Jurnal Teknik Informatika (Jutif)

Analysis of Static and Contextual Word Embeddings in Capsule Network for Sentiment Analysis of The Free Nutritious Meal Program on Twitter

Sun, 15 Feb 2026 00:00:00 +0000

Public discourse surrounding Indonesia’s Makan Bergizi Gratis (MBG) program reflects diverse opinions that have not yet been systematically examined using computational methods. This study addresses that gap by evaluating the effectiveness of static and contextual word embeddings within a Capsule Network (CapsNet) framework for sentiment analysis of MBG-related tweets on Twitter. A total of 7,133 Indonesian-language tweets were collected through web crawling, preprocessed, and manually labeled into positive, neutral, and negative categories. Four embedding techniques—Word2Vec, FastText, ELMo, and IndoBERT—were tested under two preprocessing settings, raw and stemming. The experimental results show that Word2Vec on raw text achieved the highest accuracy of 96.17%, while FastText obtained the best performance on stemmed data with 94.10%. These findings indicate that morphological normalization benefits static and subword-based embeddings, whereas contextual models maintain stable performance without extensive fine-tuning. Overall, this study demonstrates the potential of combining CapsNet with appropriate embedding strategies for Indonesian-language sentiment analysis and provides evidence that natural language processing can support data-driven evaluation of public programs such as MBG.

Insulator Defect Detection Based On Image Processing Using A Modified YOLOv8n Model

Muchamad Arfim Muzaki, Subiyanto, Andika Anantyo — Sun, 15 Feb 2026 00:00:00 +0000

Insulators are critical components in power transmission and distribution systems, where any defects can lead to severe operational failures and power outages. To enhance inspection efficiency, unmanned aerial vehicles (UAVs) are increasingly used for aerial monitoring. However, the quality of images captured by drones is often compromised due to hardware limitations, motion blur, and complex environmental backgrounds, which significantly reduces the performance of deep learning-based defect detection methods. This study proposes an improved insulator defect detection model based on the YOLOv8n architecture, optimized for accuracy and efficiency in low-quality image scenarios and suitable for deployment in resource-constrained environments. The model introduces two major modifications. First, a Slim-Neck module employing Ghost-Shuffle Convolution (GSConv) replaces standard convolutions to substantially reduce computational cost while preserving rich feature representations. Second, an Efficient Multi-Scale Attention (EMA) module is integrated into the neck to enhance multi-scale feature fusion by maintaining per-channel information without dimensionality reduction, improving the model’s ability to extract discriminative features. Experimental results demonstrate that the proposed model achieves a precision of 92.0%, recall of 88.6%, mAP@0.5 of 92.1%, and an inference speed of 161.29 FPS. Furthermore, it reduces parameter count by 10.8% and computational load by 8.6% compared to the baseline, validating its suitability for real-time UAV-based inspections. The model also outperforms existing methods in detecting insulator defects, particularly in challenging conditions involving blur and complex backgrounds.

Multi-Class Real-Time Color Classification of Coffee Beans via Fine-Tuned EfficientNetB0 and Post-Training Quantization

Siti Yuliyanti, Syamsul Maarip — Sun, 15 Feb 2026 00:00:00 +0000

The first problem faced in coffee bean classification is that the manual grading or sorting process still relies heavily on human labor, making it subjective, time-consuming, and prone to errors. Secondly, existing deep learning-based systems often require substantial computing resources, rendering them inefficient for industrial-scale implementation or on limited hardware. The research objective is to develop an efficient, lightweight, and accurate automatic classification model to recognize coffee bean color and support the automation of quality control processes in the coffee post-harvest chain. This study develops an automated system for coffee bean classification based on four color classes: light, medium, green, and dark, utilizing the lightweight EfficientNet model with fine-tuning of smaller versions of EfficientNet (B0–B3). The research stages consist of dataset acquisition, pre-processing, modeling and fine-tuning, as well as model evaluation on the detection system on low-end devices. The main innovation of this research is the efficiency and speed of real-time classification of coffee bean color images using a lightweight CNN model optimized through quantization, which supports field applications with hardware limitations without sacrificing accuracy. Fine-tuning EfficientNetB0 by unfreezing the last 30 layers achieved 97.17% training accuracy and 99.25% validation accuracy with consistent loss reduction, supported by Test-Time Augmentation (TTA) which improves prediction stability to >80% confidence against variations in field image quality. Deployment to TensorFlow Lite (TFLite) with 8-bit quantization resulted in a lighter model that maintained 99.50% accuracy and accelerated inference by up to 6x compared to the original H5 model, and excelled at multi-object detection without sacrificing classification confidence.

Classifying Public Complaints in Denpasar: a Comparative Study of CNN, RNN, LSTM, and Stacking Deep Learning Models

Sun, 15 Feb 2026 00:00:00 +0000

The process of lodging complaints represents a complex behavioral construct, influenced by the interplay of emotional states, sociocultural factors, and situational contexts. It functions as a pivotal channel for citizens to express dissatisfaction regarding the quality of governmental services. This research aims to optimize public complaint management by leveraging deep learning-based text classification on citizen submissions collected from the Denpasar City Complaint Web Portal. The methodological approach integrates several neural network models, including Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Long Short-Term Memory (LSTM) networks, further enhanced by a Stacking ensemble technique that amalgamates the strengths of each architecture. The dataset consists of 10,302 textual records, categorized into four semantic classes: Complaints, Suggestions, Inquiries, and Information. To improve the robustness and reliability of the classification, advanced preprocessing steps were implemented, including the application of the Synthetic Minority Over-sampling Technique (SMOTE) to alleviate class imbalance and the utilization of Term Frequency–Inverse Document Frequency (TF-IDF) for extracting the most informative textual features. Empirical results demonstrate that the Stacking ensemble model significantly outperforms individual baseline models, achieving an accuracy of 77.83%, with recall and F1-score values of 74.38%. These findings highlight the effectiveness of ensemble deep learning approaches in multiclass complaint classification, thereby supporting improvements in public service delivery and fostering greater governmental transparency. Ultimately, this study contributes to the field of automated text classification by illustrating the potential of advanced neural architectures to enhance citizen participation and institutional accountability.

Natural Language Processing (NLP) and Support Vector Machine (SVM) Optimization in Detecting Phishing Website URLs

Sun, 15 Feb 2026 00:00:00 +0000

Phishing remains one of the most pervasive cyber-threats, with recent reports indicating a sharp rise in both volume and sophistication of attacks. According to the Anti‑Phishing Working Group, phishing incidents reached nearly 1 million in Q4 2024. To address this evolving threat, this study aims to develop an automated phishing-URL classification system based on Natural Language Processing (NLP) and Support Vector Machine (SVM). We utilised the Kaggle "PhiUSIIL Phishing URL Dataset" comprising 256,795 URL records and applied comprehensive preprocessing, feature extraction (structural URL features plus NLP-based keyword analysis), and SVM training with grid search optimisation. Evaluation was performed via confusion matrix and standard metrics of accuracy, precision, recall and F1-score. The best model achieved an accuracy of 99.99%, precision of 99.98%, recall of 100%, and F1-score of 99.99%. These results demonstrate that the combined NLP + SVM approach can effectively distinguish phishing from legitimate URLs with very high reliability. The proposed system contributes to cybersecurity by offering a feasible AI-based solution for real-time URL screening that can be integrated into browser extensions or enterprise email filters to bolster phishing defences.

