Building of Informatics, Technology and Science (BITS)

Reversible Data Hiding Citra MRI T1-Weighted Menggunakan Spatial Fuzzy C-Means dan Selective Histogram Shifting

2026-03-07T22:24:17+07:00

The transmission of medical images over telemedicine networks increases the risk of data leakage and manipulation of sensitive information. This study develops a Reversible Data Hiding framework that integrates Spatial Fuzzy C-Means, Selective Histogram Shifting, and a measurable Distortion Control Mechanism for securing T1-weighted brain MRI images. The proposed method prioritizes the preservation of Region of Interest intensity characteristics and full reversibility over embedding capacity. SFCM is employed to generate Region of Interest and Non-Region of Interest mappings based on intensity distribution, with adaptive parameter adjustment for each slice. Data embedding is performed selectively on NROI using histogram shifting, while ROI areas remain unmodified. An Adaptive Feedback Control mechanism monitors image quality metrics SNR, CNR, GLCM with conservative thresholds (ΔSNR ≤ 2.0%, ΔCNR ≤ 1.0%) to ensure ROI stability. Experimental evaluation on the OASIS-1 dataset shows that the proposed method achieves an average PSNR of 54.13 dB, SSIM of 0.9996, and NCC of 0.9999, with an embedding capacity of 630 bits per slice (BPP 0.007-0.013 within NROI). Reversibility verification confirms perfect recovery (maximum difference = 0) for all samples. Batch testing on five slices demonstrates consistent performance across varying intensity characteristics, with ΔSNR and ΔCNR remaining at 0.0%. These results indicate that the method is capable of maintaining ROI technical integrity and pixel-perfect reversibility, although with a limited capacity suitable for lightweight metadata such as integrity hashes and patient identifiers. Limitations of the study include the technical-only evaluation without radiologist clinical validation and testing restricted to T1-weighted MRI modality.

Prediksi Periode Fosil Trilobita Menggunakan XGBoost dengan Seleksi Fitur Geologi–Geospasial dan Hyperparameter Tuning

2026-03-07T22:34:29+07:00

This study investigates the application of the Extreme Gradient Boosting (XGBoost) algorithm to predict the age period of trilobite fossils based on geological and geospatial data. The challenges addressed in this research include the high complexity of paleontological data, the presence of missing values, and class imbalance in the target variable time_period, which can negatively affect predictive performance. The objective of this study is to develop an accurate and robust fossil age prediction model through systematic data preprocessing, feature selection, and model optimization. The dataset used in this research was obtained from Kaggle and consists of the attributes longitude, latitude, lithology, environment, and collection_type as the main features. The research workflow includes data cleaning, missing value imputation, categorical feature encoding, data splitting using stratified train–test split, and class imbalance handling through a class weight adjustment approach. The XGBoost model was trained on the training dataset and further optimized using RandomizedSearchCV to obtain the optimal hyperparameter configuration. Evaluation results on the testing dataset show that the tuned XGBoost model achieved an accuracy of 95%, precision of 90%, recall of 93%, and an F1-score of 91%, outperforming the model without hyperparameter tuning. These results demonstrate that the integration of geological–geospatial feature selection and hyperparameter tuning in XGBoost is effective in improving the performance of trilobite fossil age period prediction. The results of this study are expected to serve as a computational support approach in paleontology to assist fossil period determination in a more objective, efficient, and data-driven manner.

Deteksi Malware Android Berbasis Ensemble Soft Voting LightGBM, Logistic Regression dan CatBoost

2026-03-07T23:07:00+07:00

The Android operating system faces serious challenges with increasingly complex and diverse malware evolution. This research proposes an Android malware detection system based on soft voting ensemble that integrates three algorithms (LightGBM, Logistic Regression, and CatBoost) to improve detection accuracy while maintaining computational efficiency. The dataset used is CCCS-CIC-AndMal-2020, which is highly imbalanced with over 400,000 Android application samples. The proposed model leverages hybrid features that combine static information (permissions, intents, API calls from the AndroidManifest) with dynamic behavior (memory activities, runtime API calls, logcat, and network traffic in an emulated environment), balancing low extraction cost with improved robustness against obfuscation. The methodology includes multi-stage preprocessing (IQR capping 40×, StandardScaler, RFE 150 features, SMOTE 30%) to improve data quality and reduce dimensionality by 56% without losing important information. The ensemble model is trained with F1-Macro-based weights (33.46% LightGBM, 30.99% Logistic Regression, 35.55% CatBoost) approximating 1:1:1 proportion. Evaluation results on the testing set demonstrate very high performance: Accuracy 95.58%, Balanced Accuracy 92.21%, F1-Macro 0.9208, True Positive Rate 100%, and False Alarm Rate 0.00%. The combination of these metrics indicates that the model can detect all malware samples without false positives on benign applications, making it suitable for production deployment. This research contributes by demonstrating the effectiveness of an efficient soft voting ensemble (only 3 models) for Android malware detection with multi-dimensional evaluation metrics representative of imbalanced data.

Stacking-Correspondence Analysis for Fuzzy Data: Computational Framework for Analyzing Complex Qualitative Survey Data

2026-03-07T23:32:39+07:00

Bandung Regency faces a significant challenge in achieving Sustainable Development Goal (SDG) 12, marked by a critically low score of 14.53 out of 100. Uniform policies are often ineffective due to regional diversity and uncertainty in categorical survey data, which inadequately reflects real-world conditions. This study aims to identify sub-district characteristics based on consumption and production patterns to provide precise policy recommendations. The research utilizes data from the 2024 Supporting Area Survey (SWP), covering 280 villages across 31 sub-districts. A computational framework combining stacking techniques and Correspondence Analysis for Fuzzy Data (CAFD) is implemented to analyze four qualitative variables. The stacking phase transforms the multi-way data structure into a two-way structure, while CAFD effectively handles qualitative uncertainty using membership degrees. Analysis results indicate that two principal dimensions capture 73.35% of the total information variance and successfully identify 17 sub-district clusters with similar problem profiles. The fuzzy approach unveils multi-characteristic profiles, identifying both dominant and secondary traits. This research contributes a two-dimensional perceptual map, enabling the government to transition from generic policies to tailored interventions for each sub-district. This computational solution represents a concrete step toward improving the SDG 12 achievement score through data-driven strategic planning.

Penerapan Algoritma Naïve Bayes Terhadap Sentimen Ulasan Produk Skincare Pada E-Commerce Shopee

2026-03-07T23:43:15+07:00

The rapid growth of the beauty industry has generated a large volume of consumer reviews, necessitating an automated processing system to understand public sentiment. This study aims to implement sentiment analysis on skincare product reviews using the Multinomial Naïve Bayes algorithm. The labeling process was conducted by converting star ratings into sentiment categories: ratings 4 and 5 were labeled as positive, ratings 1 and 2 as negative, while rating 3 was excluded to avoid data ambiguity. The feature representation stage utilized TF-IDF with an N-gram approach (unigram and bigram), generating 10,000 features from a dataset of 8,646 reviews. Based on the testing results of 1,730 test data, the model achieved an accuracy of 70%. The Confusion Matrix evaluation revealed that the model performed exceptionally well in the positive class, reaching a recall of 1.00. However, the model struggled to classify negative and neutral classes, with recall values approaching 0.00. This was caused by imbalanced data distribution, where positive reviews significantly dominated the dataset. Nevertheless, Multinomial Naïve Bayes proved efficient in handling large-scale frequency-based textual features. A weighted average F1-score of 0.58 suggests that dataset optimization is required to improve the model's ability to accurately recognize minority sentiments.

Explainable Aspect-Based Sentiment Analysis with Contrast-Aware IndoBERT for Indonesian Public Service Reviews

2026-03-08T00:13:48+07:00

This study presents an Explainable IndoBERT with Contrast-Aware Attention framework for Aspect-Based Sentiment Analysis (ABSA) on Indonesian public service reviews. The proposed model integrates automated aspect labeling using KeyBERT with a contrast-aware mechanism to handle mixed or opposing sentiments within a single sentence. By leveraging IndoBERT as the base transformer, the system captures context-sensitive sentiment cues while maintaining interpretability through attention-based rationale extraction. Experimental results on the SMSA dataset demonstrate an accuracy of 83.4%, with strong precision in positive and negative sentiment detection. The contrast-aware module improves clause-level understanding, while the attention-based explainability module provides transparent, token-level rationales that align with human judgments at an average rate of 87.7%. Although a modest performance decline occurs compared to non-explainable baselines, the proposed model offers significant gains in semantic transparency, making it suitable for evidence-based policy evaluation and citizen feedback monitoring. This research contributes a practical, interpretable, and linguistically grounded solution for explainable sentiment analysis in low-resource languages, advancing the application of responsible AI in public service analytics.

Evaluasi Strategi Fine-Tuning pada ConvNeXt dan Swin Transformer untuk Klasifikasi Kanker Kulit

2026-03-08T00:22:41+07:00

Skin cancer is one of the diseases whose prevalence continues to increase every year, especially in areas with high exposure to ultraviolet (UV) rays. The main challenge in diagnosing skin cancer lies in the visual similarity between benign and malignant lesions, which often leads to misdiagnosis even by experienced medical personnel. The development of deep learning technology has made significant progress in medical image classification through a transfer learning approach. This study aims to compare the performance of two architectures from Transformer and CNN, namely Swin Transformer and ConvNeXt, in the task of classifying two class benign and malignant skin cancer images. Both models use pretrained from ImageNet and are applied with three different fine-tuning strategies, namely Linear Probe (LP), Full Fine-Tuning (FT), and a combination of the two previous strategies (LP-FT). The dataset used is the ISIC Archive Dataset with an 80:20 data split for training and validation, consisting of 3.297 images divided into two classes, with 1800 benign images and 1.497 malignant images. The evaluation was performed using the accuracy, precision, recall, and F1-score metrics. Swin Transformer with the LP-FT strategy achieved the best performance, with an accuracy of 92,27%, precision of 92,24%, recall of 92,17%, and an F1-score of 92,20%. These findings indicate that the two-stage fine-tuning approach can improve model stability and generalization, as well as contribute to the development of a more accurate artificial intelligence based skin cancer diagnosis system.

