strategic alternatives based on hierarchical criteria and subcriteria. The most impactful strategies identified were: (1) developing partnerships with tour operators, and (2) promoting community cultural education and training. The Learning and Growth Perspective emerged as the most influential factor (weight = 0.5549), highlighting the importance of community empowerment and digital skills development. Sensitivity analysis and cross-validation using the Simple Additive Weighting (SAW) method confirmed the consistency and robustness of the rankings. In practice, this framework offers a participatory, data-driven guide for digital transformation in heritage tourism, supporting not only improved destination management but also long-term cultural preservation through inclusive digital initiatives.

Real-Time Traffic Density and Anomaly Monitoring Using YOLOv8, OpenCV and Pattern Recognition for Smart City Applications in Demak

2025-06-13T03:37:58+00:00

Urban traffic congestion is a persistent issue in medium-sized cities like Demak, leading to delays and potential accidents. This study presents the development of a real-time vehicle density and anomaly detection system using YOLOv8, combined with OpenCV for video analysis, to monitor traffic flow at strategic entry points of Demak City. The system classifies vehicles into four categories (cars, motorcycles, trucks, buses) and determines their direction by detecting crossing lines. A key feature is the recognition of vehicle patterns, particularly the detection of stopped vehicles, flagging anomalies after 30 seconds of stoppage, with tolerance for temporary detection losses. Traffic data is stored in CSV format, enabling periodic analysis and visualization via an interactive graphical user interface (GUI). Evaluation results show the YOLOv8n model achieves 92.5% precision, 88.3% recall, and 89.7% mean average precision (mAP@0.5), demonstrating improved accuracy and speed over previous YOLO versions. Additionally, the vehicle counting accuracy reaches 94.2% when compared with manual annotations. The proposed system provides a reliable solution for real-time traffic monitoring and early anomaly detection, supporting intelligent transportation systems (ITS) and enabling data-driven traffic management decisions. This research contributes to the advancement of real-time video analytics and pattern recognition for urban traffic control and serves as a scientific reference for the development of smart city infrastructures. Furthermore, this study strengthens the application of pattern recognition in intelligent anomaly detection, providing new insights for researchers in the fields of computer science and informatics.

Ambidextrous AI Governance Model for Advancing State-Owned Bank in Indonesia Digital Transformation Through COBIT 2019 Traditional and DevOps

2025-06-10T06:09:45+00:00

Integrating artificial intelligence into the banking sector accelerates digital transformation, but it also presents governance challenges, particularly in striking a balance between innovation and regulatory compliance, risk management, and operational control. This research proposes an ambidextrous AI governance model by combining two distinct yet complementary mechanisms from COBIT 2019: the structured, control-oriented Traditional framework and the agile, adaptive DevOps Focus Area. This dual approach enables organizations to pursue innovation and maintain governance stability simultaneously. The study investigates BankCo’s, a state-owned bank in Indonesia that is undergoing a systemic digital transformation and applies the Design Science Research (DSR) methodology with a case study approach. Collecting data through five semi-structured interviews with key IT Governance, Risk, and Compliance stakeholders and triangulated with internal policy documents, annual reports, and audit trails. The analysis identified two prioritized Governance and Management Objectives (GMOs), MEA03 (Managed Compliance with External Requirements) and APO12 (Managed Risk), based on design factors, regulatory alignment (POJK No. 11/2022 and SOE Minister Regulation No. PER-2/MBU/03/2023), and agile governance needs. A maturity gap analysis revealed areas for improvement across people, process, and technology dimensions, with the proposed model raising governance capability from 3.55 to 3.95. The proposed model applies multidimensional prioritization through Resource-Risk-Value (RRV) analysis. This study presents a practical and auditable approach to ethical AI governance that strikes a balance between innovation and accountability. The model supports digital transformation in banks and contributes to information systems governance by linking the ethical use of AI with agile yet compliant practices in regulated environments.