Artificial Intelligence in Green and Sustainable Investment: a Bibliometric and Systematic Literature Review

Antika Zahrotul Kamalia, Arief Wibowo, Deni Mahdiana — Sun, 15 Feb 2026 00:00:00 +0000

Green and sustainable investment has gained increasing global attention due to the urgency of the climate crisis, social demands, and the adoption of Environmental, Social, and Governance (ESG) principles. However, research on the application of artificial intelligence (AI) in this domain remains fragmented and lacks a comprehensive mapping. This study aims to map the trends, research directions, and key findings related to AI in green and sustainable investment using a bibliometric and systematic literature review (SLR) approach. Data were retrieved from the Scopus database and screened with the PRISMA framework, resulting in 24 articles analyzed through VOSviewer and thematic synthesis. The results indicate significant developments in energy efficiency, green buildings, machine learning, and sustainability, alongside an expanding pattern of international collaboration. Nonetheless, limitations remain, including insufficient cross-sectoral integration, limited empirical studies in developing countries, and the lack of AI models that holistically incorporate risk, ESG, and SDGs indicators. The main contribution of this study lies in providing a structured literature mapping that can serve as a foundation for developing more integrative AI frameworks and expanding research contexts to optimize sustainable green investment. These findings are expected to be valuable for researchers and practitioners in advancing innovation and strengthening the AI-driven sustainable finance ecosystem.

Comparison of the Accuracy Levels of Naive Bayes, Random Forest, and Long Short-Term Memory (LSTM) Methods in Predicting Gold Jewelry Sales

Muhammad Arfianto Pandu W, Rujianto Eko Saputro, Purwadi, Umdah Aulia Rohmah — Sun, 15 Feb 2026 00:00:00 +0000

Gold has long been recognized as a safe haven asset, especially during economic uncertainty. Accurate prediction of gold jewelry sales is essential for inventory management and business strategy, particularly in high-demand regions such as Imogiri. This study aims to compare the accuracy levels of three machine learning methods—Naïve Bayes, Random Forest, and Long Short-Term Memory (LSTM)—in predicting gold jewelry sales using historical transaction data from Toko Emas Parimas. The dataset comprises 4,595 records from January 2022 to December 2024. The research employs data preprocessing, including data cleaning, feature transformation, and normalization, followed by classification into sales categories. Two data-splitting schemes (80:20 and 70:30) were implemented to evaluate model generalization. The models were trained and tested using performance metrics such as accuracy, precision, recall, and F1-score. The results show that Random Forest achieved perfect classification with an accuracy of 1.00 in both schemes, outperforming the other models. Naïve Bayes also performed well with accuracy up to 0.98, while LSTM showed moderate results with accuracy ranging from 0.82 to 0.88. These findings indicate that Random Forest is the most reliable model for sales prediction of gold jewelry, especially for static classification tasks. The study provides practical insights for retailers and decision-makers in selecting suitable analytical models, and it highlights the importance of aligning analytical methods with data characteristics to improve decision support systems in retail.

Integrating Digital Governance for Disaster Resilience: A TOGAF 10-Based Enterprise Architecture for Coastal Villages in Eretan Wetan, Indonesia

Sofiyatun Hasanah, Asti Amalia Nur Fajrillah, ⁠Iqbal Yulizar Mukti — Sun, 15 Feb 2026 00:00:00 +0000

Indonesia's coastal regions face significant challenges due to climate change and natural disasters such as coastal abrasion, tidal flooding, and high waves, which impact the social and economic sustainability of rural communities. One of the vulnerable areas is Eretan Wetan Village, Kandanghaur Subdistrict, Indramayu Regency, which has a low score of 5.88 for SDGs Goal 13 (Climate Action Village). This studyse aims to design an Enterprise Architecture to support the implementation of a more effective, structured, and sustainable Coastal Disaster-Resilient Village (Destana). The design adopts the TOGAF 10 framework, covering the phases of Preliminary, Architecture Vision, Business Architecture, Data Architecture, Technology Architecture, Opportunities and Solutions, and Migration Planning. The outcome of this study includes an architectural blueprint and IT roadmap, which are expected to serve as a strategic guide for the village government in developing an integrated and adaptive disaster management system. Through this approach, Eretan Wetan Village is expected to enhance disaster preparedness, strengthen stakeholder coordination, and contribute to the achievement of sustainable development goals. This study shows how important it is in the field of information systems to solve real-world problems in rural regions through digital system integration.

Preventing Data Leakage in Classification via Integrated Machine Learning Pipelines: Preprocessing, Feature Transformation, and Hyperparameter Tuning

Sun, 15 Feb 2026 00:00:00 +0000

Data leakage in machine learning classification often leads to overfitting, inflated performance estimates, and poor reproducibility, which can undermine the reliability of deployed models and incur industrial losses. This paper addresses the leakage problem by proposing an integrated machine learning pipeline that strictly isolates training and evaluation processes across preprocessing, feature transformation, and model optimization stages. Experiments are conducted on the Titanic passenger survival dataset, where exploratory data analysis identifies data quality issues, followed by stratified train-test splitting to preserve class distribution. All preprocessing steps, including missing value imputation, categorical encoding, and feature scaling, are applied exclusively to the training data using a ColumnTransformer embedded within a unified Pipeline. A K-Nearest Neighbors (KNN) classifier is employed, with hyperparameters optimized via GridSearchCV and 3-fold cross-validation. Experimental results show that a baseline model without leakage control achieves only 72.62% test accuracy and exhibits a substantial overfitting gap. In contrast, the proposed pipeline-based approach improves generalization, achieving 78.21% test accuracy with an optimal configuration of k = 29 and Manhattan distance while significantly reducing overfitting. The main contribution of this work is the formulation of a reproducible, leakage-aware pipeline guideline that ensures unbiased evaluation and reliable generalization in classification tasks, providing practical methodological insights for both academic research and real-world machine learning applications.

Enhancing Question Classification in Educational Chatbots Using RASA Natural Language Understanding

Zaenur Dwi Christanto, Kristophorus Hadiono — Sun, 15 Feb 2026 00:00:00 +0000

This research develops a chatbot model based on Rasa Framework to understand and respond to questions related to informatics learning, addressing the critical need for personalized AI-driven educational tools in Indonesian secondary education. The model is trained to recognize various patterns of student questions about informatics materials, especially the topic of number conversion. Using Natural Language Understanding (NLU), the chatbot model is developed to process natural language and classify the intent of student questions. Evaluation of the model using the confusion matrix showed good performance with 91.5% accuracy, 94.4% average precision, and 100% recall. The test results showed that the model was able to correctly classify various types of intent, where eight out of nine intents achieved a perfect precision of 100%, with one intent, tutorial_calculation_octal_to_decimal, having a precision of 50%. The 100% recall across all intents demonstrates the model's comprehensive ability to identify all cases requiring responses, ensuring no student queries are missed. This research significantly contributes to computer science education by validating RASA's effectiveness for domain-specific NLU in low-resource educational settings, providing a scalable foundation for AI-based learning assistance tools that can enhance digital literacy and computational thinking skills among junior high school students.

