Building of Informatics, Technology and Science (BITS)

Deteksi Manipulasi Citra Medis MRI Menggunakan Watermarking Least Significant Bit dengan Autentikasi SHA-256 dan ECDSA

2026-06-05T23:56:39+07:00

Medical image security is a crucial aspect of maintaining the integrity and authenticity of diagnostic data, particularly during digital transmission and storage processes that are vulnerable to manipulation. Minor modifications to pixels can lead to misdiagnosis; thus, protection methods are required to verify integrity without compromising visual quality. However, previous studies still face a trade-off between system complexity, computational efficiency, and tamper detection capabilities. This research aims to develop a medical image watermarking method capable of efficiently detecting changes in diagnostic areas with minimal distortion. The proposed method integrates automated Region of Interest (ROI) segmentation based on Otsu thresholding, 1-LSB watermark embedding in the Region of Non-Interest (RONI), and authentication based on SHA-256 and ECDSA digital signatures. The primary contribution of this study is an integrated framework that combines automated segmentation and cryptographic authentication to maintain image integrity without sacrificing clinical information. Experimental results demonstrate that the method maintains high image quality, with an average PSNR of 75.04 dB, low MSE, and the highest SSIM of 0.9999975. This performance is achieved through a small payload (99 bytes) that modifies only 1.21% of pixels in the RONI. In terms of efficiency, the method exhibits relatively fast computational performance with average embedding and extraction times of 0.14 seconds and 0.095 seconds, respectively, on 256×256 pixel images using an AMD Ryzen 5 5600H and 16 GB RAM. The system is capable of detecting ROI manipulation, identifying global payload damage, and remains valid under RONI changes, although it remains limited against large-scale manipulation due to the fragile nature of the LSB technique.

Deteksi Penyakit Jantung Menggunakan SVM dan XGBoost dengan Interpretabilitas SHAP dan Integrasi LLM

2026-06-05T23:56:39+07:00

Cardiovascular disease remains the leading cause of death globally, demanding accurate early detection, yet limited access to specialist medical personnel in developing countries often hinders timely diagnosis. This study aims to address the critical gap between the high accuracy of machine learning models in academic research and the minimal adoption of practical clinical applications by developing a safe and trustworthy hybrid artificial intelligence-based heart disease triage system. The proposed methodology integrates a dual-model architecture in which Support Vector Machine serves as the primary prediction model and Extreme Gradient Boosting as a second-opinion model, both optimized with SMOTE oversampling technique to handle class imbalance, and implements SHAP to provide transparency in black-box model decisions. The system is enriched with Dynamic Prompt Engineering innovation on the Mistral-7B Large Language Model to translate numerical probabilities into safe, personalized, and empathetic medical narratives. Experimental results show that the Support Vector Machine model with RBF kernel delivers superior performance with an accuracy of 90.22% and sensitivity of 94.12%, which is crucial for minimizing false negative cases in medical screening, outperforming the Extreme Gradient Boosting model which recorded 88.04% accuracy. Interpretability analysis identified chest pain type, cholesterol level, and maximum heart rate as the primary risk indicators, validating the model's alignment with standard cardiology guidelines. A dual safety validation mechanism through programmed risk thresholds and language generation temperature control ensures the system does not produce harmful diagnostic hallucinations. In conclusion, the system implemented as a FastAPI-based microservice is proven technically feasible with low latency, offering an accurate, transparent, and communicative early screening solution to support healthcare service efficiency.

Perbandingan Naïve Bayes dan Support Vector Machine Dalam Analisis Sentimen Google Maps Pusat Perbelanjaan

2026-06-05T23:56:40+07:00

The rapid growth of user reviews on Google Maps is not always accompanied by ease in understanding the sentiment contained within them, causing tourists and the general public to face difficulties in determining shopping centers with good reputation and service quality. The lack of information regarding visitor satisfaction levels, along with various facility-related issues such as crowd density, limited parking space, and the comfort of public facilities, combined with the large number of subjective and unstructured reviews, makes manual sentiment analysis ineffective and potentially leads to less accurate conclusions. This investigation aims to analyze sentiment from Google Maps reviews of shopping centers in the city of Semarang utilizing the Support Vector Machine (SVM) and Naïve Bayes methods. The data were collected from five shopping centers with the highest number of reviews in Semarang, namely Paragon Mall, Mall Ciputra, Java Mall, DP Mall, and Queen City Mall. The investigation method includes text preprocessing, TF-IDF weighting, and sentiment classification into three classes: negative, neutral, and positive. The dataset was divided into training and testing data with a ratio of 80:20. The outcomes reveal that the Naïve Bayes method achieved an accuracy of 85.56%, while the Support Vector Machine (SVM) method achieved an accuracy of 89.20%. Considering the outcomes, the SVM method performs better in classifying sentiment from Google Maps reviews of shopping centers in Semarang.

Analisis Sentimen X Terhadap Isu Industri Sawit Prabowo Subianto Menggunakan TF-IDF dan Machine Learning

2026-06-05T23:56:40+07:00

This study aims to analyze public sentiment on the X platform regarding the palm oil industry issue associated with Prabowo Subianto and to compare the performance of Decision Tree, Support Vector Machine (SVM), and Random Forest algorithms. The dataset consisted of 3,785 tweets collected through a crawling process. The data were then processed through cleaning, case folding, text normalization, tokenizing, stopword removal, and stemming. Sentiment labeling was conducted using a lexicon-based approach, followed by feature extraction using Term Frequency-Inverse Document Frequency (TF-IDF) and traintest data splitting. The labeling results show that public opinion was dominated by positive sentiment with 3,018 tweets (79.7%), while negative sentiment accounted for 767 tweets (20.3%). The experimental results indicate that SVM achieved the best performance with an accuracy of 0.90, followed by Random Forest with 0.86 and Decision Tree with 0.84. SVM also demonstrated more stable performance based on precision, recall, and F1-score across both sentiment classes. These findings indicate that SVM is the most effective model for Indonesian-language sentiment classification on palm oil policy issues and has strong potential to support public policy evaluation based on social media data.

Optimasi Bayesian pada Gradient Boosting untuk Prediksi Niat Beli E-Commerce pada Dataset dengan Ketidakseimbangan Kelas

2026-06-05T23:56:40+07:00

Predicting consumer purchase intention in e-commerce is a crucial challenge due to the high rate of class imbalance, where the majority of visitors only browse without making a transaction. This study compares the performance of three Gradient Boosting family algorithms (XGBoost, LightGBM, and CatBoost) using the Online Shoppers Intention dataset, which has a class ratio of 84.5% to 15.5%. To overcome majority class bias, the Synthetic Minority Oversampling Technique (SMOTE) approach was implemented on the training data. This research focuses on hyperparameter optimization implementation using the Optuna framework based on the Tree-structured Parzen Estimator (TPE), which is statistically validated using the Friedman and Post-Hoc Nemenyi tests. Model evaluation using stratified 10-Fold Cross-Validation shows that all three models can handle class imbalance effectively. LightGBM achieved an accuracy of 88.36% with an ROC-AUC of 0.9138, XGBoost achieved an accuracy of 88.56% with an ROC-AUC of 0.9127, and CatBoost achieved an accuracy of 88.56% with an ROC-AUC of 0.9121. Feature importance analysis identifies ProductRelated_Duration and ExitRates as the main predictors of purchase intention. The Friedman statistical test detected global performance differences (p=0.0450), but the Nemenyi post-hoc test found insufficient empirical evidence to claim significant pairwise performance differences. This research provides a practical contribution to the e-commerce industry by demonstrating that the selection of ensemble algorithms no longer needs to rely absolutely on pseudo-accuracy margins, but can be objectively recommended based on computational latency efficiency, where the LightGBM architecture proves to be efficient.

Evaluasi Validitas Model Machine learning pada Klasifikasi Stunting Berbasis Data Antropometri dan Hubungan Deterministik

2026-06-05T23:56:40+07:00

Stunting is a chronic nutritional problem among infants and toddlers that affects children’s growth and development. Various studies have utilized machine learning for nutritional status classification based on anthropometric data; however, the validity of the resulting models has rarely been examined. This study aims to evaluate the validity of machine learning models in classifying stunting status using the XGBoost, Random Forest, and Naïve Bayes algorithms. The dataset consists of 120,999 anthropometric records of infants, with age, gender, and height as features, and nutritional status as the target variable. The research process included preprocessing, data transformation, and model evaluation using the k-fold cross-validation method with accuracy, precision, recall, and F1-score metrics. The results showed that Random Forest and XGBoost achieved very high accuracy, at 99.91% and 99.08%, respectively, while Naïve Bayes reached only 55%. This stark difference in performance indicates that ensemble-based models are capable of capturing very strong patterns in the data, while Naïve Bayes struggles due to the interdependence among features. Furthermore, the high accuracy of certain models suggests a deterministic relationship between features and labels, which could potentially make the models less robust against data containing measurement errors or noise.