Analisis Pola Temporal Penyebaran Penyakit DBD dan HIV Berbasis Time Series Clustering

2026-03-08T00:33:18+07:00

In Indonesia, including in East Java Province, infectious diseases such as Dengue Fever (DHF) and Human Immunodeficiency Virus (HIV) remain public health concerns. Incidence patterns vary by region and time of year. Variations in temporal patterns among districts and cities may lead to suboptimal identification of priority intervention areas when analyses rely solely on absolute case counts. This study aims to analyze the temporal patterns of DHF and HIV case distribution in East Java Province during the 2018–2024 period in order to cluster regions based on similarities in case dynamics over time.The analysis was conducted using a time series clustering approach to group districts and cities according to the similarity of their case development patterns. Temporal similarity was measured using the Dynamic Time Warping method and subsequently clustered using Hierarchical Clustering. Prior to analysis, the data were normalized using the Z-score method to minimize the influence of differences in case scale among regions. The results show that the temporal patterns of DHF and HIV cases were each classified into three main clusters. Cluster quality evaluation using the Silhouette index yielded a value of 0.408 for DHF, indicating a relatively clear cluster structure, whereas a value of 0.197 was obtained for HIV, suggesting a weaker cluster structure due to the complexity and heterogeneity of regional-level case data. Nevertheless, the resulting clusters still provide preliminary information on variations in temporal patterns. The identified clusters represent regions with stable, fluctuating, and increasing case patterns. Several urban areas, such as Pasuruan City, Probolinggo City, and Banyuwangi Regency, tend to belong to clusters with relatively high case levels for more than one disease, indicating challenges in disease control within these regions. These findings provide an initial overview of the temporal dynamics of DHF and HIV cases in East Java, which may serve as supporting evidence for region- and time-based disease control planning.

Analisis Sentimen Diseminasi Produk Iklim Menggunakan Metode Recurrent Neural Network (RNN) dalam Klasifikasi dan Density-Based Spatial Clustering of Applications with Noise (DBSCAN) untuk Klasterisasi

2026-03-08T00:56:29+07:00

Climate change and extreme weather events have a significant impact on various sectors of life, making the accurate and timely dissemination of climate information crucial. Public sentiment can be an indicator of public assessment of climate dissemination. The implications of the sentiment analysis itself can be used as a communication strategy from information providers to the public. This study aims to analyze public sentiment toward the dissemination of climate products by the Central Java Climatology Station through social media platforms Instagram (@bmkgjateng) and X (@bmkg_semarang). The analysis was conducted using a hybrid framework integrating the Recurrent Neural Network (RNN) method for sentiment classification and the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) for topic clustering and outlier identification. A total of 12,847 comments were collected via web scraping from 2020 to 2024. The RNN classification results revealed a dominance of neutral responses (76.41%), followed by negative (13.15%) and positive (10.44%) sentiments. The model achieved high performance with 96% accuracy and a weighted average F1-Score of 0.96. DBSCAN successfully identified 82 topic clusters and classified 74.5% of the data as noise, largely consisting of non-topical interactions or spam. The validity of the cluster structure was confirmed by a Silhouette Coefficient of 0.3675, a Davies-Bouldin Index of 0.504, and a Calinski-Harabasz Index of 191.395, indicating that the formed topic clusters possess a robust structure and are distinctly separated from one another. Integrative analysis revealed that negative sentiments were consistently focused on specific issue clusters such as floods and extreme heat, whereas positive sentiments were dispersed across service appreciation. These findings suggest the necessity of implementing an automatic filtration

Performance Analysis of Quantum Long Short-Term Memory (QLSTM) Models for TLKM Stock Price Prediction

2026-03-08T01:06:18+07:00

Stock price prediction is a challenging task due to its nonlinear, dynamic, and temporal characteristics, yet accurate forecasting models are crucial for decision-making in volatile stocks such as PT Telkom Indonesia Tbk (TLKM). Despite the rapid adoption of AI-based forecasting methods, several research gaps remain. Empirical studies on Quantum Long Short-Term Memory (QLSTM) are still relatively limited compared to classical LSTM variants, particularly for emerging market datasets. Existing research also tends to emphasize architectural comparisons rather than systematically analyzing training configurations. The joint effects of optimizer selection, epoch number, and hidden unit size on QLSTM performance have not been comprehensively evaluated, and many studies rely on limited evaluation metrics, reducing the strength of robustness assessment. To address these gaps, this study applies a QLSTM model to predict stock opening prices using historical time-series data and systematically evaluates the impact of different optimizers. The model is trained using Adam, Nadam, RMSprop, and SGD with epoch variations (50–250) and hidden units (8, 16, 32). Performance is measured using accuracy, MAE, MSE, RMSE, MAPE, and R² to ensure a comprehensive evaluation. The results indicate that adaptive optimizers consistently outperform SGD, with Adam providing the most stable and accurate predictions, highlighting the importance of optimizer choice and hyperparameter configuration in QLSTM-based stock forecasting.

Implementasi Deep Learning Berbasis MobileNetV2 untuk Deteksi Real-Time Bacterial Spot dengan Pendekatan Arsitektur Lightweight

2026-03-08T21:48:35+07:00

Bacterial spot caused by Xanthomonas campestris pv. vesicatoria is a critical disease in bell peppers that can reduce productivity by up to 50%. This study implements MobileNetV2 with two-stage transfer learning for real-time bacterial spot detection using lightweight architecture approach, with ResNet50 as baseline comparison. PlantVillage dataset (2,475 images) was used for training and in-domain evaluation, while India dataset (132 images) for domain shift assessment. Results demonstrate MobileNetV2 achieves 98.66% accuracy on PlantVillage test set, outperforming ResNet50 (89.78%) by 8.88 percentage points despite being 9.2× lighter (2.7 MB vs 24.3 MB TFLite) and 2.0× faster (22.4 ms vs 45.8 ms inference time). MobileNetV2 efficiency advantage is also evident in its inference memory footprint of only 107 MB RAM, significantly 2.3x lower than ResNet50(242 MB RAM), making it highly suitable for deployment on mid-range smartphones with limited RAM. External dataset evaluation reveals MobileNetV2 maintains superior robustness with 65.3% retention rate versus ResNet50's 52.3%. Trade-off analysis positions MobileNetV2 on the Pareto frontier, achieving optimal accuracy-efficiency sweet spot for plant disease detection applications. This research contributes empirical evidence for lightweight architecture superiority, comprehensive efficiency-oriented evaluation framework, ULTRA-LIGHT training strategy for addressing inverse overfitting, and realistic generalization assessment using tropical external dataset.

Optimizing Ensemble Learning Models with SMOTE-ENN for Early Stroke Detection in Imbalanced Clinical Datasets

2026-03-08T21:58:42+07:00

Stroke remains a leading cause of mortality and long-term disability worldwide, including in Indonesia, highlighting the urgent need for early risk identification. Machine learning models for stroke prediction often suffer from severe class imbalance, where stroke cases constitute only 4.9% of clinical datasets, leading to biased predictions that favor the majority class. This study evaluates three ensemble and kernel-based algorithms Random Forest, XGBoost, and Support Vector Machinecombined with two resampling strategies (SMOTE and SMOTE-ENN) using the Healthcare Stroke Dataset (5,110 records, 11 clinical attributes). To prevent data leakage, resampling was strictly applied within each training fold of 5-fold stratified cross-validation, while all evaluations were conducted on the original imbalanced test set. The results demonstrate that XGBoost integrated with SMOTE-ENN achieved the highest minority-class sensitivity, improving PR-AUC by 23.5% (0.1537 vs. 0.1244 with SMOTE alone), while detecting 24% of stroke cases (12 out of 50) in the test set. Although cross-validation results indicate strong class discrimination with AUC-ROC values above 0.98, the low PR-AUC reflects the operational challenge of extreme class imbalance and the inevitable trade-off between recall and precision, resulting in an increased number of false positives. Consequently, the proposed model is best positioned as a first-tier population screening tool that flags high-risk individuals for confirmatory clinical diagnostics, rather than as a standalone diagnostic system. The approach maintains computational efficiency (training time < 0.12 seconds) and substantially improves model stability, evidenced by a 73% reduction in cross-validation variance. These findings support the integration of hybrid resampling techniques with ensemble learning as a practical and scalable framework for early stroke risk screening in resource-constrained primary healthcare settings.

Klasifikasi Kesehatan Mental Menggunakan Support Vector Machine Berdasarkan Screen Time dan Interaksi Sosial Digital

2026-03-08T22:15:54+07:00

Mental health is an important aspect that influences the quality of life of individuals, especially in adolescents and young adults who are vulnerable to stress due to the increased use of digital devices. Technological developments have led to increased screen time and the intensity of digital social interactions, which have the potential to affect mentsal health conditions. This study aims to develop a mental health classification model using the Support Vector Machine (SVM) method with a Radial Basis Function (RBF) kernel based on digital behavior data, including daily device usage time, social media time, number of positive interactions, and number of negative interactions. The dataset used is secondary data obtained from Kaggle and goes through the stages of pre-processing, feature selection, data normalization, and division of training and test data with a ratio of 80:20. The built SVM model is able to classify mental health conditions into three classes, namely Healthy, Stressed, and Risky. The evaluation results show that the accuracy of the resulting model is 94.3%, with a precision value of 66.3%, a recall of 96.1%, and an f1-score of 74.1%. These results indicate that the variables of screen time and digital social interaction have strong potential to be used as a basis for objective and data-based mental health classification.

Deteksi Cyberbullying pada Komentar Media Sosial Berbahasa Indonesia Menggunakan Pendekatan Hibrida IndoBERTweet- BiLSTM

2026-03-08T22:23:59+07:00

Cyberbullying on Indonesian-language social media has become a serious issue with significant psychological and social consequences, necessitating the development of reliable automated detection systems. However, the informal, ambiguous, and highly contextual nature of social media language, including the frequent use of slang and sarcasm, poses substantial challenges for conventional text classification approaches. This study proposes a hybrid cyberbullying detection model that integrates the domain-specific pre-trained language model IndoBERTweet with a Bidirectional Long Short-Term Memory (BiLSTM) architecture. IndoBERTweet is employed to generate contextualized semantic representations aligned with the linguistic characteristics of Indonesian Twitter data, while BiLSTM is utilized to capture bidirectional sequential dependencies at the sentence level. Experiments were conducted using a publicly available, manually annotated Indonesian Twitter dataset consisting of 13,091 samples, which were reformulated into a binary classification scheme. To address class imbalance, a combination of class weighting and label smoothing was applied during model training. Model performance was evaluated using Accuracy, Precision, Recall, F1-Score, ROC-AUC, and PR-AUC metrics. Experimental results show that the IndoBERTweet–BiLSTM model achieved the best performance with an F1-Score of 87.53%, Recall of 88.80%, Precision of 86.31%, ROC-AUC of 92.91%, and PR-AUC of 94.25%. This performance consistently outperforms baseline models based on IndoBERT and IndoBERT-p1 with identical architectural configurations. These findings highlight the critical role of domain alignment in enhancing cyberbullying detection performance for Indonesian social media text.