Aspect-Based Sentiment Analysis of Access by KAI Application Reviews Using IndoBERT for Multi-Label Classification Tasks

Hilda Nur Alfiana, Afrizal Doewes, Bambang Widoyono — Sun, 15 Feb 2026 00:00:00 +0000

Ratings and reviews on mobile applications provide valuable insights into user experience and satisfaction with app features and services. However, ratings are subjective and often inconsistent with the content of the reviews. Therefore, a more in-depth analysis of the review content is necessary to identify evaluation points accurately. This study aims to evaluate the performance of IndoBERT in Aspect-Based Sentiment Analysis (ABSA) on Access by KAI application reviews. Data were collected by scraping user reviews from the Google Play Store, then annotated using a hybrid labeling approach. The resulting dataset was used to fine-tune the IndoBERT model across three ABSA tasks: aspect classification, sentiment classification for each aspect, and joint aspect-sentiment classification. We also benchmarked the model against baseline models to demonstrate its effectiveness. The results show that IndoBERT achieved the best performance across all tasks, specifically aspect classification (accuracy 0.928, F1-score 0.785), sentiment classification (accuracy 0.928, F1-score 0.752), and joint aspect-sentiment classification (accuracy 0.962, F1-score 0.549). Overall, IndoBERT successfully outperformed SVM and XGBoost with TF-IDF, BiLSTM with pre-trained IndoBERT embeddings, mBERT, and XLM-R. This study contributes a new dataset that provides resources for further research and development in Indonesian Natural Language Processing (NLP). These findings also highlight the advantages of a monolingual model trained specifically on Indonesian-language data.

Article Retrieval And Automatic Summarization System Using BERT-Based Neural Network Model On Chatbot

Sun, 15 Feb 2026 00:00:00 +0000

The rapid growth of online scientific publications presents challenges in efficiently filtering relevant information. Many search systems still rely on keyword matching, which is often ineffective in understanding the context of user queries. This study develops a chatbot system based on BERT (Bidirectional Encoder Representations from Transformers) for scientific article retrieval and automatic summarization. The system is designed to comprehend user intent and generate summaries of relevant articles. The evaluation was conducted on a dataset of 506 scientific articles, assessing search accuracy based on topic, abstract, author name, and time range. Results show 100% accuracy in searches by author and abstract, with varying performance in topic-based and time-based searches. This system is expected to enhance the efficiency and relevance of scientific literature retrieval and support the productivity of researchers across various fields.

K-Means Clustering with Elbow Method and Validity Indices for Classifying Student Academic Achievement Based on Knowledge Scores at SDN 48 Kota Jambi

M. Fikri Azmi, Dodo Zaenal Abidin, Jasmir — Sun, 15 Feb 2026 00:00:00 +0000

Student performance evaluation at SDN 48 Kota Jambi has been traditionally conducted manually, which is inefficient and often subjective. This study aims to provide an objective classification of students’ academic achievement using data-driven methods. The research applies the Knowledge Discovery in Databases (KDD) framework, which involves data selection, preprocessing, clustering, and evaluation. The dataset consists of knowledge scores from 152 elementary students across seven subjects, obtained from the Merdeka Curriculum report cards. Data preprocessing included cleaning and normalization to ensure consistency. K-Means clustering was implemented using RapidMiner, with the optimal number of clusters determined through the Elbow Method. Cluster validity was assessed using the Davies–Bouldin Index (1.226) and the Silhouette Coefficient (0.245). The results produced three clusters: high achievers (30.9%), medium achievers (27.0%), and low achievers (42.1%). Centroid analysis indicated that Mathematics and Physical Education were the most discriminative subjects across groups. These findings highlight a substantial proportion of students requiring remedial intervention and support differentiated learning strategies. The contribution of this research lies in applying educational data mining techniques to an elementary school context in Jambi, integrating both quantitative indices and qualitative validation with teachers. The study demonstrates that clustering methods can enhance educational decision-making, providing a basis for adaptive teaching, targeted interventions, and resource allocation in elementary education.

Random Forest and LLM Synergies Framework for Autonomous DDoS Mitigation

Sun, 15 Feb 2026 00:00:00 +0000

Modern Distributed Denial of Service (DDoS) attacks increasingly evade traditional defenses, and while Machine Learning (ML) has improved detection accuracy, a critical challenge remains in bridging detection with effective automated mitigation. This paper introduces a novel framework centered on a cognitive agent that synergistically combines high-speed ML detection with the advanced reasoning capabilities of a Large Language Model (LLM) for autonomous DDoS mitigation. The proposed architecture operates as a closed-loop security system. Following a data preprocessing phase that includes one-hot encoding and Standard Scaling (z-score normalization), a fine-tuned Random Forest model was identified as the optimal detector with 95.99% accuracy on the UNSW-NB15 dataset, which in turn triggers the LLM-based agent. This agent autonomously generates both human-readable incident explanations and machine-executable mitigation commands. Crucially, all generated commands undergo a syntax and logic validation step before execution to ensure operational integrity. Empirical results demonstrate the framework's efficacy, achieving a complete end-to-end detection-to-mitigation cycle in 24.20 seconds. This work validates that the unified approach presents a viable and transparent paradigm, contributing to the field of cybersecurity by enhancing automated mitigation and analytical processes through responsive and intelligent defense mechanisms.

Deep Learning-Based Detection of Potato Leaf Diseases Using ResNet-50 with Mobile Application Deployment

Cahyono Budy Santoso, Rufman Iman Akbar Effendi, Johannes Hamonangan Siregar — Sun, 15 Feb 2026 00:00:00 +0000

Plant diseases significantly reduce agricultural productivity, especially in developing regions with limited access to early detection tools. This research presents a deep learning-based approach for detecting potato leaf diseases, focusing on Early blight, Late blight, and healthy conditions. A modified ResNet-50 architecture was employed and trained using a publicly available potato leaf image dataset. Preprocessing steps included data augmentation and normalization to enhance model generalization. The model achieved a high accuracy of 99.31%, with precision, recall, and F1-score all exceeding 99%, indicating excellent classification performance. This study introduces a novel approach that improves classification performance through an optimized deep learning architecture, achieving higher accuracy compared to existing models. In addition to enhancing predictive capability, the study also addresses the practical need for accessibility by integrating the trained model into an Android-based mobile application. The application allows users to upload or capture leaf images and receive real-time predictions. The interface was designed for simplicity and usability in field conditions, making it accessible to farmers and agricultural workers. The findings demonstrate that combining deep learning with mobile technology can offer an effective and scalable solution for early disease detection in agriculture. Future work may explore cross-crop adaptability and lightweight model optimization for real-time performance on low-resource devices.

Web-Based Attendance and Leave Management System with Sequential Search Implementation at Tondano Religious Court

Gladly C. Rorimpandey, Natalia Arsel Rombon, Quido C Kainde — Sun, 15 Feb 2026 00:00:00 +0000

Based on the problems found in the leave application and attendance recording processes at the Religious Court of Tondano, this research was conducted to address these issues through the implementation of the Sequential Search Algorithm in a web-based application. This application aims to manage leave requests and employee attendance efficiently. The existing manual processes, such as paper-based attendance sheets and leave forms requiring multiple physical signatures, have proven to be inefficient, prone to manipulation, and at risk of data loss. To solve these issues, the application was developed using the Rapid Application Development (RAD) method, which enables a faster process in planning, designing, and deploying the system. The implementation of the Sequential Search Algorithm allows for efficient data retrieval, particularly in searching employee leave data without requiring data sorting, thereby simplifying the search process. In addition, the system includes a location-based attendance feature to prevent fraud, ensuring that employees are present within a defined radius before marking attendance. This online system benefits both permanent and non-permanent employees by providing a more secure, accurate, and accessible record of attendance and leave history. This research contributes to the digital transformation efforts of the Religious Court of Tondano by offering an integrated system that supports internal Administration processes and improves overall service quality through the use of information technology.