Analisis Perbandingan Metode Edas Dan Aras Dalam Pemilihan Platform Freelance Terbaik Untuk Pekerja Jarak Jauh (Remote Worker)

2026-06-05T23:56:41+07:00

The trend of remote workers has increased significantly, driving the high adoption of global freelance platforms. However, the diversity of policies regarding service fees, withdrawal limits, and levels of competition across platforms often makes it difficult for beginner remote workers to determine the most optimal choice. This study aims to analyze and compare the recommendation results of a Decision Support System (DSS) using the Evaluation based on Distance from Average Solution (EDAS) method and the Additive Ratio Assessment (ARAS) method in selecting freelance platforms. The study evaluates five platform alternatives (Upwork, Fiverr, Fastwork, Freelancer, and Projects.co.id) using a mixed-methods approach that combines factual platform policy data (Administrative Fee Deduction and Minimum Withdrawal) with user perception data (UI/UX, Security, and Level of Competition). The analysis results show a high level of consistency between the two methods for the best alternative, where Upwork (A1) ranks first with an Appraisal Score (AS) of 0.965 in EDAS and a Utility Degree (Ki) of 0.958 in ARAS. However, the comparative analysis reveals differences in rankings at the 4th and 5th positions, caused by the extreme value (outlier) sensitivity of the EDAS algorithm on cost attributes and the more tolerant stability of the ARAS algorithm in providing proportional value compensation. This study concludes that a comparative method not only provides validated recommendations but also reveals the characteristics of each algorithm in handling anomalies in cost attribute data. The main contribution of this study is to provide a valid comparative decision-making framework for remote workers in optimizing platform selection, while also enriching the academic literature regarding the disclosure of algorithmic sensitivity in the ARAS and EDAS methods when handling cost data anomalies.

Segmentation-Aware Recommendation with Cluster-Specific Item Graphs Using Pointwise Mutual Information for Market Basket Analysis

2026-06-05T23:56:41+07:00

Traditional Association Rule-based recommendation methods often exhibit limited coverage and high redundancy when applied to sparse transactional data, thereby constraining their effectiveness for product discovery in e-commerce systems. This study proposes a hybrid recommendation framework that integrates customer behavioral segmentation with graph-based item representation learning to address these limitations. Customers are first grouped into behaviorally homogeneous clusters using historical transaction features. For each cluster, an item co-occurrence graph is constructed and weighted using pointwise mutual information to mitigate sparsity bias and emphasize informative associations. Graph-based representation learning is then applied using Node2Vec to generate low-dimensional product embeddings that capture both local structural proximity and higher-order relational patterns. The proposed framework explicitly restricts the candidate item space to the Top 100 most frequent products within each behavioral cluster, thereby focusing the recommendation task on improving localized discovery within high-frequency product segments rather than global catalog exploration. The objective of this research is to assess whether segmentation-aware graph embeddings can outperform traditional FP-Growth association rules under a strict temporal split between the Historical Training Set and the Hold-out Evaluation Set, ensuring realistic and leakage-free evaluation. Model performance is evaluated using precision, recall, normalized discounted cumulative gain, and intra-list diversity on the Hold-out Evaluation Set. Experimental results indicate that the proposed graph-based approach improves ranking quality and diversity within constrained high-frequency item spaces, demonstrating more effective localized discovery within Top 100 product segments compared to FP-Growth. These results demonstrate that graph-based embeddings are more robust to sparse behavioral patterns within high-frequency product segments and better suited for exploratory recommendation scenarios within dense product subsets. The proposed framework offers a scalable and temporally valid foundation for knowledge-driven recommender systems.

Comparing TF-IDF Based SVM and Logistic Regression for Imbalanced Pertamina Corruption Tweet Sentiment Classification

2026-06-05T23:56:41+07:00

The corruption case involving PT Pertamina (Persero) in early 2025 generated widespread public reactions on social media, particularly on the X (Twitter) platform. The rapid dissemination of opinions in digital environments highlights the importance of analyzing public sentiment toward socio-political issues. This study aims to examine public sentiment regarding the Pertamina corruption case using a text classification approach based on Term Frequency–Inverse Document Frequency (TF-IDF). This study contributes a controlled comparison of TF-IDF-based Support Vector Machine (SVM) and Logistic Regression on imbalanced Indonesian-language tweets related to a nationally salient corruption issue, while also emphasizing the importance of evaluating performance beyond accuracy alone through macro-F1 and minority-class recall. Two classification algorithms, Support Vector Machine (SVM) and Logistic Regression, were employed to compare their performance in predicting lexicon-derived positive and negative sentiment labels.. To address class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was applied to the training data. A total of 3,058 Indonesian-language tweets collected between February 25 and March 10, 2025 underwent preprocessing and sentiment labeling using the INSET Lexicon. The results show that SVM achieved higher overall accuracy of 94.93% and a macro-F1 score of 0.80, while Logistic Regression achieved an accuracy of 90.52% and a macro-F1 score of 0.73. However, class-wise evaluation indicates that accuracy should not be interpreted independently because the dataset was dominated by negative sentiment. For the positive minority class, SVM obtained an F1-score of 0.64 and recall of 0.60, whereas Logistic Regression obtained a lower F1-score of 0.52 but a higher recall of 0.69. These findings indicate a trade-off between overall classification performance and minority-class sensitivity.

Perbandingan XGBoost dan Random Forest Menggunakan Seleksi Fitur ANOVA-MI Dalam Klasifikasi Kesehatan Janin Cardiotocography

2026-06-05T23:56:41+07:00

This study compares the performance of Random Forest and XGBoost algorithms in classifying fetal health problems using Cardiotocography (CTG) data. The imbalance in the amount of data between classes, the presence of less relevant features, and the challenge in identifying the Suspect class, which has characteristics between the Normal and Pathological classes, are the main problems in the CTG dataset. This condition is important because the early stage of fetal health risk determines further medical treatment represented by the Suspect class. This study uses ANOVA and Mutual Information feature selection techniques, as well as the ADASYN oversampling method to balance the data to overcome these problems. In addition, Random Search is used to optimize model parameters to improve its performance. Unlike previous studies that generally focus on improving accuracy, this study also emphasizes the model's ability to detect minority classes, especially the Suspect class. Based on the results of the study, in almost every test scenario, XGBoost consistently outperforms Random Forest. The XGBoost model obtained optimal accuracy from the combination of ANOVA, ADASYN, and hyperparameter tuning with an accuracy of 95.51%. Meanwhile, the application of Mutual Information with ADASYN and tuning was quite effective in identifying the Suspect class with a higher recall value of 81%. However, because the Suspect class attribute lies between the Normal and Pathological class attributes, the model still faces challenges in optimally distinguishing them. Overall, this study shows that a combination of appropriate feature selection, handling data imbalance, and parameter optimization in a single pipeline can improve model performance more balanced. This research is expected to support more objective medical decision-making, especially in detecting fetal risk conditions from an early stage.

Evaluasi Kinerja Naïve Bayes, Decision Tree, Dan Random Forest Serta Voting Ensemble Pada Klasifikasi Multi-Kelas Penyakit Sapi Berbasis Gejala

2026-06-05T23:56:42+07:00

Cattle are an important livestock commodity; however, farmers often face difficulties in disease diagnosis due to the similarity of clinical symptoms and limited access to veterinary experts. This study aims to compare the performance of three machine learning classification algorithms, namely Naïve Bayes, Decision Tree, and Random Forest, and to evaluate the effectiveness of an ensemble approach using a Voting Ensemble method for cattle disease diagnosis. The study adopts the CRISP-DM methodology, consisting of data preprocessing, modeling, and evaluation stages. Model performance is assessed using accuracy, precision, recall, and F1-score metrics. The experimental results show that Naïve Bayes achieves the best performance with an accuracy of 0.951 and an F1-score of 0.920. Random Forest obtains an accuracy of 0.799, while Decision Tree performs the lowest with an accuracy of 0.265. Ensemble methods, including Voting NB+RF, Voting Weighted, Voting Soft, and Voting Hard, achieve accuracies of 0.912, 0.900, 0.853, and 0.792, respectively. These findings indicate that Naïve Bayes is more suitable for high-dimensional and sparse symptom-based data, providing the most stable performance among the evaluated models. The developed system is implemented as a web-based expert system. Usability evaluation using the System Usability Scale (SUS) yields a score of 77, categorized as “Good.” This study demonstrates that machine learning can support decision-making in cattle disease diagnosis.