Komparasi Algoritma Naive Bayes dan K-Nearest Neighbor untuk Analisis Sentimen Pengguna Dompet Digital pada Google Play Store

2026-03-08T22:43:54+07:00

The rapid growth of digital wallet users in Indonesia, reaching millions of active users, has generated a massive volume of reviews on the Google Play Store. This textual data contains crucial insights regarding customer satisfaction but is often underutilized due to challenges in processing unstructured data. This study aims to perform a comparative performance analysis between the probabilistic Naive Bayes algorithm and the distance-based K-Nearest Neighbor (KNN) in classifying user sentiment for DANA, OVO, DOKU, and LinkAja applications. This study utilizes a dataset of 18,869 reviews which exhibits a mild class imbalance with a negative sentiment dominance of 57.54%. To preserve the representation of the large original data, this research applies Stratified Sampling without synthetic data balancing techniques (such as SMOTE), followed by comprehensive preprocessing stages aided by the Sastrawi library and Term Frequency-Inverse Document Frequency (TF-IDF) feature extraction. Model optimization was systematically conducted using GridSearchCV for Naive Bayes and the Elbow Method to determine the optimal k value for KNN. Empirical test results show that the Naive Bayes algorithm with a smoothing parameter alpha of 0.1 achieved the best performance with an accuracy of 88.5% and an AUC of 0.9237, outperforming KNN at k=27 which obtained an accuracy of 87.4%. The validity of this performance difference was confirmed to be significant through the McNemar statistical test with a p-value of 0.0045. Another crucial finding is computational efficiency, where Naive Bayes proved to be 129 times faster in the prediction process compared to KNN. Based on the significant advantages in accuracy and time efficiency, Naive Bayes is recommended as the superior method for real-time sentiment analysis in the financial technology ecosystem.

Optimasi Deteksi Intrusi Jaringan Menggunakan Hybrid Model Autoencoder dan Random Forest

2026-03-08T23:23:52+07:00

Conventional Intrusion Detection Systems often suffer from performance degradation due to their inability to handle the complexity of high-dimensional data and class imbalance in modern network traffic. This study aims to optimize the Network Intrusion Detection System (IDS) by addressing the limitations of the Random Forest algorithm in handling high-dimensional data and its lack of model transparency (black-box). The proposed method is a Hybrid model integrating an Autoencoder as a non-linear feature extractor and Random Forest as a classifier. The Autoencoder is trained using a semi-supervised strategy to generate latent features and Reconstruction Error (MSE), which serves as a robust anomaly indicator. Additionally, the Synthetic Minority Over-sampling Technique (SMOTE) is applied to address class imbalance in the NSL-KDD dataset. To address the challenge of interpretability, SHAP-based Explainable AI (XAI) is strategically implemented to elucidate the complex interactions between the Autoencoder-compressed latent features and the final classification decisions, thereby transforming this hybrid architecture into a transparent system. Evaluation results demonstrate that the Hybrid Autoencoder-Random Forest model outperforms the Random Forest Baseline, achieving an Accuracy increase of 2.54% (to 77.61%) and a Recall increase of 3.96% (to 62.31%). The significant improvement in the Recall metric empirically validates the effectiveness of hybrid features, specifically the Reconstruction Error, in detecting Zero-Day attacks characterized by unknown patterns. Furthermore, SHAP visualization successfully reveals the contribution of latent features, providing crucial transparency for network security forensic analysis.

Deep Fake Image Detection Using Vision Transformer with Random Oversampling Technique

2026-03-08T23:31:47+07:00

Recent developments in deep learning have facilitated the generation of visually convincing deepfake images, creating serious concerns for the reliability and security of digital media content. The primary challenge lies in detecting these sophisticated manipulations while handling imbalanced datasets, a common issue in deepfake detection research. This research focuses on designing a robust deepfake image classification model based on the Vision Transformer (ViT) architecture to differentiate between authentic and manipulated images. The main objectives are to: (1) adapt and fine-tune a pre-trained Vision Transformer for binary classification, (2) evaluate the effectiveness of Random Oversampling in addressing class imbalance while preventing data leakage, and (3) assess model performance using comprehensive metrics. Methods: A pre-trained Vision Transformer model (Deep-Fake-Detector-v2-Model) was adapted and fine-tuned using a dataset consisting of 190,335 images. To overcome the issue of class imbalance, a Random Oversampling strategy was applied exclusively to the training set after dataset splitting to prevent data leakage. The dataset was divided into training and testing subsets using an 80:20 ratio. During the training phase, data augmentation techniques such as image rotation, sharpness variation, and pixel normalization were employed. The model was trained for four epochs with a learning rate of 1×10⁻⁶ and a batch size of 32. Results: Experimental evaluation demonstrates that the proposed model achieves a classification accuracy of 94.46% on the test dataset. The model demonstrates high precision of 97.56% for fake images and 91.74% for real images, with corresponding recall rates of 91.21% and 97.72% respectively. The F1-score reaches 94.46% for both classes, indicating balanced performance. Novelty: This research presents a novel application of Vision Transformer architecture for deepfake detection, combining efficient transfer learning with strategic oversampling to handle imbalanced datasets while preventing data leakage. The study demonstrates that ViT-based models can effectively capture subtle manipulation artifacts in deepfake images, achieving superior performance compared to traditional convolutional neural network approaches.

Optimasi Hyperparameter Random Forest untuk Klasifikasi Depresi Mahasiswa Menggunakan GridSearchCV dan RandomizedSearchCV

2026-03-08T23:46:51+07:00

Student mental health is an important issue that requires a data-driven approach to support the classification process of student depression. This study aims to analyze the factors that cause depression and optimize the performance of the classification model by applying the Random Forest algorithm. The data used in this research is secondary data from the Student Depression Dataset obtained from the Kaggle platform, with a total of 27,901 data points. The research stages begin with data collection followed by Exploratory Data Analysis (EDA), which includes descriptive statistical analysis and correlation between variables using a heatmap. Data preprocessing involves removing irrelevant features, handling missing values, encoding categorical data, and splitting the data into training and testing sets. Model development is carried out through three scenarios: a baseline model, hyperparameter optimization using GridSearchCV, and RandomizedSearchCV. Model performance evaluation is measured using a Confusion Matrix to analyze accuracy, precision, recall, and F1-score. The results show that all models produce relatively stable accuracy in the range of 0.84–0.85. The model with GridSearchCV optimization provides the best performance with a recall value of 0.8869 and an F1-score of 0.8719. This increase in recall is important to minimize the risk of false negatives in identifying students experiencing depression. It is hoped that these findings can contribute as a decision support system for educational institutions in more accurately detecting and managing students' mental health.

Perbandingan Kinerja Algoritma CatBoost, XGBoost, LightGBM dan Random Forest Dalam Memprediksi Risiko Infeksi Aids Dalam Dataset Kesehatan

2026-03-09T00:04:52+07:00

This study investigates the prediction of AIDS infection risk using tree-based algorithms CatBoost, XGBoost, LightGBM, and Random Forest applied to a medical and demographic dataset consisting of 2,139 observations and 23 variables. The research process includes data exploration, cleaning, handling extreme values using the interquartile range (IQR) method, normalization with RobustScaler, and class balancing using the Synthetic Minority Over-sampling Technique (SMOTE). Due to the imbalanced nature of the dataset, model evaluation emphasizes not only accuracy but also Recall, F1-Score, and AUC-ROC to better assess infected class detection. Prior to SMOTE implementation, all models achieved high accuracy but relatively low recall for the positive class; after resampling, CatBoost demonstrated the most significant improvement, with recall increasing from 63% to 77% and F1-Score from 72% to 79%, achieving an overall accuracy of 90%. In comparison, XGBoost reached an accuracy of 88.63% with a more moderate recall improvement, while LightGBM and Random Forest showed consistent yet smaller gains, indicating that the combination of SMOTE and CatBoost is more effective in minimizing False Negatives in AIDS infection cases. The main contribution of this study lies in the integration of robust outlier handling, feature normalization, and class balancing within a structured experimental framework, with a specific emphasis on sensitivity optimization to enhance early detection reliability in clinical screening contexts.

Analisis Kinerja Decision Tree dan Naïve Bayes Pada Klasifikasi Tingkat Kepuasan Masyarakat

2026-03-09T00:18:45+07:00

The Public Satisfaction Survey (SKM) is an official instrument used by the government to evaluate public service performance as stipulated in Regulation of the Minister of State Apparatus Empowerment and Bureaucratic Reform (PermenPANRB) Number 14 of 2017. However, the use of SKM data in many government agencies is still limited to calculating satisfaction index values without further predictive analysis. This study aims to classify the level of satisfaction of service users of the Metro City Investment and Integrated Services Agency (DPMPTSP) using the Decision Tree and Naïve Bayes algorithms. The data used is SKM data from 2025 to the fourth quarter, consisting of 2,760 respondents, which consists of nine service elements (U1–U9) with satisfaction categories as class variables. The research process includes data pre-processing, classification modeling using RapidMiner, and model evaluation based on confusion matrix, accuracy, precision, and recall. The results showed that the Naïve Bayes algorithm produced an accuracy rate of 95.04%, higher than the Decision Tree, which obtained an accuracy of 84.46%, and had a better recall value in the dominant class (recall of the Satisfied class was 98.16%). These advantages demonstrate the efficiency of the Naïve Bayes probabilistic approach in handling categorical features in public service elements. This study proves that the application of Data mining techniques to SKM data can support data-based public service evaluation.