Buffalo Price Estimation Using YOLOv8 And Image Thresholding

Amelia, Wawan Firgiawan, Sulfayanti — Sun, 15 Feb 2026 00:00:00 +0000

The skin color pattern of buffaloes can determine their market price, especially for traditional ceremonial purposes that involve buffaloes. Currently, the pricing of buffaloes is still done subjectively by sellers or buyers, resulting in inconsistencies in price determination. This study proposes the development of a system to estimate the price of buffaloes based on their type and the percentage of light and dark skin, specifically for the Saleko buffalo type. The algorithm used to recognize buffalo types is YOLOv8, which was trained to detect four classes: Lotongboko, Saleko, Bonga, and Other types. The model was trained over 100 epochs using the Adam optimizer and hyperparameters. A thresholding method was applied to identify the percentage of black and white on the Saleko buffalo images that were successfully detected by YOLOv8. If the light skin percentage exceeds 80%, the buffalo is estimated to be worth 800 million rupiah. Otherwise, the Saleko buffalo is estimated at 300 million rupiah. The YOLOv8 training achieved a highest mAP value of 97.8%, with steadily decreasing loss and increasing metrics at each iteration, indicating a successful training process with strong detection performance. The price estimation model achieved an accuracy of 76.3% based on 55 tested images. Estimation errors were caused by low image resolution and poor lighting quality. This study provides insights into the application of technology for buffalo price estimation through digital image processing.

Optimized KNN Performance with PCA and K-Fold Cross-Validation for Colorectal Cancer Survival Prediction

Yuke Manza, Rika Rosnelly, Mhd Furqan, Bob Subhan Reza — Sun, 15 Feb 2026 00:00:00 +0000

Colorectal cancer remains a leading cause of global mortality, necessitating effective predictive tools for patient survival. While Machine Learning algorithms like K-Nearest Neighbors (KNN) utilize patient data for prediction, standard KNN implementations often suffer from the curse of dimensionality and overfitting, leading to unreliable performance on complex medical datasets. This study aims to evaluate and optimize the performance of the KNN algorithm by integrating Principal Component Analysis (PCA) for dimensionality reduction and K-Fold Cross-Validation (KFCV) to enhance model stability. The research utilized a quantitative approach on a global colorectal cancer dataset, processing demographic and clinical features through a rigorous pipeline of imputation, encoding, and normalization. Three model configurations were systematically compared: Standard KNN, KNN combined with PCA, and an optimized KNN model utilizing both PCA and KFCV across various neighbor values. The results demonstrate a distinct trade-off between predictive sensitivity and model stability. While the Standard KNN and PCA-enhanced models achieved higher recall, indicating a strong ability to identify survivors in a single data split, the fully optimized KNN+PCA+KFCV model provided the most stable and generalized accuracy with minimal deviation. These findings indicate that while PCA effectively reduces computational complexity without information loss, the integration of cross-validation is crucial for obtaining an honest assessment of model performance. This research contributes to clinical informatics by highlighting the necessity of prioritization between high sensitivity and generalization stability when developing survival prediction models for complex, inseparable medical data.

Performance Comparison of SVM in Sentiment Analysis of Israel-Palestine Comments Using Lsa and Word2vec

Sun, 15 Feb 2026 00:00:00 +0000

This study compares two feature extraction techniques, namely Latent Semantic Analysis (LSA) and Word2Vec, in the sentiment classification of comments related to the Israeli-Palestinian conflict using Support Vector Machine (SVM). The dataset consists of 1000 YouTube comments and 158 news paragraphs, categorized into pro and con Palestinian sentiments. The preprocessing process includes casefolding, special character and stopword removal, lemmatization, and tokenization. The results show that SVM with Word2Vec has better performance than SVM with LSA in the classification of positive and negative comments. SVM model with Word2Vec recorded a precision value of 92% and F1-Score of 93% on negative comments. Meanwhile, SVM with LSA recorded 90% precision and 92% F1-Score. On positive comments, SVM with Word2Vec recorded 92% recall and 93% F1-Score. While SVM with LSA recorded 89% recall and 91% F1-Score. Word2Vec's strength lies in its ability to capture word context and nuance more effectively, thanks to training using richer contextualized comment and news data. In conclusion, although both methods show good ability in sentiment classification, the use of Word2Vec provides more consistent and accurate results. This research contributes to the advancement of sentiment classification methods in the context of complex socio-political issues and can serve as a reference for applying machine learning to more accurate and contextual public opinion analysis.

IoT-Based Smart Detector with SVM and XGBoost for Real-Time Child Growth Monitoring and Stunting Risk Prediction

Fajar Mahardika, Lutfi Syafirullah, Adlan Nugroho — Sun, 15 Feb 2026 00:00:00 +0000

Stunting is a major public health issue, particularly in developing countries, causing long-term physical and cognitive impairments in children that reduce their quality of life and future productivity. To address this challenge, this study aims to develop an IoT-based smart detection system for child growth monitoring, enabling quicker and more accurate detection of stunting risks. The proposed system combines both hardware and intelligent software components to measure key growth indicators—height, weight, and BMI—using digital sensors and microcontrollers, transmitting the collected data to a cloud platform for real-time analysis. Machine learning algorithms, such as Support Vector Machine (SVM) and Extreme Gradient Boosting (XGBoost), are employed to predict stunting risk. Experimental results show that the XGBoost model outperforms SVM, achieving an accuracy of 80%, precision of 82%, recall of 78%, and F1-score of 79.9%, compared to SVM’s accuracy of 70%, precision of 68%, recall of 65%, and F1-score of 66.4%. This research provides a scalable technological framework for real-time stunting monitoring and early intervention, with the potential for implementation in resource-limited settings. By supporting national stunting reduction initiatives, the system enhances public health innovation and child welfare.

Improved Contrast and Clarity in Plant Microscopic Images using Contrast Limited Adaptive Histogram Equalization

Eka Wahyu Hidayat, R Reza El Akbar, Muhammad Adi Khairul Anshary — Sun, 15 Feb 2026 00:00:00 +0000

This research aims to enhance the quality of microscopic plant images which often suffer from low contrast and noise, hindering both visual and automated analysis. We propose the application of the Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm to address this issue. Implementation was carried out using MATLAB, processing a dataset of microscopic images from the Biology Laboratory of Siliwangi University. The research methodology includes image pre-processing, applying CLAHE with a Tile Grid Size of 8×8 and a Clip Limit of 0.02, and a quantitative evaluation using full-reference metrics such as MSE, PSNR, SSIM, RMSE, and FSIM. The results show that the application of CLAHE consistently demonstrated a significant improvement in image quality. Based on calculations, the lowest MSE value was found in the “monokotil (L.S)” image with 644.046 and the highest in the Monocotyledon Stem image with 6,298,683. The highest PSNR value was achieved by the “monokotil (L.S)” image with 46.225 dB, while the lowest was in two Monocotyledon Stem images, at 25.174 dB and 23.422 dB. The highest SSIM value was also in the “monokotil (L.S)” image with 0.946, indicating a very high structural similarity. Likewise, the highest FSIM value was also found in the “monokotil (L.S)” image with 0.979. This enhancement is crucial for botanical analysis and bioinformatics applications, as it effectively increases contrast, reduces noise, and preserves structural integrity, thereby facilitating the identification of fine details in microscopic images. These results establish a reproducible enhancement baseline that strengthens downstream botanical analytics.