Optuna-Driven Hyperparameter Optimization in Tsukamoto Fuzzy Logic for House Price Estimation

2026-06-05T23:56:42+07:00

The property sector faces challenges in determining accurate house selling prices due to subjectivity and market uncertainty. The relationship between physical attributes, such as land area and building area, and price is not always linear, making conventional methods often less precise in estimation. This study aims to design a decision support system to objectively estimate house prices in the Plamongan area, Semarang. The method used is Fuzzy Tsukamoto Logic. This preliminary study explores the integration of the Tree-structured Parzen Estimator (TPE) algorithm through the Optuna framework to automatically optimize membership function limits, replacing manual trial and error methods. The dataset was collected via scraping techniques, providing a pilot dataset of 26 data points. Final model performance evaluation showed a Mean Absolute Percentage Error (MAPE) value of 11.39%, which falls into the 'Good Forecast' category. However, given the highly limited sample size, these findings primarily serve as a proof-of-concept that requires further validation with larger, multi-variable datasets. These results prove that integrating the Fuzzy Tsukamoto method with hyperparameter optimization is effective in reducing subjectivity and providing reliable property price estimates. The primary contribution of this research is providing a mathematical proof-of-concept for an automated, objective property valuation system that eliminates human bias in fuzzy parameter configuration, offering a practical baseline tool for localized real estate markets.

Classification of School Students Lifestyle Risks Based on Smoking Behavior Using Naïve Bayes

2026-06-05T23:56:42+07:00

This study aims to classify students' lifestyle risks based on smoking behavior using the Naïve Bayes algorithm within a knowledge management framework. The research was conducted on students at a vocational high school within the coverage area of a local community health center. The dataset consisted of 277 valid records after undergoing data selection, cleaning, and transformation stages. The modeling process was carried out using RapidMiner software with an 80:20 data split for training (221 students) and testing (56 students). The evaluation metrics used included accuracy, precision, recall, and confusion matrix. The experimental results demonstrate that the Naïve Bayes model achieved an accuracy of 85.92%, precision of 86.12%, and recall of 92.86% for the unhealthy class. Furthermore, the classification results were integrated into a knowledge management framework to support decision-making processes in schools and community health centers. This study contributes to the application of predictive data mining in adolescent health and demonstrates how classification models can serve as effective tools for early detection, preventive interventions, and evidence-based policy formulation in educational and health settings.

Sentiment Classification on Indonesian Game Sequels: A Comparative Analysis of SVM and Naive Bayes on Coffee Talk Franchise Reviews

2026-06-05T23:56:42+07:00

User reviews on Steam are a critical source of feedback for game developers, yet manual sentiment analysis at scale is impractical. This study aims to compare Support Vector Machine (SVM), Multinomial Naive Bayes (MNB), and Complement Naive Bayes (CNB) for binary sentiment classification and to analyze sequel reception patterns through cross-game evaluation. Reviews were preprocessed with negation-aware stopword removal and WordNet lemmatization, then vectorized with TF-IDF unigram and bigram features. Four scenarios were evaluated: two within-game baselines, a cross-game generalization, and a combined evaluation. Class imbalance was handled at the model level via class weighting for SVM and the CNB variant. Macro-averaged F1-Score was the primary metric. SVM consistently outperformed both Naive Bayes variants, achieving macro-F1 of 0.81 within-game and 0.75 cross-game. MNB collapsed to majority-class prediction across all scenarios; in S2, all three models also failed on the minority class due to the small test partition (n=6). The cross-game result indicates that sentiment patterns transfer reasonably from the original game to its sequel, with the performance drop concentrated in the minority class. These findings offer practical guidance for Indonesian game developers monitoring sequel reception through automated sentiment analysis.

Komparasi Random Forest dan Artificial Neural Network dalam Prediksi Dampak AI terhadap Pekerjaan 2030

2026-06-05T23:56:42+07:00

The development of Artificial Intelligence (AI) is expected to affect the future employment structure, particularly regarding automation risks in 2030 as a period of accelerated AI adoption across various industrial sectors. This study aims to compare the performance of the Random Forest and Artificial Neural Network (ANN) algorithms in predicting the impact of AI on jobs. The study employed two modeling approaches, namely regression to predict job automation probability and classification to determine job risk categories into Low, Medium, and High classes through a discretization process. The dataset was obtained from Kaggle with a total of 3,000 records and processed through preprocessing, feature engineering, and train-test splitting with an 80:20 ratio. Regression evaluation was conducted using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and coefficient of determination (R²), while classification evaluation used accuracy and F1-score. The results showed that Random Forest achieved the best regression performance with an MAE of 0.0786, RMSE of 0.0932, and R² of 0.8640, outperforming ANN with an MAE of 0.0949, RMSE of 0.1137, and R² of 0.7973. In the classification task, both algorithms achieved an accuracy and F1-score of 99.33%. This study shows that Random Forest is more stable on tabular data and contributes to the comparative analysis of ensemble learning and neural network approaches for predicting the impact of AI on jobs.

Klasifikasi Risiko Bencana di Indonesia Menggunakan SVM dan Random Forest

2026-06-05T23:56:43+07:00

Indonesia is a country with a high level of disaster vulnerability, requiring effective methods to accurately classify disaster risk levels. This study aims to analyze and compare the performance of Support Vector Machine (SVM) and Random Forest algorithms in disaster risk classification. The dataset used consists of disaster event data from 2019–2024, including disaster type, region, number of victims, and population density. Disaster risk levels were classified into three categories, namely low, medium, and high, based on the total impact calculated from the number of victims. The proposed method includes data preprocessing, normalization, and train-test data splitting. The results show that both models achieved high performance, where Random Forest obtained an accuracy of 95.66% and SVM achieved 95.28%, with ROC-AUC values of 0.9823 and 0.9769, respectively. Random Forest demonstrated slightly better performance with an accuracy difference of 0.38% and more consistent prediction results. The high performance indicates that the models were able to recognize the main patterns within the dataset, although the results were also influenced by the characteristics of the data used. Overall, Random Forest is more suitable for disaster risk classification on data with complex characteristics.

Hybrid CNN-BiLSTM untuk Analisis Sentimen Multi-Platform terhadap Insiden Keamanan Pangan Program Makan Bergizi Gratis

2026-06-23T00:28:09+07:00

The decline in stunting prevalence in Indonesia has not been accompanied by improvements in the quality of nutritional intervention program implementation, including the Free Nutritious Meal Program (MBG), which sparked public controversy following food safety incidents in several regions. The high volume of cross-platform public opinion on social media requires an analytical approach capable of simultaneously capturing diverse linguistic styles from various sources. This study proposes a hybrid Convolutional Neural Network-Bidirectional Long Short-Term Memory (CNN-BiLSTM) classification model to analyze public sentiment regarding the incidents, with CNN extracting local feature patterns and BiLSTM modeling bidirectional word-sequence dependencies. A total of 3,416 comments were collected from five social media platforms (X, Instagram, TikTok, YouTube, and Facebook), then processed through text preprocessing and initial lexicon-based labeling into three sentiment classes: negative, neutral, and positive. To strengthen label validity, the labeling quality was validated through manual annotation by two independent annotators, yielding a Cohen’s Kappa value of κ = 0.828. The dataset was split using an 80:20 stratified scheme, with class weight applied to reduce bias caused by class imbalance without changing the number of samples in each class. The hybrid model was compared with two baseline models, CNN and BiLSTM, using macro F1-score as the primary metric, while accuracy was used as a supporting metric. The experimental results show that the hybrid CNN–BiLSTM model achieved a macro F1-score of 90.38% and an accuracy of 94.59%, outperforming both baseline models. Misclassification analysis revealed that most errors occurred in argumentative comments, negation, and contrastive sentences, reflecting the limitations of lexicon-based labeling in capturing nuanced language. Overall, this approach demonstrates the potential of cross-platform deep learning-based sentiment analysis as an initial component for monitoring public opinion on national-scale government policies. This study contributes by providing a manually validated multi-platform Indonesian dataset, developing a hybrid CNN-BiLSTM architecture with a class weight scheme effective for three-class sentiment classification on informal text, and opening opportunities for applying deep learning as a means of data-driven public opinion monitoring.

Perbandingan Kinerja Naïve Bayes, SVM, dan Random Forest dalam Klasifikasi Risiko Kehamilan

2026-06-05T23:56:43+07:00

Classifying pregnancy risk levels is a crucial aspect in supporting early detection of potential complications in pregnant women. However, most previous studies have focused on a single algorithm and relied solely on accuracy metrics, thus failing to provide a comprehensive picture of model performance in multiclass classification. Furthermore, performance comparisons between algorithms using more comprehensive evaluation approaches are still limited. This study aims to analyze and compare the performance of the Naïve Bayes, Support Vector Machine (SVM), and Random Forest algorithms in classifying pregnancy risk levels using the Maternal Health Risk Dataset from the UCI Machine Learning Repository, which consists of 1,014 data sets with six maternal health attributes. The methods used include data preprocessing, hyperparameter optimization using GridSearchCV, and model evaluation using Stratified K-Fold Cross Validation with k = 10. Model performance was measured using accuracy, precision, recall, and F1-score metrics to provide a more comprehensive evaluation. The results showed that the Random Forest algorithm had the best performance with an accuracy value of 0.8629, precision of 0.8704, recall of 0.8629, and F1-score of 0.8635, followed by SVM and Naïve Bayes. The superiority of Random Forest is due to its ability to combine several decision trees and capture non-linear relationships between features, resulting in more accurate and stable predictions. Thus, Random Forest is recommended as the most effective method in pregnancy risk classification based on maternal health data.