A Comparative Study of LSTM and BiLSTM Performance in Predicting XAU/USD Prices

2026-03-09T23:26:09+07:00

Gold price forecasting in the XAU/USD market is challenging due to nonlinear dynamics, high volatility, and sensitivity to global macroeconomic factors. This study compares the performance of Long Short-Term Memory (LSTM) and Bidirectional Long Short-Term Memory (BiLSTM) architectures in forecasting XAU/USD closing prices using historical data from 2023–2026. Data preprocessing includes cleaning, chronological ordering, normalization, and transformation using a sliding window approach. A window size of 60 time steps is selected to represent approximately three months of daily trading activity, enabling the models to capture short- to medium-term temporal dependencies while limiting excessive noise and computational burden. The dataset is divided chronologically into training and out-of-sample testing sets to ensure proper generalization assessment. Both models employ identical architectures with two recurrent layers (50 hidden units each) and are trained using the Adam optimizer with epoch variations (20–100). Evaluation on unseen test data uses MAE, MSE, RMSE, MAPE, and R² metrics. LSTM achieves its lowest MAE of 21.26 at 40 epochs, while BiLSTM attains its best performance at 80 epochs with an MAE of 20.86 and R² of 0.9981. However, extending training to 100 epochs leads to performance degradation in BiLSTM, indicating sensitivity to overtraining. Overall, optimal performance is achieved through balanced training duration rather than increased architectural complexity.

Market-Adaptive Stock Trading through B-WEMA Driven Proximal Policy Optimization

2026-03-09T23:36:12+07:00

Developing automated trading strategies that achieve stable returns while controlling risk remains a central threat in quantitative finance. Many reinforcement learning-based trading systems focus on reward maximization but provide limited justification for the choice of forecasting indicators and often lack comprehensive benchmarking against alternative strategies and risk measures. This essay addresses the problem of integrating a statistically grounded price-smoothing technique with a policy optimization scheme to improve sequential trading decisions under market uncertainty. We propose a hybrid model that combines Brown’s Weighted Exponential Moving Average (B-WEMA) as a trend-sensitive forecasting indicator with a Deep Reinforcement Learning agent trained using Proximal Policy Optimization (PPO). The role of B-WEMA is to provide structured price signals that reduce noise sensitivity, while PPO determines buy and sell actions through policy updates constrained for stable learning. The performance of the proposed model is evaluated over a 10-month trading horizon and compared with a buy-and-hold benchmark and an alternative reinforcement learning method, Advantage Actor-Critic (A2C), both implemented under the same experimental conditions. Empirical results show that the proposed B-WEMA-PPO framework achieved a cumulative return of 23.43% over the test period, outperforming both the benchmark and the A2C-based agent. In addition to cumulative return, risk-adjusted performance metrics, namely volatility and maximum drawdown, are reported to provide a balanced assessment of profitability and risk exposure. These findings suggest that incorporating structured exponential smoothing into policy optimization may enhance the stability and effectiveness of reinforcement learning-based trading strategies.

Implementasi Arsitektur MobileNetV2 Berbasis Citra untuk Deteksi Penyakit Dropsy dan Popeye pada Ikan Cupang

2026-03-09T23:43:31+07:00

The identification of diseases in betta fish based on visual symptoms remains a challenge, particularly for beginners who lack experience in recognizing disease characteristics. This study aims to implement an image-based MobileNetV2 architecture as a diagnostic support system to detect dropsy and popeye diseases in betta fish that have already exhibited visual symptoms. The dataset used in this study consists of 600 betta fish images divided into three classes: healthy, dropsy, and popeye, with 200 images in each class, collected from the internet. Data preprocessing was conducted through image ratio adjustment, normalization, and data augmentation to increase data variability. A transfer learning approach was applied by freezing most layers of the MobileNetV2 feature extractor and fine-tuning several of the final layers. Model evaluation was performed using 5-Fold Cross Validation to ensure experimental stability and reproducibility. The best model from each fold was then combined using an ensemble method based on average probability to improve prediction performance on the test dataset. Experimental results show that the average 5-Fold Cross Validation accuracy reached 74.71% with a standard deviation of ±4.57%, while the Macro-F1 score achieved ±74.43%. The ensemble approach produced a test accuracy of 85.56% with balanced classification performance across all classes. Grad-CAM visualizations indicate that the model is able to focus on image regions relevant to disease symptoms. These findings demonstrate that the MobileNetV2 architecture is effective as an image-based diagnostic support tool for betta fish diseases.

Analisis Komparatif Arsitektur Convolutional Neural Network untuk Klasifikasi Kualitas Cabai dengan Implementasi Perangkat Mobile

2026-03-09T23:58:39+07:00

Manual chili quality sorting is susceptible to subjectivity and inter-assessor inconsistency, which can reduce product market value. This study conducts a comparative analysis of three Convolutional Neural Network (CNN) architectures—Custom CNN, MobileNetV3-Small, and EfficientNetV2-B0 for binary chili quality classification (Good/Bad) using a primary dataset of 1,383 chili images (684 Good-class, 699 Bad-class) captured with a smartphone camera. The Good class includes chili with a smooth surface, fresh color, and no decay spots, while the Bad class includes chili showing signs of decay, physical defects, or deformation. Evaluation was conducted based on accuracy, precision, recall, F1-Score, AUC, inference time, and post-quantization model size. The results show that EfficientNetV2-B0 achieved the highest accuracy of 92.0% (precision 92.4%, recall 92.0%, F1-Score 92.0%, AUC 0.961), MobileNetV3-Small obtained an accuracy of 87.7% with the lowest server-side inference latency (2.39 ms), and Custom CNN achieved 87.3% accuracy with the most compact model size (118 KB post-quantization). All three models were integrated into a Flutter-based Android application prototype as a proof-of-concept, displaying the classification result (Good/Bad), confidence score, and inference latency, with end-to-end response times ranging from 80 to 120 ms on a Xiaomi 13T device. This study contributes empirical comparative data on three CNN architectures in the chili quality classification domain, accompanied by the construction of a local dataset and technical validation of model deployment on a mobile device. The results of this study are expected to serve as a reference in selecting CNN architecture for the development of a mobile-based chili quality classification system, particularly as a first step toward the implementation of simple small-scale sorting at the farmer level.

Machine Learning Comparative Analysis of SVR Method with RBF Kernel and Random Forest for Bitcoin Price Prediction

2026-03-10T00:08:48+07:00

This study aims to determine how accurate machine learning predictions are for predicting Bitcoin prices using the SVR With RBF Kernel and Random Forest methods. This study was conducted because Bitcoin’s volatility is so high that it is difficult to predict. Therefore, this study uses two different methods to allow for a more objective evaluation of model characteristics on volatile data. The dataset was obtained through Kaggle with a Bitcoin price dataset from 2018 to October 2025, totaling 2,856 datasets in CSV format. After training both methods on the same dataset, price prediction results were obtained. Support Vector Regression (SVR) With RBF Kernel achieved a relatively high data evaluation result with an MAE of 10866.882878735294, MSE of 204836847.5591309, and RMSE of 14312.12239883138, while the Random Forest method achieved a low data evaluation result with an MAE of 19342.47, MSE of 659671833.13, and RMSE of 25684.08. The result of these two methods show a significant difference, with Random Forest more closely aligning with the acual data, with a lower evaluation value and producing values closer to the actual data. This research was conducted to determine the accuracy of the Support Vector Regression (SVR) with RBF Kernel and Random Forest algorithms. It is concluded that both methods make good predictions, only the Random Forest method is closer to the actual Bitcoin price.

Comparative Analysis of K-NN and Naïve Bayes Algorithms for Early-Stage Chronic Kidney Disease Classification

2026-03-10T00:37:07+07:00

Chronic Kidney Disease (CKD) is a global health issue characterized by low early detection rates and high diagnostic costs. Artificial intelligence, particularly machine learning, offers a promising solution as a rapid and cost-effective decision support system. This study aims to comprehensively analyze and compare the performance of two simple and interpretable classification algorithms, K-Nearest Neighbor (K-NN) and Naïve Bayes (NB), for predicting CKD based on clinical data. The dataset was sourced from the UCI Machine Learning Repository, comprising 400 instances and 25 clinical attributes such as blood pressure and serum creatinine. The methodology included data preprocessing (median imputation for numerical features, mode imputation for categorical features), encoding, Min-Max normalization, data splitting (70:30 ratio), model training, K parameter optimization for K-NN via 5-fold cross-validation, and evaluation using accuracy, precision, recall, F1-Score, and Confusion Matrix metrics. Experimental results demonstrated that the Naïve Bayes algorithm achieved superior performance with an accuracy of 95.83%, precision of 95.95%, recall of 97.26%, and F1-Score of 96.60%. The K-NN algorithm with an optimal K=5 attained an accuracy of 91.67%. Statistical analysis using a paired t-test (α=0.05) with p-value=0.012 confirmed that this performance difference was significant. It is concluded that Naïve Bayes is more effective for this CKD dataset, likely due to its robustness in handling feature independence assumptions and varied data scales. This model holds strong potential for development into an early-stage CKD screening tool to assist healthcare professionals.

Analisis Sentimen Persepsi Publik Terhadap Program MBG Pada Komentar YouTube Menggunakan Naïve Bayes dan Resampling

2026-03-10T00:44:48+07:00

The Free Nutritious Meal Program (MBG), launched by the Indonesian government in 2025, has generated diverse public responses on social media, particularly on YouTube as an open digital discussion space. This study aims to analyze public perception of the MBG program through sentiment classification of YouTube comments using the Multinomial Naïve Bayes algorithm combined with Term Frequency–Inverse Document Frequency (TF-IDF) weighting. The dataset consists of 1,082 comments categorized into three sentiment classes: negative, neutral, and positive. The data distribution reveals significant class imbalance, with negative sentiment dominating at 70.61%. The baseline model achieved an accuracy of 70.67% with a macro F1-score of 27.60%, indicating bias toward the majority class. To address this imbalance, Random Oversampling (ROS) and Synthetic Minority Over-sampling Technique (SMOTE) were applied. Although overall accuracy decreased to approximately 51% after resampling, the macro F1-score improved to 36.24% (SMOTE) and 37.09% (ROS), indicating enhanced performance in detecting minority classes. In the context of public policy evaluation, improved sensitivity to minority sentiment is considered more representative than high but biased accuracy. These findings highlight the importance of handling class imbalance in social media–based sentiment analysis for public policy monitoring.

Analisis Komparatif Kinerja Algoritma Support Vector Machine, Random Forest, dan Naive Bayes untuk Klasifikasi Sentimen pada Komentar YouTube

2026-03-19T22:11:37+07:00

The rise of social media platforms like YouTube has made them a primary medium for public discourse on socio-political issues, such as the "August 25th protests," which triggered massive polarization in the digital space. The vast volume of comments necessitates a computational approach for sentiment analysis. This study aims to classify public sentiment into positive and negative categories while comparing the performance of Naive Bayes, Random Forest, and Support Vector Machine (SVM). These algorithms were selected for their computational efficiency on high-dimensional text data compared to Deep Learning models. The methodology involved collecting 2,917 comments via the YouTube Data API v3, followed by text preprocessing, lexicon-based automated labeling, and TF-IDF feature weighting. To address the dataset's imbalance, where negative sentiment dominated at 78.8%, stratified sampling was applied to maintain class proportions. Results indicate that SVM achieved the highest accuracy at 88.2%, outperforming Random Forest (83.1%) and Naive Bayes (81.2%). SVM's superiority stems from its ability to find an optimal hyperplane that maximizes class margins, ensuring stability in imbalanced datasets. This research contributes a robust classification framework for understanding public opinion dynamics on specific political issues in Indonesia.