Predicting Anxiety of STMIK Palangkaraya Students Using K-Means Clustering and Gaussian Naïve Bayes

Maura Widyaningsih, Rosmiati, Paholo Iman Prakoso — Sun, 15 Feb 2026 00:00:00 +0000

Academic anxiety is a common psychological problem experienced by students, especially before final exams, which impacts learning performance and mental well-being. This study aims to identify and predict students' anxiety levels using a Machine Learning approach, specifically the web framework Gradio, through a combination of the K-Means Clustering and Gaussian Naïve Bayes (GNB) methods. The research instrument used a Google Form-based questionnaire modified from the Zung Self-Rating Anxiety Scale (ZSAS) with 20 items (K1–K20) on a Likert scale (0–3). Data were obtained from 110 students of the Information Systems and Informatics Engineering Study Program at STMIK Palangkaraya. The research process consisted of five main stages: pre-processing, clustering using the K-Means algorithm, training the GNB classification model, evaluation, and prediction of new data. The clustering results categorized the data into three levels of anxiety: Low, Median, and High. The GNB model showed 95% accuracy with a balanced distribution of evaluation metrics (precision, recall, and F1 score). Comparison with other algorithms shows that while SVM achieved the highest accuracy (100%), GNB was more balanced in handling uneven class distributions and more practical for implementation in web-based systems. This prediction system has the potential to be used as an early detection tool for student anxiety, while also supporting educational institutions in designing more targeted psychological interventions. Further improvements can be made by expanding the scope of respondents, balancing the data distribution, and testing other machine learning methods to improve model generalization. The program and data are available at: https://github.com/maurawidya75/StudentAnxiety2025.

Abstractive Summarization of Indonesian Islamic Stories Using Long Short-Term Memory (LSTM)

Aisya Gusti Savila, Supriyono, Roro Inda Melani — Sun, 15 Feb 2026 00:00:00 +0000

The length of narratives in stories often poses a challenge for many readers, especially those with time constraints or difficulty understanding the entire story. In this case, summarization offers a solution, but manual summarization is not always efficient in meeting the need for quick and concise information. This study aims to develop an automatic text summarization system for Islamic stories using the Long Short Term Memory (LSTM) algorithm. The study employs three data splitting scenarios for training and testing: 90:10, 80:20, and 70:30. Testing results show that the highest training accuracy was achieved in the 80:20 scenario with a value of 89.44%. This does not entirely indicate that a smaller proportion of training data will always result in higher accuracy, as this improvement can be influenced by data variation, overfitting conditions, and early stopping performance. Therefore, the data division ratio influences the training process. Although the highest training accuracy was obtained in the 80:20 scenario, the best semantic summary quality was found in the 90:10 scenario. In the 90:10 scenario, the ROUGE-1 evaluation score achieved a precision of 0.4147, a recall of 0.2516, and an F1-score of 0.3027. Meanwhile, ROUGE-2 achieved a precision of 0.1022, a recall of 0.0568, and an F1-score of 0.0684. Meanwhile, ROUGE-L achieved a precision of of 0.2017, recall of 0.1209, and F1-score of 0.1459.

Implementation and Analysis of QR Code Phishing Attacks on Indonesian Internet Banking Using Attack Tree and Time-Based Metrics

Shavira Eka Yuniati, Adityas Widjajarto, Umar Yunan Kurnia Septo Hediyanto — Sun, 15 Feb 2026 00:00:00 +0000

The development of technology in Internet banking services facilitates customers’ financial transactions. However, this can also create opportunities for cybercrime threats, including a quishing attack. A quishing attack is a type of phishing attack that uses a QR Code to redirect victims to a fake website to steal sensitive information. This research formulates an attack tree model for quishing attacks by combining OSINT, social engineering, and QR Code exploitation, structured using data flow diagrams and evaluated with time-based metrics. The attack was simulated as a Proof of Concept (PoC) to realistically depict the stages of exploitation. Results from the experiments show that the fastest attack path using the OSINT tool Truecaller, the social engineering tool SEToolkit, and the QR Code tool Qrencode takes 248.31 seconds. This path is considered more efficient, outperforming the second fastest combination, which uses the OSINT tool Find Mobile Number Location by 25.15 seconds, with a total time of 273.46 seconds. Truecaller’s advantage lies in its ability to obtain data quickly without requiring a geographic location process like the Find Mobile Number Location tool. This approach shows that banking institutions can integrate time-based metric attack trees to assess vulnerability response times, simulate realistic threat scenarios, and develop more effective incident response strategies to prevent unauthorized access during quishing attacks.

Hybrid LSTM Forecasting Framework with Mutual Information and PSO–GWO Optimization for Short-Term SARS-CoV-2 Prediction in Indonesia

Faulinda Ely Nastiti, Shahrulniza Musa, Imam Riadi — Sun, 15 Feb 2026 00:00:00 +0000

SARS-CoV-2 remains an endemic challenge in Indonesia, requiring reliable short-term forecasting tools that support informatics, digital epidemiology, and data-driven public health systems. Standard LSTM models, while widely used for epidemic forecasting, face notable limitations such as sensitivity to poor weight initialization, and reduced ability to capture interactions within heterogeneous high-dimensional data—resulting in inconsistent performance. This research introduces ADELMI (Adaptive Deep Learning Metaheuristic Intelligence), a unified hybrid forecasting framework specifically designed not only to enhance forecasting accuracy but also to overcome core weaknesses of traditional LSTM architectures when applied to complex epidemic datasets. ADELMI integrates Mutual Information and Pearson Correlation for dual feature selection with a hybrid Particle Swarm–Grey Wolf Optimization (PSO–GWO) approach for optimizing LSTM parameters. The dataset includes 657 daily observations and 82 epidemiological, vaccination, and meteorological variables sourced from the Ministry of Health and BMKG (2020–2021). Feature selection reduced the dataset to 20 relevant predictors for recovery and death and one dominant predictor for positive cases. The optimized 50-unit LSTM with early stopping achieved highly accurate 7-day forecasts, producing MAPE scores of 0.01% (positive cases), 1.44% (recoveries), and 3.00% (deaths) across 5-fold cross-validation. These results significantly outperform ARIMA, SIR, and baseline LSTM models. By unifying dual feature selection with hybrid PSO–GWO optimization, ADELMI improves LSTM stability, weight initialization, and multivariate interaction modeling, delivering more reliable forecasts across heterogeneous datasets. This advancement strengthens informatics through DL-metaheuristic multivariate epidemic modeling and enables proactive, adaptive surveillance against evolving threats such as influenza hybrids.

Natural Language Understanding for School Bullying Detection and Consultation: A DIET Classifier Approach in RASA Framework

Yoan Freddy Irawan, Kristophorus Hadiono — Sun, 15 Feb 2026 00:00:00 +0000

This research presents the development and implementation of a DIET classifier-based chatbot system using the RASA Framework to handle bullying reports at SMP Negeri 3 Ungaran. The system aims to provide 24/7 automated counseling support service, addressing the limitations of traditional human-to-human support systems that often result in delayed responses and reduced user satisfaction. The model was trained using a structured dataset comprising 61 dialogue examples collected through interviews with experienced guidance and counseling teachers, capturing authentic student communication patterns related to bullying issues. The evaluation results demonstrate exceptional performance, achieving 100% accuracy across 12 intent categories, with perfect precision and recall scores. The system successfully distinguishes between various emotional states and counseling needs, providing appropriate responses with high confidence levels. The intent categories include emotional expressions (merasa_dibully, merasa_sedih, merasa_takut), support-seeking behaviors (butuh_nasihat, ingin_bicara_dengan_guru), and conversational elements, ensuring comprehensive coverage of bullying-related communication scenarios. This implementation proves that AI-driven solutions can effectively support educational institutions in providing immediate, accessible counseling assistance while maintaining accuracy in emotional support and bullying prevention. This research contributes to the field of computer science by demonstrating the practical application of natural language understanding frameworks in sensitive educational contexts, advancing AI-driven counseling systems that can be scaled across educational institutions. The study provides a replicable methodology for developing culturally-sensitive AI applications in educational environments, particularly valuable for institutions in developing countries with limited digital mental health resources.