Perbandingan Kinerja Model ARIMA-GARCH dan LSTM Dalam Peramalan Volatilitas Bitcoin

2026-06-05T23:56:43+07:00

Bitcoin is a cryptocurrency aset with extreme volatility, necessitating precise forecasting models for investment risk mitigation. This study aims to analyze and forecast Bitcoin price volatility using an integrated Autoregressive Integrated Moving Average - Generalized Autoregressive Conditional Heteroskedasticity (ARIMA-GARCH) approach and compare its performance with a Deep Learning method, specifically Long Short-Term Memory (LSTM). The data used is the daily closing price of Bitcoin for the period 2018 to 2025. The results indicate that the ARIMA(1,1,1)-GARCH(1,1) model effectively captures the volatility clustering phenomenon, with a significant beta parameter value of 0.8691, indicating long-term volatility persistence. However, in terms of price prediction accuracy, the LSTM model significantly outperforms the conventional statistical model. Based on the testing, the ARIMA-GARCH model produced a Mean Absolute Percentage Error (MAPE) of 18.11%, which falls into the "good forecasting" category. In contrast, the LSTM model achieved a MAPE of 3.09%, categorized as "highly accurate forecasting." The significant difference in Root Mean Square Error (RMSE) values also reinforces that the LSTM architecture is more adaptive in processing non-linear data patterns and complex Bitcoin price fluctuations. This study concludes that while ARIMA-GARCH excels in risk structure analysis, the LSTM model provides more reliable price projection results for crypto market participants.

Leakage-Aware Random Forest Regression for Predicting Job Automation Risk Using Structured Labor Market Data

2026-06-05T23:56:43+07:00

This study aims to predict job automation risk in the era of artificial intelligence (AI) using a leakage-aware Random Forest Regression approach. The automation risk score, defined as a composite index derived from task exposure to AI, occupational routine intensity, and technological susceptibility indicators sourced from the AI Impact Jobs Dataset, serves as the target variable. The dataset comprises 5,000 job vacancy records from 44 countries across 9 industries spanning 2010 to 2025. A rigorous methodological framework is applied by systematically identifying and eliminating potential data leakage features, including ai_intensity_score, reskilling_required, and ai_mentioned, which were found to share mathematical or conceptual derivation paths with the target variable. The model is evaluated using R², RMSE, MAE, and MAPE with 5-fold cross-validation. The results show that the model achieves an R² score of 0.8087 on testing data, with RMSE of 0.1129 and MAE of 0.0893. Feature importance analysis reveals that salary_change_vs_prev_year_percent is the most influential predictor (55.85%), which, although indicative of dominance bias typical in synthetic datasets, aligns with economic theories linking wage dynamics to automation incentives. The findings demonstrate that leakage control significantly reduces inflated performance estimates (from R² = 0.8857 to 0.8087), and that Random Forest Regression provides a robust predictive framework for tabular socio-economic data when combined with rigorous preprocessing. This study contributes a methodological template for preventing data leakage in labor market prediction tasks.

Segmentasi Pelanggan E-Commerce Berbasis Integrasi Text Mining dan RFM untuk Deteksi Dini Churn

2026-06-05T23:56:43+07:00

The growth of transactions on e-commerce platforms generates a massive volume of unstructured customer review data. However, traditional Customer Relationship Management (CRM) models such as RFM often only focus on quantitative transaction data and ignore the emotional dimension contained in customer reviews. This study aims to analyze the relationship between purchase frequency and customer comment polarity through the integration of Text Mining and CRM Analytics approaches. The novelty offered is the development of a hybrid method that combines Lexicon Refinement-based sentiment extraction with the Random Forest algorithm to overcome rating bias in global e-commerce platform data (Kaggle). The proposed method includes the use of Natural Language Processing (NLP) techniques, topic modeling based on Latent Dirichlet Allocation (LDA), and sentiment analysis to extract polarity scores. The test results show that the initial lexicon model has limitations with an accuracy of 52.14% due to noise in neutral reviews (3-star rating). However, after optimization using the Random Forest algorithm and neutral data filtering, the classification accuracy increased significantly to 74.62%. These results prove that sentiment integration is able to provide more accurate loyalty mapping and help e-commerce management detect potential churn in the At-Risk customer segment.

Aspect-Based Sentiment Analysis on Skintific Product Reviews Using IndoBERT

2026-06-05T23:56:43+07:00

The rapid growth of the beauty industry has generated a massive volume of online reviews where traditional sentiment analysis fails to capture contradictory opinions across specific product features. This study implements Aspect-Based Sentiment Analysis (ABSA) using the IndoBERT-base-p1 architecture on 2,139 review data points from Female Daily, integrated with a specialized slang normalization stage to mitigate linguistic noise. The novelty lies in evaluating IndoBERT’s bidirectional attention robustness in processing technical medical terminology alongside Indonesian social media slang—a complexity often overlooked in prior beauty domain studies. This study contributes a novel methodological pipeline that bridges deep learning architectures with domain-specific linguistic preprocessing, providing a benchmark dataset for Indonesian beauty product reviews. The results showed that IndoBERT was able to distinguish nuances of sentiment, with superior performance in the Effectiveness (F1-Score 72.57%) and Texture (F1-Score 71.10%). Although the average score was affected by sample limitations in certain aspects, the model proved effective in capturing the semantics of Indonesian consumer slang. Ultimately, this research provides a practical contribution for consumers in validating product quality specifically and for producers as a basis for evaluating product performance in the public eye.

Arrhythmia Detection Using XGBoost with Recursive Feature Elimination: A Two-Stage Machine Learning Approach

2026-06-05T23:56:43+07:00

Arrhythmia is a cardiac rhythm disorder that can lead to severe complications, including heart failure and sudden cardiac death. Accurate electrocardiogram (ECG)-based arrhythmia detection remains challenging due to high-dimensional features and class imbalance. Therefore, this study aims to develop a two-stage machine learning approach for arrhythmia detection using Recursive Feature Elimination (RFE) and Extreme Gradient Boosting (XGBoost). The proposed approach performs binary classification to distinguish normal and arrhythmia conditions, followed by multi-class classification to identify arrhythmia subtypes. SMOTE is applied to address class imbalance, while Grid Search with cross-validation is used for hyperparameter optimization. Furthermore, the trained model is implemented in a web-based application for interactive prediction and visualization. Experimental results show that the optimized binary classification model achieves an accuracy of 0.89 and an F1-score of 0.87. Meanwhile, the multi-class classification model achieves an accuracy of 0.69 and a weighted F1-score of 0.66. The results indicate that the proposed approach performs effectively for binary arrhythmia detection. However, performance in multi-class classification remains limited due to imbalance and insufficient samples in several arrhythmia subtypes. This study contributes by proposing an integrated framework that combines Recursive Feature Elimination (RFE) for feature selection, SMOTE for imbalance handling, XGBoost with GridSearchCV-based hyperparameter optimization, and a two-stage classification approach for ECG-based arrhythmia detection and subtype classification. In addition, the proposed model is implemented in a web-based application to support interactive prediction and visualization. Overall, this study demonstrates the potential of integrating RFE, XGBoost, and SMOTE for ECG-based arrhythmia detection and practical web-based implementation.

Komparasi Naïve Bayes, SVM, dan Decision Tree untuk Klasifikasi Komentar Provokatif pada Instagram Terkait Aksi Demonstrasi Agustus 2025

2026-08-05T17:42:26+07:00

The accelerating growth of social media has transformed digital platforms into spaces where people express opinions on social and political issues. One event that generated numerous public comments was the demonstration held on August 25–31, 2025. Many comments contained harsh language, provocation, insults, and calls for conflict that had the potential to trigger negative emotions in digital spaces. This study aims to classify provocative and non-provocative comments on Instagram using the Naïve Bayes algorithm and compare its performance with Support Vector Machine (SVM) and Decision Tree algorithms. The data were collected through a web scraping process, resulting in 3,396 comments. After the cleansing and preprocessing stages, the dataset was reduced to 2,490 comments. The preprocessing stages included transform case, tokenizing, stopwords removal, filter tokens, and stemming. Furthermore, word weighting was carried out using the TF-IDF method and implemented in RapidMiner with an 80:20 data split ratio. Based on manual labeling, 1,279 provocative comments and 1,211 non-provocative comments were obtained. The evaluation results showed that Naïve Bayes achieved an accuracy of 72.15%, SVM achieved 69.44%, and Decision Tree achieved 72.91%. Although Decision Tree produced a slightly higher accuracy, Naïve Bayes demonstrated a more balanced performance in detecting both comment classes, even though the accuracy value was still in the moderate category. The findings provide insights into the effectiveness of machine learning algorithms for identifying provocative comments and may support the development of automated content moderation on social media platforms.