Comparison of Geothermal Well Productivity Using KNN, SVM and Gradient Boost Methods

2026-03-19T22:10:21+07:00

Manual and conventional processing of geothermal well production data is computationally inefficient and requires several hours to days to generate productivity assessments, particularly when dealing with large-scale and non-linear operational datasets. The complexity of geothermal production parameters this paper such as wellhead pressure (WHP), enthalpy, steam flow, brine flow, total flow, and generated power this paper creates challenges for accurate and timely productivity classification at the well level. This study utilizes 74,912 daily production records collected from January 2018 to June 2024, comprising 13 operational and production-related attributes. The objective is to identify the most effective machine learning algorithm for classifying geothermal well productivity levels to support faster and more reliable operational decision-making. A comparative machine learning classification approach was conducted using K-Nearest Neighbors (K-NN), Support Vector Machine (SVM), and Gradient Boosting. Model evaluation was performed using three train–test split ratios: 70:30, 80:20, and 90:10. Two modelling scenarios were implemented: with and without Synthetic Minority Over-sampling Technique (SMOTE) to address class imbalance. The results indicate that the K-NN model achieved the highest classification performance, reaching 94.22% accuracy using the 90:10 split ratio without SMOTE. Gradient Boosting demonstrated stable performance across all ratios, with its best accuracy of 91.39% at the 70:30 split without SMOTE. In contrast, SVM produced the lowest performance, with a maximum accuracy of 79.78% at the 90:10 ratio without SMOTE. The application of SMOTE improved minority class recall, particularly for SVM, but generally reduced overall model accuracy for K-NN and Gradient Boosting. These findings demonstrate that classical machine learning algorithms, particularly K-NN, provide an efficient and accurate solution for geothermal well productivity classification. The proposed approach significantly reduces processing time compared to conventional analytical methods and supports data-driven decision-making in geothermal production forecasting and development planning.

Optimasi Deteksi Malware Android pada Dataset Drebin Menggunakan Ensemble Learning

2026-03-19T22:32:35+07:00

The increasing number and complexity of Android malware require detection systems that are accurate, efficient, and capable of handling high-dimensional data. Machine learning–based approaches have become one of the widely adopted solutions in cybersecurity research. However, the performance of classification models is often affected by feature redundancy and suboptimal hyperparameter configurations. This study aims to evaluate the effectiveness of combining Random Forest–based feature selection with modern boosting classification algorithms for Android malware detection. The dataset used in this study is the Drebin 215 dataset, which was selected because it is one of the most widely used benchmark datasets for Android malware detection based on static analysis, enabling more objective comparison with previous studies. Feature selection was performed using the Random Forest feature importance method to reduce data dimensionality prior to the classification stage. The classification models employed include XGBoost, Light Gradient Boosting Machine (LightGBM), and CatBoost. The experiments were conducted under two scenarios: without hyperparameter optimization (non-tuning) and with hyperparameter optimization using the Grid Search method. Model performance was evaluated using accuracy, precision, recall, F1-score, and ROC-AUC metrics, as well as computational time analysis. The experimental results show that all models achieved very strong classification performance on the Drebin benchmark dataset, with accuracy values exceeding 0.98. Among the evaluated models, LightGBM achieved the best performance, with an accuracy of 0.9900 and an F1-score of 0.9865. This performance advantage is likely influenced by the efficiency of its histogram-based learning mechanism and leaf-wise tree growth strategy, which enables faster and more effective learning on high-dimensional data. Nevertheless, the high performance observed on this benchmark dataset still requires further evaluation on more diverse datasets or dynamic environments to ensure the generalization capability of the model in real-world scenarios. The findings of this study indicate that the combination of Random Forest–based feature selection and boosting algorithms can serve as an effective approach for improving the efficiency and performance of Android malware detection systems.

Analisis Ketahanan Model ResNet-50 pada Klasifikasi Bahasa Isyarat Arab terhadap Degradasi Citra Bawah Air

2026-03-19T22:44:37+07:00

Automatic sign language recognition using deep learning, particularly Convolutional Neural Networks (CNNs), has shown significant potential. The ResNet architecture, through transfer learning, is frequently reported to achieve high accuracy for Arabic Sign Language Alphabet classification under ideal conditions. However, the robustness of these models against real-world visual distortions remains a significant, yet under-explored challenge. This research aims to develop a ResNet-50-based classification model while comprehensively analyzing its robustness. The primary contribution of this research is mapping the tolerance limits and the extent of performance degradation of the ResNet architecture when facing image degradation. Evaluation was conducted on both ideal test data and test data digitally modified to simulate underwater visual effects. This underwater simulation was selected as an extreme stress test scenario because it technically represents an accumulation of simultaneous real-world optical distortions, such as contrast reduction, turbidity (haziness), and light refraction. Quantitative evaluation results show that the model performs excellently with an accuracy of 96.95% under ideal conditions. However, exposure to underwater distortion resulted in an accuracy drop of 4.24%, reducing it to 92.71%. Despite this noticeable performance reduction, the model maintained an F1-Score of 92.79%. These findings provide empirical evidence regarding the capability limits of the ResNet architecture when facing visual degradation, while also emphasizing the importance of robustness testing before deep learning models can be reliably deployed in non-ideal environments full of visual uncertainties.

Perbandingan Kinerja Model ARIMA dan LSTM dalam Peramalan Harga Crypto Solana (SOL-USD) Berbasis Data Yahoo Finance

2026-03-19T22:58:00+07:00

The extreme volatility and non-linear patterns of Solana (SOL) data, driven by its unique consensus mechanism and massive transaction volume, demand accurate forecasting methods to mitigate investment risks. This study compares the statistical method Autoregressive Integrated Moving Average (ARIMA) and Deep Learning Long Short-Term Memory (LSTM) using daily closing price data of SOL-USD from April 2020 to March 2025 obtained from Yahoo Finance. The ARIMA model was developed with optimal parameters (0,1,0), while the LSTM architecture utilized 50 hidden layer units with a 60-day timestep. Evaluation results indicate that the LSTM model significantly outperforms ARIMA, achieving an RMSE of 13.1352 and a MAPE of 6.07% (classified as highly accurate), compared to ARIMA's RMSE of 31.1241 and MAPE of 14.03%. The study concludes that neural network approaches are more effective and adaptive than traditional statistical methods in capturing the highly volatile price dynamics of crypto assets.

Klasifikasi Motif Batik Nitik Berbasis Fitur Ekstraksi SqueezeNet dengan Reduksi Dimensi PCA–LDA

2026-03-19T23:11:38+07:00

Batik nitik motif classification faces significant challenges due to high intra-class variability and complexity of geometric dot patterns, along with limited samples per class in available datasets. Previous research using handcrafted feature extraction methods such as GLCM and MTCD achieved only 53% accuracy, while BSIF with data augmentation reached 97.70%. This study aims to develop a batik nitik classification method using feature extraction based on SqueezeNet trained on ImageNet to achieve superior accuracy without additional external data augmentation techniques. The Batik Nitik 960 dataset consisting of 960 images (60 classes × 16 samples) inherently contains natural visual diversity for each motif as curated by Minarno et al., enabling deep feature extraction from SqueezeNet to be optimized without extra augmentation. A 1000-dimensional feature vector extracted from SqueezeNet's pool10 layer then underwent dimensionality reduction using PCA, LDA, or PCA+LDA, and was classified with Random Forest, SVM, or KNN. These three classifiers were selected to represent distinct learning paradigms: ensemble method (Random Forest), margin-based classifier (SVM), and instance-based learning (KNN), enabling a comprehensive analysis of the extracted feature space characteristics. Experiments were conducted across various training data sizes (4-14 samples per class). Results showed that 8 out of 9 model combinations achieved perfect 100% accuracy, with LDA+SVM, LDA+KNN, PCA+LDA+SVM, and PCA+LDA+KNN requiring only 4 training samples per class. Only LDA+Random Forest failed to reach 100% (maximum 95.14%). The method's advantages lie in the deep feature extraction capability of SqueezeNet, which produces far more discriminative representations than handcrafted features, combined with the efficiency of supervised dimensionality reduction (LDA) in optimizing class separability. Inference time analysis shows that all model combinations are capable of performing predictions within the range of 0.013–0.173 ms per image, and stability evaluation using 5 random states confirms result consistency with mean accuracy ≥99.70% across 8 combinations (standard deviation ≤0.25%), confirming real-time implementation feasibility. This research establishes a new state-of-the-art for the Batik Nitik 960 dataset and opens opportunities for practical applications in authentication, quality control, and preservation of Indonesian batik cultural heritage. The primary contributions of this research encompass the application of SqueezeNet as a fixed feature extractor without fine-tuning for batik nitik classification a previously unexplored approach in this domain a comprehensive comparative analysis of nine dimensionality reduction and classifier combinations, and the establishment of a new state-of-the-art benchmark for the Batik Nitik 960 dataset, validating that CNN-based deep feature extraction surpasses handcrafted methods even with as few as four training samples per class. These findings pave the way for practical real-time batik identification systems applicable to authentication, quality control, and Indonesian cultural heritage preservation

Integrasi Model Hibrida TOPSIS-BORDA untuk Penentuan Prioritas Strategi Digitalisasi Pesantren yang Berkelanjutan

2026-03-19T23:40:49+07:00

Digital transformation in Islamic boarding schools (pesantren) holds a specific urgency as these institutions face the challenge of integrating administrative governance modernization with the preservation of salaf traditions, a dilemma rarely found in general formal education. Resource limitations and preference differences among leaders and administrators often trigger strategic deadlocks. This study aims to determine sustainable digitalization strategies at Pondok Pesantren YPMI Al-Firdaus Semarang by integrating the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) and Borda Count. The TOPSIS model is utilized to objectively evaluate the technical feasibility of seven criteria, while BORDA facilitates quantitative deliberation to accommodate the preferences of five decision-makers. The analysis results indicate that alternative A1 (Digitalization of Administrative Management) becomes the main priority with a relative closeness value of 0.738 and a BORDA consensus score of 8.7664. This figure significantly outperforms other alternatives, proving that the improvement of basic administration is the most urgent and mutually agreed-upon foundation before the pesantren advances to more complex digitalization. Although this study is on a local scale, the integration of these methods proves effective in mapping social and technical compromises, and can be adapted by other pesantren with similar characteristics.