Evaluating Ensemble Versus Non-Ensemble Machine Learning Performance with Preprocessing Techniques for IoT Intrusion Detection on CICIoT2023

Febrian Sabila Firdaus, Puspanda Hatta, Basori — Sun, 15 Feb 2026 00:00:00 +0000

The rapid expansion of the Internet of Things (IoT) introduces significant security vulnerabilities, exposing networks to sophisticated attacks. Developing effective Intrusion Detection Systems (IDS) is critical, yet many machine learning benchmarks rely on outdated datasets. This study provides a comprehensive comparative evaluation of ensemble and non-ensemble machine learning models for multiclass attack classification using the modern and complex CICIoT2023 dataset. The methodology involves robust preprocessing, including random undersampling to address extreme class imbalance and a hybrid feature selection approach combining Mutual Information (MI) and Random Forest Feature Importance (RFFI). Models, including Naive Bayes, Logistic Regression, SVM, Random Forest, and XGBoost, were evaluated using stratified 5-fold cross-validation (K=5) with default hyperparameters. The results demonstrate that ensemble models consistently and significantly outperform non-ensemble models. XGBoost achieved the highest and most stable performance, yielding a mean F1-score of 0.8889 ± 0.0008 across the K-folds, and a final macro F1-score of 0.8891 on the test set. This research confirms the superiority of ensemble methods for complex IoT traffic and quantitatively highlights the critical role of preprocessing. Notably, scaling was proven essential for non-ensemble models, drastically improving Logistic Regression's F1-score from an unstable 0.6280 to 0.7691.

Comparative Analysis of Face Mask Detection using Lightweight CNN and Bag of Visual Word-based Classifier for Real-Time Surveillance

Ika Candradewi, Bakhtiar Aldino Ardi S, Agus Harjoko, Andi Dharmawan — Sun, 15 Feb 2026 00:00:00 +0000

Face mask detection has become increasingly important across various sectors, including healthcare, food processing industries, and public safety, to ensure adherence to health and hygiene protocols and minimize the risks of contamination. Manual supervision of mask usage is often inefficient, labor-intensive, and prone to inconsistency. To address this challenge, this study proposes an automated face mask detection system utilizing computer vision technology, designed for real-time monitoring on resource-limited edge devices, such as the Raspberry Pi 4 with a Google Coral USB Accelerator.

The system integrates Multi-task Cascaded Convolutional Neural Networks (MTCNN) for face detection and a modified lightweight Convolutional Neural Network (CNN) for binary mask classification. Deployed via a web-based platform, it captures images of non-compliant individuals and triggers alerts. To evaluate model performance, the modified CNN is compared with the Bag of Visual Words (BoVW) method using SVM and MLP classifiers on two datasets: the 12k-Face Mask Dataset (Kaggle) and a newly proposed dataset. The CNN model demonstrated higher classification performance than both BoVW-SVM and BoVW-MLP, with AUC improvements of 49% and 43% on the proposed and 12k-Face Mask datasets, respectively.

This study contributes to the advancement of computer vision-based public health monitoring by offering a robust, scalable, and real-time face mask detection system. The findings highlight the practical advantages of deep learning approaches over traditional feature extraction techniques, supporting the development of intelligent, automated surveillance systems and policy enforcement in high-risk environments, which will facilitate future advancements in AI-driven public safety solutions.

Sales Forecasting Model for Indonesian Clothing MSMEs For Sales Strategy Optimization Using The Long Short-Term Memory Method

Sun, 15 Feb 2026 00:00:00 +0000

Micro, Small, and Medium Enterprises (MSMEs) in the clothing industry are one of the key pillars of the economy, contributing significantly to Gross Domestic Product (GDP) and employment. However, MSMEs face considerable challenges related to market competition, shifting consumer trends, and fluctuating demand. Advances in data analytics and machine learning offer solutions to improve sales forecasting accuracy, thereby supporting more effective business strategies. This study aims to develop a sales forecasting model based on Long Short-Term Memory (LSTM) tailored to the characteristics of clothing MSMEs in Indonesia. The research was conducted at Ananda Kids MSME in Purbalingga, using 30,885 daily transaction records collected over 23 months. The dataset included product categories, sales volume, and revenue, which were further processed through normalization, handling of missing values, and the addition of seasonal features. The LSTM model was designed with 128 neurons and evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). The findings indicate that the LSTM model achieved high accuracy for certain product categories. The “Set” and “Children’s Fashion” categories recorded MAPE values below 10%, demonstrating the model’s effectiveness in forecasting stable sales patterns. In contrast, categories with high volatility, such as accessories, produced larger prediction errors. These results highlight that data quality and sales pattern stability are crucial factors in enhancing model performance. Overall, the study demonstrates that the application of LSTM holds significant potential in supporting strategic decision-making for MSMEs through more accurate sales forecasting. Beyond its practical contributions for business actors, the study also provides a basis for the development of digitalization policies for the MSME sector in Indonesia.

Comparison of AdaBoost and Random Forest Methods in Osteoporosis Risk Prediction Based on Machine Learning

Edwardo Parlindungan H, Setiawan Assegaff, Jasmir Jasmir — Sun, 15 Feb 2026 00:00:00 +0000

This study aims to determine the most effective ensemble machine learning algorithm for osteoporosis risk prediction in resource-constrained healthcare settings, specifically comparing AdaBoost and Random Forest performance on Southeast Asian population data. We implemented nested 5-fold cross-validation on a dataset of 1,958 records with 15 lifestyle and demographic attributes. Both algorithms underwent hyperparameter optimization, and performance was evaluated using accuracy, precision, recall, F1-score, and clinical utility metrics including cost-effectiveness analysis. AdaBoost achieved superior performance with 86.90% accuracy (95% CI: 84.2-89.6%) and perfect precision (1.00) compared to Random Forest's 84.69% accuracy and 0.92 precision. Statistical significance testing confirmed AdaBoost's advantage (p=0.032). Clinical implementation in three health centers demonstrated 60% reduction in unnecessary referrals. This is the first study to compare these algorithms specifically for Southeast Asian populations with clinical validation and cost-effectiveness analysis, providing a ready-to-deploy model for resource-limited healthcare settings.

Efficient ECG-Based Sleep Apnea Detection Using CNN-GRU and Sparse Autoencoder

Ramadhian Eka Putra, Sani Muhamad Isa — Sun, 15 Feb 2026 00:00:00 +0000

Sleep apnea is a serious and common breathing disorder that occurs during sleep, characterized by repeated pauses in breathing that can increase the risk of hypertension, heart disease, and stroke. Early detection of sleep apnea is crucial, but conventional methods, such as polysomnography, are expensive, complex, and inefficient for mass screening. Therefore, an automated system based on physiological signals such as an electrocardiogram (ECG) is needed for a more practical and efficient approach. This study proposes a sleep apnea classification model utilizing a combination of 1D Convolutional Sparse Autoencoder (1DCSAE), Convolutional Neural Network (CNN), and Gated Recurrent Unit (GRU) architectures, referred to as the SAE-DEEP model. This method is designed to automatically extract features while minimizing the need for preprocessing. Four testing scenarios were conducted to evaluate the impact of signal reconstruction and preprocessing on classification performance. Experimental results show that the CNN-GRU model with signal reconstruction using 1DCSAE achieves an accuracy of 89.8%, a sensitivity of 90.1%, and a specificity of 89.2%, demonstrating balanced and stable classification performance. Additionally, this model was proven to work effectively without complex preprocessing steps, making it a potential solution for efficient sleep apnea detection systems. These findings could contribute to the development of more straightforward, reliable, and clinically viable ECG-based classification systems, as well as wearable devices. In doing so, the proposed model addresses a critical gap in sleep apnea screening, underscoring the urgent need for accessible and cost-effective diagnostic tools.