Comparing TabNet and CatBoost Models for Explainable Student Depression Prediction

2026-06-22T23:20:07+07:00

Early identification of depression risk among students is increasingly important for educational institutions because mental health problems may affect academic engagement, well-being, and learning continuity. This study aims to compare TabNet and CatBoost in predicting student depression risk and to examine their explainability in identifying influential predictors from structured tabular data. The dataset used in this study consists of student-related variables covering personal characteristics, academic conditions, psychological indicators, and lifestyle factors. The experimental procedure included data cleaning, missing value treatment, categorical feature transformation, feature scaling, model training, testing, and interpretation. Model performance was assessed using accuracy, precision, recall, F1-score, and ROC-AUC. Meanwhile, model explainability was examined through attention-based feature importance for TabNet and SHAP-based interpretation for CatBoost. The experimental results indicate that CatBoost produced better overall classification performance, achieving 84.54% accuracy compared with 83.44% for TabNet. CatBoost also obtained higher precision, F1-score, and ROC-AUC values. In contrast, TabNet showed slightly better recall, suggesting stronger sensitivity in detecting students classified as at risk. The interpretation results show that suicidal thoughts, financial stress, academic pressure, sleep duration, study satisfaction, dietary habits, and workload-related variables were consistently relevant to the prediction process. These findings indicate that model selection for student depression prediction should consider not only accuracy, but also sensitivity and interpretability.

Klasifikasi Pesan Penipuan pada Platform WhatsApp Menggunakan Metode Naïve Bayes Berbasis TF-IDF, N-Gram, dan Chi-Square

2026-06-22T23:40:11+07:00

The rapid development of digital communication has led to an increase in message exchanges across various platforms, accompanied by the widespread spread of fraudulent messages (scams). This situation demands an automated system capable of identifying and classifying messages quickly and accurately. This study aims to develop a text-based message classification system on the WhatsApp platform using the Naïve Bayes algorithm. The research stages include text preprocessing consisting of case folding, cleaning, normalization, stopword removal, and stemming to improve data quality. Next, feature extraction is carried out using Term Frequency-Inverse Document Frequency (TF-IDF) combined with the N-Gram (unigram) approach to represent each word in the text, and Chi-Square feature selection is applied to obtain the most relevant features in the classification process. The dataset used consists of three categories of WhatsApp messages: normal, promotional, and fraudulent. In addition, this study also applies a data balancing method using Random Oversampling to increase the number of minority class samples in the training data for optimal model performance. The main contribution of this research is the application of a combination of TF-IDF unigram, Chi-Square feature selection, and Random Oversampling in the Naïve Bayes algorithm to improve the classification performance of Indonesian WhatsApp messages, especially in conditions of unbalanced class distribution. Model evaluation is carried out using a Confusion Matrix with accuracy, precision, recall, and F1-score metrics. The test results show that the model built is able to achieve an accuracy level of 95.63%, so the method used is proven to be effective in classifying WhatsApp messages accurately and consistently.

Stereo Camera-Based Motor Vehicle Dimension Measurement with Luxmeter-Based Light Intensity Detection

2026-06-23T00:22:54+07:00

Over Dimension and Over Loading (ODOL) vehicles require accurate and efficient dimension inspection systems to support transportation safety and regulatory compliance. This study proposes an automatic motor vehicle dimension measurement system based on stereo vision integrated with YOLOv8 object detection and luxmeter-based illumination monitoring. The system was developed using a Research and Development (R&D) approach involving stereo camera calibration, hardware-software integration, experimental testing, and validation against manual measurements. Two Logitech C270 USB cameras with a fixed 50 cm baseline were calibrated using a 9 × 6 checkerboard pattern and processed using OpenCV and Python. Vehicle and wheel objects were detected using YOLOv8 models with stereo disparity estimation performed using Semi-Global Block Matching (SGBM) and triangulation methods to calculate Overall Length (OAL), Front Overhang (FOH), Wheelbase (WB), Rear Overhang (ROH), and Overall Height (OAH). Environmental lighting conditions were monitored using a luxmeter under illumination ranges of 5,000-100,000 lux. Experimental results showed a stereo calibration success rate of 96% from 100 stereo image pairs. The developed system achieved average measurement accuracies of 98.58%, 98.92%, and 98.89% at testing distances of 7 m, 8 m, and 9 m, respectively, while the highest accuracy of 99.44% was obtained at the T_8M_LC configuration under stable illumination conditions. Operational efficiency analysis showed that the automatic measurement process, including image acquisition and computational processing, reduced total measurement time from 187 seconds in manual measurements to 5 seconds in the automated system, corresponding to an efficiency improvement of 97.32%. The results show that the proposed stereo vision system provides accurate, efficient, and lighting-robust vehicle dimension measurements suitable for automated motor vehicle inspection applications.

Classification of Swiftlet Nest Quality Based on SNI 8998:2021 Using Deep Learning

2026-06-23T00:42:39+07:00

The quality of swiftlet nests is a key factor in determining the market value and quality standards of this commodity in both domestic and international markets. The quality classification process, which is currently dominated by manual methods, has fundamental weaknesses, namely high subjectivity and inconsistency in sorting results. This study aims to evaluate the performance of deep learning architectures in automatically classifying the quality of swiftlet nests based on visual characteristics. The main contribution of this study is to address the research gap in previous publications by strictly aligning quality class labels with the formal document of the Indonesian National Standard (SNI) 8998:2021, as well as presenting a cross-architecture comparative analysis to map model performance trade-offs. Evaluations were conducted on the MobileNetV2, and presents a cross-architecture comparative analysis to map model performance trade-offs. Evaluations were conducted on the MobileNetV2, ResNet50, and YOLOv8n-cls architectures using accuracy, precision, recall, and F1-score metrics. The research dataset includes visual images of swiftlet nests grouped into three quality classes (good, moderate, and poor) through self-documentation and augmentation techniques. Test results show that YOLOv8n-cls achieved the highest performance in this scenario with an accuracy of 99.5%, precision of 98.78%, recall of 98.72%, and an F1-score of 98.71%. Meanwhile, MobileNetV2 achieved a competitive accuracy of 98.37% with good computational efficiency, while ResNet50 demonstrated the lowest performance (66% accuracy) due to network complexity on the limited dataset. This research indicates that lightweight architectures exhibit good stability for limited-size visual datasets; however, external validation using larger datasets remains necessary to test the models’ generalization capabilities more broadly.

Analisis Spasio-Temporal Berbasis Data Video untuk Identifikasi Bangunan Melayu Menggunakan Metode Hybrid CNN-LSTM

2026-06-23T00:53:19+07:00

Malay buildings have distinctive architectural characteristics and require technology-based identification systems to support cultural documentation and preservation. This study aims to develop an identification system for Malay and non-Malay buildings using video data extracted into frame-by-frame images. The use of video data in this study is not intended to analyze the physical movement of buildings, but to utilize visual variations caused by changes in camera angle, recording distance, lighting, object composition, and visible building elements. The proposed method is CNN-LSTM, where CNN extracts visual features from each frame, while LSTM learns inter-frame feature relationships as a sequence of visual information. To reduce redundant information between adjacent frames and minimize the risk of excessive similarity between training and testing data, the number of frames was limited to a maximum of 15 frames per video folder, and data splitting was performed by considering video source groups. The dataset consists of Riau Malay buildings, Kalimantan Malay buildings, and non-Malay buildings. The research stages include frame extraction, image resizing to 224×224 pixels, normalization, data augmentation, class labeling, train-test splitting, modeling, evaluation, GroupKFold validation, and web-based system implementation. The best testing scenario was obtained using an 80:20 data split, 80 maximum epochs, and a batch size of 16, achieving an accuracy of 0.9916 and a test loss of 0.0852. GroupKFold validation produced an average accuracy of 99.1% with a standard deviation of 0.5%. These results indicate that the model can recognize architectural visual patterns, such as roofs, windows, doors, ornaments, and overall building appearance, while the performance should still be interpreted within the scope of the dataset and evaluation scenario used in this study.

Analisis Sentimen Twitter Terhadap Isu Royalti Lagu di Industri Musik Indonesia Menggunakan Naive Bayes dan Support Vector Machine Berbasis TF-IDF

2026-06-23T13:59:32+07:00

The development of digital platforms in Indonesia’s music industry has triggered various debates regarding the song royalty system, particularly those related to copyright and income distribution for songwriters. Public opinions on these issues are widely expressed through Twitter, making it a valuable data source for sentiment analysis. This study aims to analyze public sentiment toward song royalty issues in the Indonesian music industry and compare the performance of Multinomial Naive Bayes and Support Vector Machine (SVM) algorithms using TF-IDF weighting. This study contributes through the implementation of semi-manual labeling, the use of a stratified 5-fold cross-validation approach, and multi-metric evaluation to obtain more representative sentiment classification results on song royalty issues in Indonesian social media. The initial dataset was collected through Twitter scraping using keywords related to song royalties and music copyright. The data were then processed through preprocessing stages, including case folding, cleaning, tokenization, stopword removal, and stemming. Sentiment labeling was conducted using a semi-manual approach, involving lexicon-based pre-labeling followed by manual verification into three sentiment categories: positive, negative, and neutral. Model evaluation was performed using stratified 5-fold cross-validation with accuracy, precision, recall, and F1-score metrics. The results indicate that the SVM algorithm outperformed Multinomial Naive Bayes, achieving an accuracy of 93.21%, while Multinomial Naive Bayes obtained an accuracy of 82.53%. These findings demonstrate that SVM is more effective in handling high-dimensional textual data represented using TF-IDF for Indonesian sentiment analysis. This study is expected to provide insights into public perceptions regarding song royalty issues and serve as a reference for sentiment analysis applications on Indonesian social media data.