Sentimen Analisis Pengguna Jasa Layanan Kereta Api dengan Menggunakan Metode CNN (Convolutional Neural Network)

2026-03-19T23:53:55+07:00

Train services are a popular mode of transportation in Indonesia, especially in the Greater Jakarta area. However, the quality of train services is often debated among users. This study aims to analyze the sentiment of train service users using the Convolutional Neural Network (CNN) method with a focus on the DAOP 1 Jakarta area. The data used are reviews or comments of train users taken from Indonesian Railways social media. The results of the study show that the CNN method can classify user sentiment analysis with accurate results or high accuracy. This sentiment analysis shows that train users in DAOP 1 Jakarta have positive sentiments towards aspects such as punctuality, service, comfort and safety. The results of this study can help the railway to understand user needs and complaints so that they can improve service quality with a final value of 89.29% accuracy, 88.73% precision, 90.00% recall, and 89.36% F1-score.

Perbandingan Naïve Bayes dan Support Vector Machine Berbasis Term Frequency−Inverse Document Frequency pada Analisis Sentimen Ulasan Produk Afiliasi Lintas Platform TikTok dan Shopee

2026-03-20T01:09:11+07:00

The growth of affiliate marketing on digital platforms, particularly TikTok and Shopee, has led to a rapid increase in consumer reviews that can be leveraged as actionable insights for businesses. However, reviews across platforms exhibit different linguistic characteristics: Shopee reviews tend to be more repetitive and transactional, whereas TikTok reviews are more informal, rich in slang, and noisier. This difference creates a research gap because sentiment classification performance may vary across platforms, while comparative studies on cross-platform affiliate reviews remain limited. This study aims to analyze and compare the performance of Multinomial Naïve Bayes and Support Vector Machine in identifying positive and negative sentiment polarity in TikTok and Shopee affiliate product reviews. Data were collected via web scraping during December 2025–January 2026, yielding 5,502 raw reviews. After text preprocessing (case folding, regex-based cleaning, normalization, stopword removal, and stemming using Sastrawi), 4,593 clean reviews were obtained. Lexicon-based automatic labeling with negation handling produced a binary dataset of 3,314 reviews (2,729 positive and 585 negative), indicating class imbalance; therefore, no data balancing was applied and evaluation emphasized precision, recall, and F1-score in addition to accuracy. Feature representation used Term Frequency–Inverse Document Frequency, and the dataset was split using an 80:20 hold-out scheme (2,651 training and 663 testing instances). Experimental results show that the Support Vector Machine achieved higher performance (95.93% accuracy; 0.81 negative-class F1) than Multinomial Naïve Bayes (89.14% accuracy; 0.12 negative-class F1). This superiority is related to the ability of Support Vector Machine to learn a maximum-margin hyperplane in the high-dimensional and sparse Term Frequency–Inverse Document Frequency feature space, making it more robust to linguistic variation and noise than the probabilistic Naïve Bayes approach, which is more sensitive to majority-class dominance.

Analisis Performa K-Nearest Neighbor dengan Optimasi F1-Score dan Teknik SMOTE dalam Klasifikasi Risiko Serangan Jantung

2026-03-25T23:51:34+07:00

Heart attack is one of the leading causes of death worldwide, making early risk prediction essential for improving patient outcomes. However, many medical datasets suffer from class imbalance, where the number of high-risk cases is significantly smaller than normal cases. This condition may cause machine learning models to be biased toward the majority class and reduce their ability to detect high-risk patients. This study aims to analyze the performance of the K-Nearest Neighbor (KNN) algorithm optimized using F1-score and combined with the Synthetic Minority Over-sampling Technique (SMOTE) for heart attack risk classification. The dataset used is the Heart Attack Dataset, which consists of numerical and categorical features. The research applies an experimental approach by developing a machine learning pipeline that includes data preprocessing, missing value handling, feature standardization, oversampling using SMOTE, and hyperparameter optimization through GridSearchCV with F1-score as the main evaluation metric. Model evaluation is conducted using Stratified 5-Fold Cross-Validation with accuracy, precision, recall, F1-score, and ROC-AUC metrics. The results show that the baseline KNN model achieves an accuracy of 98.50%, precision 95.27%, recall 81.47%, and ROC-AUC 0.9278. Meanwhile, the KNN model integrated with SMOTE attains a recall of 87.27% and ROC-AUC of 0.9484, indicating improved detection of heart attack cases and a reduction in false negatives by 31%, although precision decreases to 72.15%. These findings demonstrate that the integration of SMOTE and hyperparameter optimization effectively improves model sensitivity, making it more suitable for medical applications that prioritize patient safety.

Comparative Analysis of VGG16 Transfer Learning Fine-Tuning Strategies for Automated Concrete Crack Classification

2026-03-20T01:07:22+07:00

Identifying cracks in concrete structures is critical for structural health monitoring, as undetected cracks can lead to catastrophic infrastructure failure. Conventional manual inspections are labour-intensive, subjective, and costly, necessitating automated solutions capable of consistent and scalable deployment. This paper presents a systematic comparative study of four VGG16 transfer learning strategies for automated binary classification of concrete surface cracks. VGG16 was selected for its proven effectiveness in binary image classification tasks, well-established pre-trained feature representations from ImageNet, and low trainable parameter count that reduces overfitting risk on domain-specific datasets. A dataset of 40,000 concrete surface photographs was utilised, divided 80:20 for training and validation. Four training configurations were evaluated: Baseline CNN, Full Freeze, Partial Fine-Tuning, and Full Fine-Tuning, all trained using the Adam optimiser (learning rate 0.001), binary cross-entropy loss, and early stopping. Partial Fine-Tuning achieved the highest accuracy at 99.90%, followed by Full Freeze (99.84%) and Baseline CNN (99.69%). Full Fine-Tuning collapsed to 50.00% due to catastrophic forgetting. The best-performing Partial Fine-Tuning configuration achieved an AUC of 0.9998, precision of 0.9990, recall of 0.9990, and F1-score of 0.9990, with only 15 misclassifications out of 8,000 validation samples. These results confirm that Partial Fine-Tuning is the recommended strategy for concrete crack classification in structural health monitoring application.

Hybrid Music Recommendation System Using K-Means Clustering and Neural Collaborative Filtering for Spotify Playlist Personalization

2026-04-07T04:43:49+07:00

Personalizing music recommendations has become a significant challenge on music streaming platforms such as Spotify due to the vast number of available songs and the limitations of conventional recommendation systems in accurately capturing user preferences. In addition, traditional single-method recommendation approaches often face the cold start problem, which reduces the effectiveness of generated recommendations. Therefore, this study aims to develop and evaluate a hybrid recommendation system that integrates the K-Means Clustering algorithm and Deep Collaborative Filtering based on Neural Matrix Factorization to improve the relevance of music playlist recommendations. The dataset used in this study consists of more than 15,151 Spotify songs obtained from the Spotify dataset available on Kaggle. The dataset was processed through several stages including data inspection, data cleaning, feature selection, and standardization. Audio features used in the analysis include danceability, energy, acousticness, instrumentalness, valence, tempo, and duration. The optimal number of clusters was determined using the Elbow Method and Silhouette Score, resulting in five clusters with a relatively balanced data distribution. The clustering results were then used as the basis for Cluster-Based Filtering to narrow the search space of candidate songs before being processed by the Neural Matrix Factorization model. Performance evaluation was conducted using Hit Ratio at rank 10 and Normalized Discounted Cumulative Gain at rank 10. The proposed model achieved values of 0.1110 and 0.0507, respectively, indicating that the integration of clustering and deep collaborative filtering can improve the effectiveness and personalization of music recommendation systems. This study contributes by proposing a hybrid recommendation framework that integrates clustering-based item grouping with deep collaborative filtering to improve recommendation efficiency and playlist personalization in large-scale music streaming platforms.

Komparasi Model Ensemble dan Algoritma Machine Learning Untuk Memprediksi Penyakit Jantung

2026-03-20T01:05:02+07:00

This study compared the performance of nine machine learning algorithms in predicting heart disease using a dataset dating back to 1988 and consisting of four databases: Cleveland, Hungary, Switzerland, and Long Beach totaling 1025 data. The dataset used includes medical features that reflect physiological states, clinical examination results, and cardiovascular risk factors, namely age, gender, type of chest pain, resting blood pressure, serum cholesterol levels, fasting blood sugar levels, resting electrocardiography results, maximum heart rate, chest pain during physical activity, ST segment depression, ST segment slope, number of major blood vessels visible by fluoroscopy, and thalassemia status. The stages of this study include data cleaning, data transformation, and evaluation carried out using the data splitting method for training and testing as well as K-fold cross-validation with metrics of accuracy, precision, recall, F1 score, and AUC-ROC. The algorithms used in this study are Decision Tree, Random Forest, Support Vector Machine, MLP Classifier, Bagging Classifier, Gradient Boosting, CatBoost, XGBoost, and LightGBM with ensemble-based models, such as CatBoost, Random Forest, XGBoost, and LightGBM, showing consistent performance on various evaluation metrics when compared to non-ensemble models. Among all models tested, CatBoost showed the best performance, with an accuracy reaching 98%, an F1-Score of 0.980, and a Recall of 0.9875 then followed by other ensemble algorithms such as Random Forest, XGBoost and LightGBM. The results of this study indicate that ensemble models are proven to be more effective in predicting heart disease. This study aims to present an in-depth comparative study of the performance of ensemble algorithms and modern machine learning in predicting heart disease, as well as enriching the literature related to the application of Knowledge Discovery in the health sector and providing a basis for selecting more reliable prediction algorithms to support clinical decision making and the development of machine learning-based heart disease diagnosis support systems.