Integration of Thermal Images and Agricultural Data for Multi-Class Classification of Palm Seed Origin using MobileNet

Yusuf Abidin Nurrahman, Rifki Wijaya, Tjokorda Agung Budi Wirayuda — Sun, 15 Feb 2026 00:00:00 +0000

This research develops a palm kernel origin classification model by combining thermal images and numerical agricultural data using MobileNet architecture. The quality of palm kernels is highly influenced by origin and environmental conditions, but manual visual identification is difficult. Therefore, a machine learning-based approach is applied to improve classification accuracy. The dataset consists of 7.257 thermal images representing 73 seed origin classes, as well as supporting data in the form of soil, fruit, and socioeconomic information collected from plantations in Aceh, Indonesia. The MobileNet model was tested in two scenarios: using only thermal images, as well as a combination of thermal images with agricultural data. Results show that data integration provides significant performance improvement. The best model was obtained from MobileNet V3-Large with the optimal hyperparameter configuration (batch size 16, learning rate 0.001, and optimizer Adam), resulting in test accuracy of 99.04%, validation 97.25%, and training 98.77%. This finding opens up opportunities for the application of real-time classification technology in the plantation environment, especially to support precision and sustainable agriculture.

Benchmarking Relational and Array-Based Models for Genealogical Data Storage in PostgreSQL

Suwanto Raharjo, Ema Utami — Sun, 15 Feb 2026 00:00:00 +0000

Genealogical information systems manage inherently hierarchical data structures that represent family relationships across multiple generations. Traditional implementations predominantly rely on normalized relational database designs using junction tables to model parent–child relationships. While this approach ensures strong referential integrity, it often incurs substantial performance overhead due to complex join operations during deep hierarchical traversal. Recent versions of PostgreSQL provide native support for array data types This study compares two genealogical database models implemented in PostgreSQL: a normalized relational model using a junction table and a denormalized model that stores child identifiers directly as UUID arrays. To evaluate their performance, we conducted controlled benchmarking experiments using synthetically generated genealogical datasets with varying generational depth and branching patterns. The comparison focuses on storage efficiency, recursive traversal performance, and write operation costs under realistic hierarchical workloads. Results obtained from a large-scale dataset containing more than 7 million individual records show that the UUID array–based model reduces disk space usage by 31%. During deep recursive traversal involving over 12 million nodes at the tenth generation, the array-based model demonstrates improved data locality, leading to a 5.2% reduction in execution latency and 7% fewer shared buffer accesses compared to the relational model. Interestingly, contrary to common expectations in normalized database design, the array-based model achieves 22% faster single-insert performance because it avoids foreign key validation and multiple index updates. This improvement comes with slightly higher write amplification, reflected in a 6.6% increase in buffer usage due to PostgreSQL’s multi-version concurrency control mechanism. These findings contribute to the field of Informatics by providing empirical evidence on how database internal mechanisms influence performance trade-offs in hierarchical data management, offering guidance for designing scalable and read-efficient information systems beyond genealogical applications.

User Experience Analysis of Learning Management System (LMS) SINAU to Support Learning with MERDEKA Flow Using UX Curve Method

Sri Yarsasi, Imam Tahyudin, Taqwa Hariguna — Sun, 15 Feb 2026 00:00:00 +0000

The rapid development of information technology has driven transformation in education, including the use of Learning Management Systems (LMS) to facilitate independent and flexible learning aligned with the Merdeka Curriculum. This study aims to evaluate the user experience (UX) of the Sinau LMS at SMA Negeri 1 Sidareja using the UX Curve method, which tracks changes in user perceptions over time. The research involved 20 grade XII students who had used the LMS for at least three months. Data were collected through initial questionnaires, interviews, UX curve drawings, and final questionnaires, focusing on five main UX aspects: General UX, Attractiveness, Ease of Use, Utility, and Degree of Usage. The analysis of 100 curves revealed that more than half of the respondents experienced a decline in user experience quality, particularly in Ease of Use, General UX, and Degree of Usage, due to issues such as an unattractive interface, navigation challenges, and limited feature relevance. Conversely, a minority showed improved perceptions as they adapted and became more familiar with the system. These findings highlight the need for continuous improvement of the LMS's interface design and features to enhance user satisfaction and learning effectiveness. The study contributes theoretically by demonstrating the application of the UX Curve in educational systems and practically by providing recommendations for refining LMS development to better support the Merdeka Curriculum.

Hybrid Heuristic Algorithms for Optimizing University Graduate-Job Matching: A Quantitative Study in Indonesia's 2025 Labor Market

Ikbal Nidauddin, Kresno Murti Prabowo, Abdullah Alim — Sun, 15 Feb 2026 00:00:00 +0000

University graduate unemployment in Indonesia reached critical levels with 1,010,652 unemployed graduates in 2025 (BPS data), representing approximately 15% of national unemployment due to severe skills mismatch between education outcomes and labor market demands. This research develops and validates a novel hybrid heuristic algorithm integrating Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Simulated Annealing (SA) with adaptive diversity-based switching mechanisms to optimize graduate-job matching through multi-objective competency profile alignment. The quantitative experimental study collected data from 200 university graduates across five academic disciplines and 5 major recruiting companies through structured surveys and competency assessments. The proposed GA-PSO-SA hybrid algorithm with adaptive switching achieved 92.4% matching accuracy (35% improvement over traditional methods), 42% faster convergence compared to single algorithms (10.6s vs. 18.4s for pure GA), and solution quality of 8.9/10. Statistical validation through paired t-tests demonstrated highly significant improvements (p < 0.001, Cohen's d > 2.0) across all comparisons. The system successfully reduces average job search duration by 40% (from 6+ months to 3.6 months) and improves graduate placement success rates by 28%. This research contributes a theoretically-grounded and empirically-validated intelligent recommendation system addressing Indonesia's graduate employment crisis through computational optimization, with implications for national workforce development and recruitment efficiency enhancement.

Deep Learning-Based Autism Detection Using Facial Images and EfficientNet-B3

Sun, 15 Feb 2026 00:00:00 +0000

This study presents a novel deep learning approach for early detection of Autism Spectrum Disorder (ASD) using facial image analysis. Leveraging the EfficientNet-B3 model, the research addresses limitations in traditional diagnostic methods by autonomously extracting discriminative facial features associated with ASD. A balanced dataset of 2,940 facial images (1,470 autistic and 1,470 non-autistic children) from Kaggle was pre-processed to 200x200 pixels and evaluated under three dataset-splitting scenarios (80:10:10, 70:15:15, and 60:20:20) to assess generalisability. The model, trained with the Adam optimiser over 10 epochs, achieved optimal performance in the 80:10:10 scenario, with 84.67% precision, 84.35% recall, and 84.32% F1 score. Results demonstrate high confidence (>90% probability) in distinguishing autistic from non-autistic individuals on unseen data. The study underscores the potential of integrating deep learning into clinical decision-support systems for ASD detection, offering a robust, scalable, and efficient solution to improve diagnostic accuracy and reduce reliance on manual methods.