Perbandingan Kinerja Model Long Short-Term Memory (LSTM) dan Gated Recurrent Unit (GRU) Untuk Prediksi Harga Cryptocurrency

2026-06-23T14:14:09+07:00

Extreme volatility in the cryptocurrency market poses substantial financial risks, necessitating precision forecasting systems. The limitations of conventional statistical models in capturing non-linear dynamics have prompted the adoption of deep learning approaches. This study evaluates the comparative performance of Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures in predicting the closing prices of Bitcoin (BTC), Ethereum (ETH), and Solana (SOL). Experiments utilized historical data from January 2023 to April 2026, partitioned with an 80:20 train-test ratio using a 60-day sliding window sequence. Dropout-based regularization and Early Stopping were implemented to prevent overfitting. As an empirical contribution, this research univariately examines the trade-off between memory cell complexity (LSTM) and architectural efficiency (GRU) on high-volatility data. The results demonstrate that GRU consistently outperforms LSTM across all instruments, reducing the Mean Absolute Percentage Error (MAPE) to a range of 2.93%-4.89%. Regarding computational efficiency, the GRU architecture reduced training duration by 13.33% to 36.92% compared to LSTM. Practically, these findings recommend GRU as an effective and efficient algorithmic foundation for algorithmic trading systems and digital portfolio risk management.

Evaluasi Kinerja Sensor FC-28 Untuk Monitoring Kelembapan Tanah Tanaman Indoor Menggunakan Confusion Matrix

2026-07-07T22:58:32+07:00

Technology plays an important role in transforming complex challenges into simpler and more practical processes, including soil moisture monitoring for indoor plants. Soil moisture is one of the most critical factors affecting the survival and growth of indoor plants. However, indoor plant owners often face difficulties in determining the moisture condition of the soil, creating the need for a simple and easy-to-use tool for monitoring soil moisture. The FC-28 soil moisture sensor is designed as a simple drought detection device equipped with an LED indicator to provide information regarding soil moisture conditions. The sensor operates by comparing the detected moisture level with a predefined comparator threshold. When the soil moisture falls below the threshold, the LED indicator turns on, indicating that the soil is dry. Conversely, when the moisture level remains above the threshold, the LED stays off, indicating that the soil is sufficiently moist. This study aims to evaluate the performance of the FC-28 sensor in detecting soil moisture conditions in indoor plants. An experimental approach was employed by conducting tests under dry and wet soil conditions. The sensor detection results were compared with the actual soil conditions (ground truth) using the Confusion Matrix method to obtain accuracy, sensitivity, specificity, and error rate values. The results indicate that the FC-28 sensor can consistently distinguish between dry and wet soil conditions in accordance with the actual conditions observed during the testing scenarios. This study contributes by providing a quantitative performance evaluation of the FC-28 sensor using the Confusion Matrix approach through accuracy, sensitivity, specificity, and error rate parameters for indoor soil moisture monitoring applications. The evaluation results demonstrate a high level of agreement between the sensor outputs and the actual soil conditions within the limited testing scope. These findings suggest that the FC-28 sensor can be utilized as a simple and cost-effective soil moisture monitoring tool for indoor plants under controlled environmental conditions.

Classification of Cancer Epitope Mutations Using Random Forest, SVM and Sequence-Derived Features

2026-06-23T14:45:07+07:00

Cancer remains a major global health challenge, with increasing incidence and mortality rates worldwide. One promising approach in cancer immunotherapy is the identification of epitopes, particularly mutated epitopes that play a critical role in immune recognition. This study aims to classify cancer epitope mutations using machine learning approaches based on sequence-derived physicochemical features. A dataset consisting of 234 samples was used, comprising into tumor and mutated tumor epitopes. Feature extraction was performed using physicochemical properties such as molecular weight, isoelectric point, aliphatic index, aromaticity, and hydrophobicity. Two machine learning models, namely Support Vector Machine and Random Forest were selected due to their proven effectiveness in biological sequence classification tasks and their robustness in handling small-to-medium-sized datasets. The results show that Random Forest achieved the best performance with an accuracy of 83% and a macro average F1-score of 0.70, while consistently outperforming the Support Vector Machine model across all data partition scenarios. However, further analysis revealed that the model exhibits limitations in detecting mutated epitopes, as indicated by a relatively high false-negative rate. This issue is likely due to the use of global sequence-derived features, which may not effectively capture local variations caused by mutations. This study contributes to the field of computational immunology by providing a comparative evaluation of Random Forest and Support Vector Machine for mutation epitope classification using sequence-derived physicochemical features. In addition, the integration of machine learning analysis with structural bioinformatics interpretation offers further biological insight into mutation-associated epitopes and their potential relevance in cancer immunotherapy.

Perbandingan Kinerja Random forest dan SVM Pada Klasifikasi Tingkat Kekumuhan Permukiman Menggunakan SMOTE

2026-06-23T15:00:05+07:00

Classifying slum levels is essential for a structured, data-driven analysis of settlement conditions. This study compares the performance of Random forest and Support vector machine (SVM) in classifying slum levels in Pekanbaru City across two scenarios with and without SMOTE using slum indicator scoring data. Its contributions include analyzing SMOTE's impact on model performance and evaluating the top 10 features against the full feature set. The dataset comprises 992 RT-level records from Disperkim Pekanbaru City (2020, 2021, and 2023) featuring 16 slum indicator scores based on PUPR Ministerial Regulation No. 14/2018, categorized into three classes: Non-Slum, Low Slum, and Moderate Slum. Following the KDD process (selection, preprocessing, transformation, data mining, evaluation, and analysis), the data was split 80:20 using stratified sampling and evaluated based on accuracy, precision, recall, F1-score, and confusion matrix. Results show that the Linear SVM without SMOTE achieved perfect evaluation metrics (1.0000); however, this is interpreted cautiously as the class labels derive from strict regulatory scoring rules, making class boundaries inherently linear. Random forest saw its F1-score rise from 0.9660 to 0.9700 after SMOTE, while the most significant improvement occurred in SVM RBF, jumping from 0.9214 to 0.9779. Testing the top 10 features led to a decreased F1-score across models, indicating that utilizing all 16 features remains optimal for this dataset.

Komparasi Kinerja Arsitektur MobileNetV2 dan EfficientNetB0 Untuk Klasifikasi Penyakit Daun Tanaman Kedelai

2026-06-23T15:12:41+07:00

Soybean leaf diseases can reduce the quality and productivity of plants, so an accurate and efficient detection method is needed. This study aims to compare the performance of the MobileNetV2 and EfficientNetB0 architectures in classifying soybean leaf diseases using a deep learning-based transfer learning approach. The dataset used consists of soybean leaf images grouped into several disease classes, then divided into training (80%), validation (10%), and testing (10%) data. The pre-processing stage includes resizing the images to 224 × 224 pixels, normalizing pixel values, and data augmentation in the form of rotation, shifting, zooming, and horizontal flipping. The training process is carried out using the Adam optimizer with a learning rate and applying Early Stopping to reduce the risk of overfitting. Model evaluation is carried out using a confusion matrix, accuracy, precision, recall, and F1-score. The results show that MobileNetV2 obtains an accuracy of 81%, higher than EfficientNetB0 which obtains an accuracy of 70%. The contribution of this study is to provide a comparative analysis of the effectiveness of both architectures in classifying soybean leaf diseases and to show that MobileNetV2 is more optimal for application to the dataset used.