Pemodelan Pola Temporal Action Unit untuk Pengenalan Ekspresi Wajah Berbasis Bidirectional LSTM

2026-03-21T00:49:25+07:00

This study develops a facial expression recognition system based on Facial Action Units (AU) data using a Bidirectional Long Short-Term Memory (BiLSTM) model. The dataset consists of AU data obtained from a supervisor, sourced from DCAP-SWOZ (USC Institute for Creative Technologies), a multimodal corpus containing AU values extracted from human interaction videos. A total of 188 AU files were used in this research. Initial labeling was performed using Facial Action Coding System (FACS)-based rules as pseudo-labels serving as a starting point for training the BiLSTM model. This approach was chosen because the dataset lacks inherent emotion labels, necessitating a label initialization mechanism. The BiLSTM model functions as a temporal smoother designed to reduce noise and label inconsistencies that commonly occur in frame-by-frame rule-based approaches. The trained model then performs inference on the same data to generate final labels with improved temporal stability. Evaluation was conducted by measuring model consistency against FACS rules and qualitative analysis of temporal stability in generated labels. Data were processed into 30-frame sequences with a 1-frame sliding window to effectively capture expression dynamics patterns. The BiLSTM model was trained using two hidden layers with dropout regularization. Evaluation results showed 96.61% consistency against FACS rules with high performance across all emotion classes, including anger (99.11%), disgust (97.98%), fear (94.08%), happiness (99.29%), neutral (96.42%), sadness (98.31%), and surprise (99.16%). Qualitative analysis demonstrated that the model successfully reduced frame-by-frame label fluctuations by 73% compared to pure rule-based approaches, producing more stable and realistic emotion segmentation. These results demonstrate that the combination of FACS-based labeling and the BiLSTM model can produce a temporally consistent automated labeling system capable of accelerating labeled dataset creation, although validation against human ground truth remains necessary as future research.

Hybrid DBSCAN - K-Means Clustering for Financial Loss Identification in INA-CBG Claims Based on Medical Treatment Patterns

2026-03-21T01:10:51+07:00

Hospital financial deficits due to INA-CBG claim discrepancies pose a critical challenge to healthcare sustainability in Indonesia. The difference between hospital operating costs and INA-CBG rates often results in significant financial deficits, which can threaten the sustainability of healthcare providers, especially hospitals. However, existing studies lack a systematic approach to identify distinct patterns of financial losses based on clinical treatment characteristics. This study aims to identify clusters of patients with different financial loss characteristics using a hybrid DBSCAN-K-Means clustering approach based on medical procedure frequency patterns. The DBSCAN algorithm was employed to detect and separate noise from data, while K-Means was used to identify medical treatment patterns. The data were obtained from electronic medical records of inpatients during the 2023–2024 period at a private hospital (N = 6,021 cases). The final clustering results revealed two main clusters with a highly significant difference in deficits between clusters (p = 6.21 × 10⁻³⁸, Cliff's Delta = −0.216). Cluster 0 represents patients with intensive care who have a higher frequency of routine procedures, with an average deficit of 1.51 times (51.3% greater) and an average length of stay of 1.76 times (76% longer) than Cluster 1. Cluster 1 represents patients with a focus on obstetrics/neonatology with a predominance of Doppler procedures. Meanwhile, the noise cluster (13.39%) represents more extreme cases with an average loss of −7.82 million IDR and high mortality. Of the total 315 treatment features, 114 were confirmed to be statistically significant. This study contributes a novel hybrid clustering framework for identifying financial loss patterns in INA-CBG claims, providing actionable insights for hospital management to optimize service utilization, adjust procedure fees for complex cases, and strengthen financial risk management strategies.

Hybrid Entropy and CRADIS Method Approach in Decision Support System for Selecting the Best Employees

2026-03-23T22:43:16+07:00

Selecting the right employees is a key factor in improving organizational performance and productivity. However, in many organizations, the employee selection process is still conducted through manual assessments and subjective judgments, which may lead to bias and inconsistent decisions. Therefore, a systematic and objective approach is needed to support the evaluation process. This study integrates the Entropy method and the CRADIS method within a decision support system to determine the best employee candidates. The Entropy method is applied to calculate objective criteria weights based on the variation of information in the data, while the CRADIS method is used to rank candidates according to their proximity to the ideal solution and distance from the anti-ideal solution. The integration of these two methods provides a framework that reduces subjectivity in determining criterion importance and produces more discriminative ranking results. The findings indicate that candidate GF achieved the highest score of 0.6848, followed by EY with 0.6835 and AR with 0.6528, showing that these candidates have performance profiles closest to the defined criteria. In addition, sensitivity analysis using several scenarios of criteria weight changes demonstrates that the proposed model is relatively stable, with an overall ranking consistency of 81.8%, while alternatives AR, DI, and FR show 100% ranking stability. These results indicate that the Entropy–CRADIS approach can improve the accuracy, objectivity, and reliability of employee selection decisions in multi-criteria decision-making environments.

Deep Learning-Based Early Detection Optimization for Rice Leaf Diseases to Support Sustainable Local Agriculture

2026-04-03T13:59:10+07:00

Rice leaf diseases such as Bacterial Blight and Blast are major threats to rice productivity that directly impact food security and the sustainability of local agriculture. This study aims to develop and optimize a deep learning-based early detection system for rice leaf diseases using a Convolutional Neural Network (CNN) architecture, specifically the Inception_v3 model. The research method includes five main stages, namely collecting rice leaf image datasets, data pre-processing (resize, normalization, and augmentation), CNN model design, model training and evaluation, and performance optimization through the application of different optimizer algorithms. Two model variants were tested and compared, namely Inception_v3 Basics with the RMSprop optimizer and Inception_v3 Optimization with the Adam optimizer. Experimental results showed that the Inception_v3 Optimization model provided the best performance, with a Precision value of 0.9672, Recall of 0.8939, F1-score of 0.9291, Balanced Accuracy of 0.9297, Matthews Correlation Coefficient (MCC) of 0.8578, Cohen's Kappa of 0.8573, and AUC ROC of 0.98. These results indicate that the Adam optimizer is able to accelerate convergence and improve model accuracy compared to RMSprop, while producing a more stable and efficient classification system. Thus, this study successfully demonstrated that the optimized Inception_v3 architecture can be used effectively for early detection of rice leaf diseases and has high potential for integration into smart farming systems to support sustainable, technology-based local agricultural practices.

Evaluasi KNN dan Logistic Regression untuk Klasifikasi Diabetes dengan Preprocessing Terstandarisasi: Trade-off Kinerja dan Interpretabilitas

2026-04-06T14:19:11+07:00

Although K-Nearest Neighbors (KNN) and Logistic Regression have been widely used in diabetes classification, studies that systematically combine a standardized preprocessing pipeline—including median imputation, feature standardization, and stratified data splitting—and evaluate the trade-off between predictive performance and model interpretability remain limited. This study aims to compare the performance of both algorithms in classifying diabetes status using the Pima Indians Diabetes dataset, which consists of 768 samples with eight numerical attributes. The research stages include data exploration, handling missing values using median imputation, feature standardization using StandardScaler, and stratified data splitting with a ratio of 80:20. Model evaluation is conducted using accuracy, precision, recall, F1-score, confusion matrix, and ROC-AUC metrics. The experimental results show that KNN with an optimal parameter of K=21 achieves an accuracy of 75.97%, an F1-score of 61.86%, and a ROC-AUC of 0.8120, while Logistic Regression achieves an accuracy of 70.78%, an F1-score of 54.55%, and a ROC-AUC of 0.8130. Although KNN demonstrates higher predictive performance, Logistic Regression provides advantages in interpretability through model coefficients, where the variables Glucose (β=1.1825) and BMI (β=0.6887) are identified as the main predictors of diabetes risk. These findings indicate a clear trade-off between accuracy and interpretability, suggesting that KNN is more suitable for high-accuracy prediction tasks, while Logistic Regression is more appropriate in clinical contexts requiring transparency and model accountability.

Perbandingan Kinerja XGBoost dan Naive Bayes dalam Analisis Sentimen Komentar TikTok Terhadap Ibu Kota Nusantara (IKN) pada Data Tidak Seimbang

2026-04-08T00:35:50+07:00

The growth of social media has generated diverse public responses regarding the development of Indonesia’s new capital city, Ibu Kota Nusantara (IKN), particularly on TikTok, a platform with high user interaction. This study aims to compare the performance of Naive Bayes and eXtreme Gradient Boosting (XGBoost) algorithms in sentiment analysis of TikTok comments related to IKN development under imbalanced data conditions. The dataset consists of 1,132 comments that underwent preprocessing, including case folding, text cleaning, tokenization, normalization, and stemming. Feature extraction was performed using the Term Frequency–Inverse Document Frequency (TF-IDF) method, generating 1,926 features to represent word importance. The classification process used an 80:20 split for training and testing data. The results show that Naive Bayes achieved an accuracy of 61.23%, while XGBoost obtained a slightly higher accuracy of 62.11%. XGBoost improved recall in the negative class (from 0.21 to 0.40) and neutral class (from 0.11 to 0.26), although the improvement remains limited. The difference in accuracy between the models is relatively small and does not indicate a significant overall performance improvement. This study is limited by the relatively small dataset size and imbalanced class distribution, which may affect data representativeness and model generalization. Therefore, the results are not yet optimal for broader real-world applications.

Hybrid Feature Selection with Metaheuristics for Improving the Accuracy of Diabetes Disease Prediction

2026-04-08T22:20:49+07:00

Early diagnosis of diabetes mellitus is crucial to prevent severe complications and reduce long-term healthcare costs, making accurate and efficient predictive models an important research focus in medical data analytics. However, one of the main challenges in diabetes prediction lies in the presence of irrelevant and redundant features within medical datasets, which can degrade classification accuracy, increase computational complexity, and reduce model generalizability. To address this issue, this study proposes a Hybrid Feature Selection (HFS) approach that integrates filter-based methods and meta-heuristic optimization to identify an optimal subset of features for diabetes prediction. In the proposed framework, statistical filter techniques combining Chi-square and Mutual Information are first employed to rank and reduce feature dimensionality by selecting the most relevant attributes. Subsequently, a Genetic Algorithm (GA) is applied to further optimize the feature subset by maximizing classification accuracy while minimizing the number of selected features. The effectiveness of the proposed HFS approach is evaluated using the Pima Indian Diabetes Dataset, consisting of 768 instances and 8 clinical features, and tested across multiple machine learning classifiers, including Random Forest, Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and XGBoost. Experimental results demonstrate that the proposed HFS significantly improves predictive performance compared to baseline models without feature selection. Specifically, the Random Forest classifier achieved the highest accuracy of 79.22%, compared to 74.03% in the baseline model, representing an improvement of approximately 5.2%. Additionally, notable improvements were observed in F1-score and AUC, with AUC increasing from 0.8336 to 0.8403. Beyond accuracy gains, the proposed method reduced feature dimensionality from 8 to 5 features, resulting in lower computational cost and faster model training time. These findings indicate that the hybrid integration of filter-based selection and meta-heuristic optimization provides a robust and efficient solution for feature selection in medical prediction tasks. Overall, the proposed HFS framework offers a promising approach for developing accurate, efficient, and reliable decision-support systems for early diabetes diagnosis.