Optimizing Bag of Words and Word2Vec with Vocabulary Pruning and TF-IDF Weighted Embeddings for Accurate Chatbot Responses in Indonesian Treasury Services

Eko Aprianto, Deni Mahdiana, Arief Wibowo — Sun, 15 Feb 2026 00:00:00 +0000

The high volume of support tickets submitted to the HAI DJPb Service Desk has caused delays and inconsistent response quality in payroll-related inquiries across Indonesian treasury work units (Satker). To improve the accuracy and efficiency of public service responses, this research proposes an optimized text-vectorization framework for chatbot development using a hybrid combination of Bag of Words (BoW), Word2Vec, vocabulary pruning, and TF-IDF weighted embeddings. The dataset consists of 2024 ticket logs, curated FAQs, and questionnaire data related to the Satker Web Payroll Application. The method includes preprocessing (snippet removal, normalization, tokenization, stopword removal, stemming), vocabulary pruning based on empirical frequency thresholds (<5 and >80) while preserving domain-specific technical terms, and semantic weighting through TF-IDF. Four vectorization models—BoW, BoW with pruning, Word2Vec, and Word2Vec + TF-IDF—were evaluated using cosine similarity, response time, and accuracy. Results show that BoW achieved the highest accuracy of 88.32%, while Word2Vec produced the most stable response time with an average of 47.32 ms and a cosine similarity of 0.99. The findings demonstrate that frequency-based representations remain highly effective for structured administrative datasets, while weighted embeddings improve semantic relevance. This study contributes to the field of Informatics by providing an efficient hybrid vectorization framework tailored for Indonesian administrative language, enabling more accurate and scalable chatbot solutions for e-government services.

Comparative Analysis of Machine Learning-Based Software Defect Prediction in Object-Oriented and Structured Paradigms Using Apache Camel and Redis Datasets

Asro Nasiri, Arief Setyanto, Prof Ema Utami, Kusrini — Sun, 15 Feb 2026 00:00:00 +0000

Software Defect Prediction (SDP) is a crucial component of software engineering aimed at improving quality and testing efficiency. However, the majority of SDP research often overlooks the fundamental influence of the programming paradigm on the nature and causes of defects. This study presents a comparative analysis to identify the most influential software metrics for predicting defects across two distinct paradigms: Object-Oriented (OOP) and Structured. To ensure modern relevance and reproducibility, we constructed two new datasets from large-scale, open-source projects: Apache Camel (Java) for OOP and Redis (C) for Structured which exhibited realistic defect rates of 14.4% and 21.8%, respectively. The dataset creation process involved mining Git repositories for defect labeling and automated metric extraction using the CK and Lizard tools. Correlation analysis and baseline modeling using Random Forest revealed significant differences between the paradigms. In the OOP system, dominant defect predictors were related to the complexity of the class interface and features (e.g., uniqueWordsQty, totalMethodsQty, WMC, CBO). Conversely, defects in the structured system were strongly correlated with size and algorithmic complexity (e.g., file_tokens, file_loc, file_ccn_sum). Although the baseline models performed well (ROC–AUC = 0.82–0.87), the significant class imbalance resulted in low recall (44–50%). This motivates the need for more context aware approaches. These findings underscore that effective SDP strategies must be tailored to the underlying programming paradigm.

Alphabet Gesture Classification of Indonesian Sign Language Using Convolutional Neural Networks

Yanuar Gideon Simalango, Anindita Septiarini, Masna Wati, Hamdani, Rajiansyah — Sun, 15 Feb 2026 00:00:00 +0000

Indonesian Sign Language (BISINDO) serves as a communication medium for deaf individuals to engage with their environment. Alphabet gestures in BISINDO play a crucial role in the formation of words and sentences. Nonetheless, the automatic recognition of BISINDO alphabet movements remains a difficulty in the advancement of accessible technology. This research intends to categorize BISINDO alphabet gestures via the Convolutional Neural Network (CNN) model. The CNN approach was used due to its proficiency in recognizing visual patterns and images. The dataset comprises BISINDO alphabet gesture photos captured from diverse perspectives and lighting conditions. The data processing procedure encompasses pre-processing phases, including picture normalization, data augmentation, and the segmentation of the dataset into training, validation, and test subsets. The constructed CNN model has multiple convolutional and pooling layers to thoroughly extract visual characteristics. The study's results indicate that the CNN model can classify BISINDO alphabet gestures with a high accuracy of 90% on the test data. This model's deployment is anticipated to aid in the creation of automatic sign language translation programs, hence enhancing communication between the deaf community and the general populace. This study demonstrates the potential of CNN models to support the development of inclusive communication technologies for the hearing impaired in Indonesia, particularly for under-researched sign languages like BISINDO.

Empirical Evaluation of IndoBERT and LSTM for Sentiment Analysis of Tourism Reviews: A Data-Driven Study on Kenjeran Park

Devi Dwi Purwanto — Sun, 15 Feb 2026 00:00:00 +0000

Tourism plays a pivotal role in Indonesia’s economic and cultural landscape, contributing significantly to job creation, regional development, and international recognition. This study evaluates the performance of IndoBERT, a state-of-the-art Indonesian language model, and Long Short-Term Memory (LSTM) networks for sentiment classification of 2,560 Google reviews of Kenjeran Park in Surabaya, consisting of 54% positive, 28% neutral, and 18% negative sentiments. Preprocessing steps included slang replacement, stemming, stopword removal, and tokenization, with class imbalance addressed through weighted loss adjustments. IndoBERT was fine-tuned using contextual embeddings with a learning rate of 0.00005, while the LSTM model employed a 128-unit architecture trained over 150 epochs with the Adam optimizer. Experimental results show that IndoBERT achieved 87.50% accuracy, 0.7697 precision, 0.7643 recall, and 0.7643 F1-score, outperforming LSTM’s 77.93% accuracy, 0.6826 precision, 0.6812 recall, and 0.6826 F1-score. This research establishes a comparative benchmark of transformer-based and RNN-based architectures for Indonesian tourism review sentiment analysis, introduces a domain-specific preprocessing pipeline with imbalance handling, and provides actionable insights for digital tourism analytics. Beyond its technical contributions, the study highlights the urgency of advancing robust natural language processing approaches for low-resource languages, thereby strengthening the field of informatics and supporting data-driven decision-making in the tourism sector.

A Morphology Processing Approach For Image Processing In Cancer Diagnosis

Jonner Hutahaean, Yudi Widhiyasana, Algi Fari Ramdhani — Sun, 15 Feb 2026 00:00:00 +0000

Early tumor detection is critical for improving cancer treatment outcomes, enabling less invasive and more cost-effective interventions. However, limited access to pathologists and high patient volumes reduce diagnostic efficiency, particularly in underserved regions, underscoring the urgency for computational support tools. While deep learning has shown promise in tumor detection, it requires extensive annotated datasets, high computational resources, and long processing times, making it less feasible in certain contexts.This study introduces a lightweight image processing approach for detecting tumors in Hematoxylin and Eosin (H&E)–stained histopathology images without deep learning. Using data from the PAIP 2023 Tumor Cellularity challenge, the proposed method applies histogram equalization, bilateral filtering, morphological transformations, bitwise operations, and an improved algorithm adapted from prior research. The method achieves IoU (Intersection of Union) of 0.93 compared to pathologist-determined ground truth. The results indicate that this approach can serve both as a standalone segmentation tool and as a preprocessing stage for deep learning pipelines, enhancing accessibility, reducing computational costs, and supporting broader adoption of computer-aided pathology in resource-limited settings.