Perbandingan Metode Elbow dan Silhouette Coefficient pada K-Means untuk Pengelompokan Wilayah Berdasarkan Indeks Pembangunan Manusia

2026-06-23T15:24:28+07:00

One of the primary metrics for evaluating the effectiveness of efforts to improve people’s well-being through human development is the Human Development Index (HDI). Although Indonesia’s HDI has continued to improve, disparities in human development remain evident across regions, particularly in Bali, West Nusa Tenggara (NTB), and East Nusa Tenggara (NTT). Identifying regions with similar HDI characteristics is important for supporting more targeted development policies. However, the performance of K-Means clustering is highly influenced by the number of clusters used, making the selection of an appropriate cluster number essential. This study compares the Elbow Method and Silhouette Coefficient in determining the optimal number of clusters for 2024 HDI data covering 41 regencies and municipalities based on Life Expectancy, Expected Years of Schooling, Mean Years of Schooling, and Per Capita Expenditure. The results show that the Elbow Method produces three clusters, while the Silhouette Coefficient produces two clusters with a silhouette value of 0.5312. Evaluation using the Davies–Bouldin Index (DBI) indicates that the two-cluster solution achieves a lower DBI value (0.7350) than the three-cluster solution (1.0382). These findings suggest that the HDI structure in Bali, NTB, and NTT tends to form two major groups: regions with high human development and regions with medium-to-low human development. The results also indicate that the Silhouette Coefficient is more representative for determining the optimal number of clusters in HDI data with relatively similar regional characteristics. The clustering results may support policymakers in prioritizing development programs in education, health, and community welfare

Analysis of Public Opinion on TikTok Regarding the MBG Program Controversy Using the Support Vector Machine Algorithm

2026-07-01T13:51:25+07:00

This study examines the analysis of public opinion regarding the controversy surrounding the Free Nutritional Meal Program (MBG), a strategic policy of the Indonesian government aimed at reducing stunting rates and improving child nutrition. Despite its important social objectives, the program has sparked various public reactions concerning budget transparency, equitable distribution of aid, and food security. TikTok, as a social media platform with high levels of interaction, has become a primary platform for the public to express opinions on public policy. However, its use in sentiment analysis research remains relatively limited compared to other platforms such as Twitter and Instagram. This study aims to analyze public perceptions of the MBG program using a combination of the Support Vector Machine (SVM) algorithm and Word2vec. Research data was obtained through the collection of 2,381 TikTok comments, followed by preprocessing steps such as data cleaning, tokenization, slang normalization, stop-word removal, and stemming. After the data selection process, 2,376 comments were used in the lexicon-based sentiment labeling and classification process using SVM. The test results show that the SVM model achieved an accuracy of 80% before class imbalance handling, whereas after applying class imbalance handling techniques, the accuracy increased to 83%, with a weighted precision of 0.84, a recall of 0.83, and an F1 score of 0.83. This improvement indicates that processing the data in the database enhances the model’s ability to recognize all sentiment classes more evenly, particularly positive sentiment, which previously had a smaller dataset. The sentiment analysis results show that the majority of opinions are dominated by neutral and negative sentiments, reflecting public concerns regarding the program’s implementation effectiveness, budget management transparency, and equitable distribution. These findings suggest that public opinion on social media can be leveraged as a real-time source for evaluating government policies to help the government develop public communication strategies that are more transparent, responsive, and targeted toward the implementation of the MBG Program.

Comparative Study of Agglomerative Hierarchical Clustering and K-Means for Student Academic Stress Grouping

2026-07-01T15:24:49+07:00

Academic stress is a common problem experienced by college students due to high academic demands, parental expectations, and social pressures during their college years. The high levels of academic stress experienced by students underscore the need for a data-driven approach to more accurately identify and map students’ stress levels. This research aims to compare the performance of the Agglomerative Hierarchical Clustering (AHC) and K-Means methods in clustering students’ academic stress levels and to determine which method produces the best clustering quality. Data were obtained from the distribution of the Perception of Academic Stress Scale (PAS) questionnaire, consisting of 18 statement items, with 361 valid respondents from the Informatics Engineering Program at UIN SUSKA Riau, class of 2022–2025. The selection of the best linkage method in AHC was performed using the Cophentic Correlation Coefficient (CCC), where Ward Linkage was selected with the highest CCC value of 0.8180. Comparative evaluation was conducted using the Silhouette Coefficient, Davies-Bouldin Index, and Calinski-Harabasz Index for variations in the number of clusters from K=2 to K=7. The test results showed that AHC Ward Linkage with K=2 was the best configuration with a Silhouette Coefficient of 0.4407 and a Davies-Bouldin Index of 0.8373, outperforming K-Means, which only excelled in the Calinski-Harabasz Index with a value of 419.7405 The clustering resulted in two clusters: High Stress with 244 students (67.6%) and Low Stress with 117 students (32.4%). The 2023 and 2024 cohorts had the highest proportions of high stress at 90.4% and 90.6%, respectively. This research contributes empirical evidence comparing hierarchy-based and partition-based clustering methods for academic stress data, while also demonstrating the use of the Cophenetic Correlation Coefficient as an objective basis for linkage method selection in AHC. It is hoped that the results of this study can serve as a basis for the institution in designing targeted mental health intervention programs for students.

ShelfMind: Demand Forecasting and Stock Control for Retail SMEs using Global XGBoost and Context-Injected AI Chatbot

2026-07-01T15:33:59+07:00

SME retail stores in Indonesia face serious challenges in inventory management: stockouts of popular products, waste from perishable items that expire before being sold, and restocking decisions relying on owner intuition. This research develops ShelfMind, a web-based inventory management system integrating three main components: a demand prediction model using a global XGBoost algorithm, an AI chatbot based on real-time data context injection, and an active notification system for low-stock and expiry risk alerts. Dataset from Toko Tika Baru covered 47 products in 11 categories with 8,225 daily sales records over 205 days. The XGBoost model was trained with 29 time-engineered features and achieved an MAE of 0.4107 units/day and RMSE of 0.5081 units/day on a 40-day test set. The global XGBoost method and context injection approach were selected due to their computational efficiency, avoiding the high costs associated with LLM fine-tuning, thus making it an ideal solution for SMEs with limited budget and infrastructure data. The AI chatbot, using context injection from inventory, sales, and XGBoost forecast data, achieved an average relevance and data accuracy of 4.94/5 across 20 test scenarios with an average response time of 2.94 seconds. Black-box functional testing of 31 scenarios passed completely. Usability evaluation scored 3.77/5, with notes on improving the communication of prediction features to non-technical users. This system proves that integrating XGBoost and LLM in an SME inventory management platform is technically feasible and provides tangible value for retail operations.

Comparative Evaluation of Local Llama and Mistral Models in RAG for Helpdesk Documentation

2026-07-02T10:51:57+07:00

The increasing volume of technical documentation and repetitive support requests at PT XYZ has made it difficult for helpdesk personnel to retrieve accurate information efficiently and consistently. The 114 internal applications that PT XYZ oversees have different documentation, troubleshooting protocols, and frequently asked questions, which leads to dispersed knowledge sources and drawn-out problem solving. This study uses locally deployed Large Language Models (LLMs), specifically Llama and Mistral, to design and assess a Retrieval-Augmented Generation (RAG) based chatbot for internal helpdesk knowledge retrieval in order to solve this issue. The suggested solution combines generative language models with semantic document retrieval in PostgreSQL using PGVector. The study compares the effectiveness of local LLMs in RAG and non-RAG configurations, employing 1,068 internal question-answer pairs for knowledge retrieval and 214 evaluation questions for performance assessment. ROUGE, BLEU, and Cosine Similarity measures are used for evaluation. According to experimental findings, RAG greatly enhances both models' performance. While Mistral improved from 0.0984 to 0.8983, Llama ROUGE-1 score rose from 0.1710 to 0.7345. With a ROUGE-1 score of 0.8983, a BLEU-1 score of 0.8161, and a Cosine Similarity score of 0.8916, Mistral with RAG performed the best of all setups. These results show that integrating RAG with locally installed LLMs greatly improves contextual correctness and response relevance, enabling safe and scalable implementation for enterprise helpdesk knowledge management.

Comparison of Random Forest and XGBoost Methods Based on Hyperparameter Tuning for Classification of Customer Churn Rate of Telecommunication Providers

2026-07-02T11:00:37+07:00

Customer churn represents one of the most critical challenges in the telecommunications industry, as the cost of acquiring new customers significantly outweighs the expense of retaining existing ones. High churn rates directly impact corporate revenue stability and market competitiveness, necessitating the development of precise predictive systems. This study presents a comprehensive comparative analysis of two prominent ensemble learning algorithms, Random Forest (RF) and Extreme Gradient Boosting (XGBoost), to establish a robust predictive framework for identifying potential churners using a large-scale Telco subscriber dataset. To ensure the reliability and scientific validity of the comparison, the research methodology incorporates the Synthetic Minority Over-sampling Technique (SMOTE) to rigorously address the inherent class imbalance within the dataset, ensuring that the minority churn class is adequately represented during the training phase to avoid model bias. Furthermore, a systematic hyperparameter tuning process was executed via GridSearchCV, exploring multiple combinations of estimators, depth, and learning rates to identify the optimal configurations for both algorithms. The experimental results reveal that while both models are highly effective, Random Forest slightly outperformed XGBoost, achieving an overall accuracy of 77.54% and a balanced F1-score of 0.616, compared to XGBoost’s accuracy of 76.54% and F1-score of 0.605. Notably, although both models demonstrated an identical recall rate of 67.64%, Random Forest exhibited superior precision (56.47% vs. 54.76%), which is vital for minimizing false positives and ensuring cost-effective retention campaigns. Feature importance analysis, conducted through Gini impurity and gain metrics, further identified tenure, total charges, and month-to-month contract types as the primary drivers of customer attrition. This study concludes that an optimized Random Forest model provides a more stable and accurate framework for telecommunication providers to proactively mitigate customer turnover. The findings offer valuable business intelligence, allowing stakeholders to transition from reactive measures to proactive, data-driven loyalty programs that enhance long-term business sustainability.