Predictive Modeling of National University Rankings Using Ensemble Machine Learning and Multi-Dimensional Institutional Performance Indicators: Evidence from Japan

2026-04-09T14:23:45+07:00

The global higher education landscape is becoming increasingly competitive in attracting outstanding students, qualified faculty, and international research collaborations. University ranking systems serve as strategic instruments for assessing institutional performance and as a basis for public policy. However, traditional ranking approaches employing linear aggregate scores often oversimplify the complex relationships among indicators such as research, internationalization, and graduate outcomes. This study develops a data-driven predictive model to map the non-linear relationships among university performance indicators. The research employs a quantitative predictive analytics approach using a dataset of 52 Japanese universities from the 2024–2026 period, encompassing the variables Research_Impact_Score, Employment_Rate, Intl_Student_Ratio, Institution_Age, Institution_Type, and Region, with National_Rank as the target variable. The research stages include data preprocessing (handling missing values, encoding, scaling), feature engineering (including Institutional Age), regression model development (Linear, Ridge, Lasso, SVR) as well as ensemble models (Random Forest and Gradient Boosting), evaluation using RMSE, MAE, and R², and explainable analysis based on feature importance. The results indicate that the Gradient Boosting model delivers the best performance with an RMSE of 1.175117, MAE of 1.087856, and R² of 0.994988, followed by Random Forest with an RMSE of 1.436536 and R² of 0.992510. Traditional linear regression models demonstrate significantly lower performance (R² 0.657519), confirming the superiority of non-linear approaches in modeling complex relationships among indicators. Stability testing using K-Fold Cross Validation yields an average RMSE of 1.1045 with a difference of 0.4493 between folds, indicating model consistency. Feature contribution analysis reveals that Research_Impact_Score is the dominant factor with a contribution of 97.94%, followed by Employment_Rate at 1.81%, while internationalization indicators and geographical factors contribute minimally. These findings confirm that research performance constitutes the primary determinant of university rankings, whereas employability and internationalization serve as supporting factors. This study demonstrates that ensemble-based machine learning models are effective in predicting national rankings accurately and interpretably. This approach offers a multidimensional evaluation framework that is more representative than linear aggregate scores, and provides policy implications for enhancing research quality, curriculum relevance, and internationalization strategies of higher education institutions.

Sentiment Analysis on the Allocation of the MBG Program Budget Using Support Vector Machine

2026-04-22T00:50:09+07:00

Sentiment analysis is one of the applications of artificial intelligence and machine learning used to automatically identify and classify public opinions, particularly those expressed on social media. This approach is important for understanding public perceptions of a policy, as it provides a systematic, fast, and data-driven overview. With the increasing use of social media, sentiment analysis can be utilized as an evaluation tool to support transparency and more objective decision-making. One issue that has attracted public attention is the MBG (Free Nutritious Food) Program, a government initiative aimed at improving community nutrition. The budget allocation for this program has generated various responses from the public, including both support and criticism regarding its implementation and policy priorities. Therefore, an analysis that can comprehensively capture these diverse perspectives is necessary. This study aims to analyze public sentiment toward the MBG Program budget using data from the social media platform X (Twitter), which is known for its ability to represent real-time and dynamic public opinion. The dataset collected through crawling consists of 2,487 entries, and after preprocessing, 1,686 valid data points were obtained for analysis. Feature extraction was performed using the TF-IDF method, while sentiment classification was conducted using the Support Vector Machine (SVM) algorithm. Model evaluation was carried out using 5-Fold Cross Validation and Confusion Matrix. The results show that the developed model achieved an accuracy of 81.17%, indicating good performance in sentiment classification. For the negative class, the precision reached 85.48% and recall 98.76%. For the neutral class, the precision was 57.58%, recall 44.19%, and F1-score 49.98%. For the positive class, the precision was 75.00%, recall 15.79%, and F1-score 26.09%. These findings indicate that a machine learning-based approach can contribute to understanding public opinion and support more effective, data-driven government policy evaluation. This study contributes by demonstrating the effectiveness of the SVM algorithm in classifying public sentiment on policy-related issues, as well as by applying k-fold cross-validation and confusion matrix to provide a more comprehensive and reliable evaluation. The findings are expected to support data-driven policy evaluation and enhance understanding of public opinion toward government programs.

Klasifikasi Opini Pengguna TikTok terhadap Keamanan dan Efektivitas Produk Skincare Lokal menggunakan Metode Naïve Bayes, Decision Tree, dan Random Fores

2026-04-22T00:49:37+07:00

This study aims to analyze and compare the performance of Naïve Bayes, Decision Tree, and Random Forest algorithms in classifying TikTok users’ opinions regarding the safety and effectiveness of local skincare products. The results show that these algorithms exhibit significant differences in performance for sentiment classification tasks. Before applying SMOTE, Random Forest achieved the highest accuracy of 87%, followed by Decision Tree at 79% and Naïve Bayes at 65%. The main weakness was observed in minority classes such as Safe and Unsafe, which had low recall values. After applying SMOTE, all models showed improved performance, particularly in recognizing minority classes, resulting in more balanced accuracy, precision, recall, and F1-score across all sentiment categories. The TF-IDF analysis revealed that the extracted features were still dominated by common words and brand names, indicating that they did not fully represent the specific aspects of safety and effectiveness. This suggests that the preprocessing and feature selection stages can be further improved to generate more relevant feature representations. The classification visualization showed that most comments were categorized as Effective and Ineffective, while the Neutral category contained fewer instances. The implementation of SMOTE improved model performance in handling imbalanced data; however, it must be applied carefully only to the training data to avoid evaluation bias. Overall, Random Forest demonstrated the best performance among the evaluated algorithms. This study contributes to the development of a multi-class sentiment analysis model capable of distinguishing between safety and effectiveness aspects of skincare products, and demonstrates that the application of SMOTE effectively improves classification performance on imbalanced datasets. Future research is recommended to enhance sentiment labeling methods, improve feature quality, and explore more advanced approaches such as deep learning to achieve more accurate and robust classification results.

Analisis Segmentasi Pelanggan Menggunakan RFM dan K-Means Clustering sebagai Dasar Penyusunan Aturan Pendukung Keputusan

2026-04-26T22:23:07+07:00

One of the important methods in supporting data-driven Customer Relationship Management (CRM) initiatives is customer segmentation. However, in practice, segmentation results are often limited to descriptive analysis and are not further utilized in decision-support processes. This study aims to utilize customer segmentation results based on the Recency, Frequency, Monetary (RFM) approach and the K-Means algorithm as a basis for developing decision-support recommendations. The research stages include data preprocessing, RFM value calculation, normalization using the Min-Max Scaling method, and determining the optimal number of clusters using the Elbow Method and Silhouette Score. The evaluation results indicate that the optimal number of clusters is four, with a Silhouette Score of 0.61, which reflects a moderately good level of cluster separation. The segmentation results classify customers into four categories: High Value/VIP Customers, Loyal Customers, Potential Customers, and Low Value/Dormant Customers, each exhibiting distinct transactional behavior characteristics. These characteristics are then interpreted into decision rules using IF–THEN logic; for example, customers with low Recency, high Frequency, and high Monetary values are recommended strategies such as loyalty rewards and upselling. The findings suggest that customer segmentation can be extended beyond descriptive analysis and utilized as a practical basis for marketing decision-making, although the approach remains relatively simple and heuristic-based. The contribution of this study is to integrate RFM-KMeans segmentation results with IF–THEN decision rules to generate more applicable marketing strategy recommendations in supporting data-driven decision making.

Analysis of Air Pollution Standard Index Using Support Vector Machine Algorithm

2026-04-26T22:41:49+07:00

Air pollution is one of the major environmental problems in urban areas, including Medan City, Indonesia. The Air Pollution Standard Index (Indeks Standar Pencemar Udara / ISPU) data provided by the Environmental Agency is often difficult for the public to interpret due to its numerical format. This study aims to analyze and classify air quality using the Support Vector Machine (SVM) algorithm and present the results through data visualization. The dataset used in this research is secondary data obtained from the Environmental Agency of Medan City, including pollutant parameters such as PM10, PM2.5, SO₂, NO₂, CO, O₃, and HC. The research method follows a quantitative descriptive approach, including data preprocessing, ISPU calculation based on government regulations, classification using SVM, and visualization using graphical methods such as line charts, bar charts, and heatmaps. The results indicate that SVM is effective in classifying air quality categories into Good, Moderate, Unhealthy, Very Unhealthy, and Hazardous. Additionally, visualization techniques improve the interpretability of air quality data, making it easier for stakeholders and the public to understand environmental conditions. This study contributes to decision support systems for environmental monitoring and public awareness.

Penerapan Regresi Logistik, K-NN, dan Naïve Bayes Berbasis Pendekatan CRISP-DM dalam Memprediksi Penyakit Jantung

2026-05-05T17:04:20+07:00

Heart disease remains the leading cause of mortality globally, despite having significant potential to be controlled through early detection and effective risk-factor management. To improve the accuracy and efficiency of early detection, machine learning technology is employed to develop predictive models for heart disease risk. The research aims to compare the performance of three classification algorithms in predicting heart disease risk to identify the most optimal model. This research applies the CRISP-DM methodology to build and compare predictive models for heart disease risk using three supervised learning algorithms: K-Nearest Neighbors (K-NN), Naïve Bayes, and Logistic Regression. The dataset used is a heart disease dataset obtained from the Kaggle platform, consisting of 10,000 records with variables such as Age, Blood Pressure, Smoking, Diabetes, Cholesterol, Triglyceride Level, Fasting Blood Sugar, and CRP Level. For the K-NN model, experiments were conducted using three values of k (k = 5, k = 10, and k = 20) to examine the effect of the number of neighbors on model performance. Meanwhile, the Naïve Bayes and Logistic Regression models were implemented using default parameters without additional tuning to ensure a consistent performance comparison. Model performance was evaluated using Accuracy and F1-Score metrics. The evaluation results indicate that the K-NN model with k = 5 achieved the best performance, with an accuracy of 0.7203 and an F1-Score of 0.7598, outperforming the Naïve Bayes and Logistic Regression models.