Peramalan Volume Transaksi Uang Elektronik Menggunakan Hybrid SARIMA-LSTM dengan Integrasi Klasifikasi Berita IndoBERT

2026-07-03T12:47:21+07:00

Amid the accelerating transformation of digital payments, the volume of electronic money transactions in Indonesia has grown rapidly, exhibiting volatile patterns and frequent structural breaks. These characteristics make forecasting particularly challenging, even though accurate projections are needed to support monitoring and planning. Existing approaches generally rely on univariate models or macroeconomic variables that suffer from publication lag, rendering them less responsive to rapidly evolving economic shocks. To address these limitations, this study proposes a hybrid SARIMA-LSTM framework that incorporates online news classification, generated using IndoBERT, as an exogenous variable providing more timely signals than macroeconomic indicators. Approximately 13,000 news headlines were collected through web scraping and classified into three impact categories pendorong (driving), penghambat (inhibiting), and informatif (informative) using a fine-tuned IndoBERT model. The classification results were aggregated monthly into exogenous features, with SARIMA capturing the linear trend and seasonal patterns while LSTM modeled the nonlinear residuals. The best-performing model, a hybrid SARIMA-LSTM with the informative news variable, achieved a MAPE of 14.42%, outperforming the hybrid model without exogenous variables (15.24%), the standalone SARIMA (17.77%), and the standalone LSTM (23.37%). The main contribution is to evaluate the hybrid SARIMA-LSTM architecture and asses whether integrating online news classification can serve as a complementary indicator, reducing reliance on lag-affected data. Although the improvement from the news variable was marginal and more pronounced during volatile periods, its consistent superiority indicates that informative news carries nonlinear signals beneficial for the forecasting model.

Klasifikasi Sentimen Program Makan Bergizi Gratis di Platform X dengan TF-IDF dan Class-Weighted LinearSVC

2026-07-03T22:18:45+07:00

This study investigates the optimization of public sentiment classification related to the Free Nutritious Meal Program on Platform X by focusing on social media data that are brief, informal, noisy, and class-imbalanced. The data were collected through scraping from January 5, 2025, to February 26, 2026, producing 7,659 initial posts. After selection, preprocessing, cleaning, and labeling, 5,307 texts were used as the modeling dataset, consisting of 2,183 positive, 2,391 neutral, and 733 negative sentiments. The texts were transformed into numerical features using Term Frequency-Inverse Document Frequency with 5,000 features and unigram-bigram settings. This study evaluated Multinomial Naive Bayes, Logistic Regression, LinearSVC, and Random Forest as comparison models. Optimization was performed using GridSearchCV, while class_weight balanced was applied to improve the model’s ability to identify the smaller negative class. The evaluation results show that LinearSVC with class_weight balanced produced the most balanced performance, with an accuracy of 0.9011, macro F1-score of 0.8989, weighted F1-score of 0.9011, negative recall of 0.8367, and negative F1-score of 0.8913. The contribution of this study lies in emphasizing minority-class evaluation through negative recall and F1-score, making the model a preliminary component for digital public opinion monitoring that is more sensitive to criticism and complaints.

Komparasi Decision Tree, Random Forest, dan XGBoost untuk Deteksi Dini Risiko Akademik, Kehadiran, dan Karakter Siswa

2026-07-03T23:01:25+07:00

Schools routinely collect academic, attendance, and character data, but their use remains limited to administrative recording. As a result, changes in student conditions are not always transformed into early risk signals, such as academic decline, repeated absenteeism, or accumulated character-related violations. The problem addressed in this study is how to evaluate an early detection model that can capture these three risk dimensions simultaneously using imbalanced and temporally ordered monthly longitudinal data. This study compares Decision Tree, Random Forest, and XGBoost for detecting Academic Risk, Attendance Risk, and Character Risk in the following month. The dataset consists of longitudinal student records from SMKS YUM Pesantren Teknologi Riau in 2025, covering 206 students across 12 months. The raw dataset contained 2,472 rows and was reduced to 2,266 valid rows after feature engineering and time-shifting target construction. Model evaluation used rolling-origin temporal evaluation, majority baseline, temporal persistence baseline, and imbalance handling strategies: no balancing, class_weight, and partial SMOTE on training data. Macro F1 was used as the primary metric because the class distribution was imbalanced. The best models were Random Forest with class_weight for Academic Risk, with Macro F1 of 0.5005 and High-Risk recall of 0.6185; Decision Tree with class_weight for Attendance Risk, with Macro F1 of 0.4586 and High-Risk recall of 0.5339; and Random Forest with class_weight for Character Risk, with Macro F1 of 0.4761 and High-Risk recall of 0.5698. These findings indicate that machine learning models can provide early High-Risk signals, but predictions should be verified by teachers or homeroom teachers.

Recommendation of Vocational Major Recommendation Using Artificial Neural Networks with ADAM Optimization

2026-07-03T23:12:31+07:00

This study proposes an Artificial Neural Network (ANN)-based recommendation framework for vocational major selection using student academic-feature profiles and ADAM optimization. The dataset consisted of 9,984 student records from West Java, derived from the school proposal recapitulation linked to DAPODIK. The predictive task was formulated as a six-class classification problem covering major categories in vocational education. Data preprocessing included data checking, target encoding, feature standardization, and train-test splitting. The ANN model used two hidden layers with ReLU activation and a softmax output layer, and was trained for 50 epochs using the ADAM optimizer. To enhance practical usability, predicted class probabilities were combined with major-specific qualification thresholds to enable the system to produce both primary and alternative recommendations. Experimental results showed that the model achieved 82% accuracy, an average precision of 0.88, a recall of 0.82, and an AUC of 0.8748, indicating useful discriminative performance for vocational major recommendation. The study provides a clearer distinction among the recommendation basis, predictive model, and operational qualification screening in educational decision support. However, the system still relies mainly on academic features and has not yet been benchmarked against alternative classifiers. Future work should incorporate interest and aptitude variables, class-wise error analysis, and comparative baseline evaluation.

Multiclass Herbal Plant Classification Using CNN Architectures: A Comparative Study of MobileNetV2, EfficientNetV2B0, NASNetMobile, and InceptionV3

2026-07-04T20:41:55+07:00

Indonesia is a country with an exceptionally rich biodiversity; herbal plants offer a wide range of benefits in the fields of health and traditional medicine. However, the process of identifying herbal leaves is still done manually and is often prone to errors due to similarities in shape, color, and texture among leaves. This study aims to develop a multi-class herbal plant leaf image classification system based on a Convolutional Neural Network (CNN) by comparing four transfer learning architectures: MobileNetV2, EfficientNetV2B0, NASNetMobile, and InceptionV3. The dataset used consists of 10 classes of herbal plant leaves. The contributions of this study include a comparative analysis of four CNN architectures for multi-class classification, an evaluation of the effectiveness of preprocessing and data augmentation on a limited dataset, and recommendations for the most optimal model based on accuracy and computational efficiency. The experimental results show that all models achieved validation accuracy above 98%. InceptionV3 delivered the best performance with a test accuracy of 97%, precision of 90%, and accuracy, recall, and F1-score of 89% respectively, demonstrating good generalization ability. Meanwhile, MobileNetV2 offers the best balance between accuracy and computational efficiency, making it a promising candidate for herbal plant identification systems based on mobile devices or in environments with limited computational resources.

Perbandingan Kinerja Model LSTM dengan dan tanpa Indeks Fear & Greed dalam Prediksi Harga Bitcoin Berbasis Data Time Series

2026-07-04T20:53:16+07:00

Bitcoin is a highly volatile digital asset whose price movements are influenced by historical market patterns and investor sentiment, making accurate prediction a significant challenge for conventional forecasting approaches. This study compares the predictive performance of the Long Short-Term Memory (LSTM) model in forecasting Bitcoin prices under two configurations: without and with the integration of the Crypto Fear & Greed Index as a market sentiment feature. The dataset consists of time series data including opening, highest, lowest, and closing prices obtained from the CoinGecko API, along with daily sentiment scores collected from the Alternative.me API. The data preprocessing stages include normalization using the Min-Max Scaler and sequence construction using the sliding window method, both of which are standard practices in deep learning-based time series forecasting. Model performance was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE) as quantitative error metrics. The experimental results show that the LSTM model without the Fear & Greed Index produced an MAE of 1011.34, an RMSE of 1340.55, and a MAPE of 1.20%, while the model incorporating the Fear & Greed Index achieved an MAE of 979.95, an RMSE of 1349.07, and a MAPE of 1.18%. These findings indicate that the inclusion of market sentiment improves overall prediction accuracy, as reflected by the reduction in MAE and MAPE, despite a marginal increase in RMSE. The main contribution of this study is providing empirical evidence regarding the impact of integrating the Crypto Fear & Greed Index as a market sentiment indicator on improving the performance of LSTM models for Bitcoin price prediction, as well as offering a direct comparison between models based solely on historical price data and models that combine historical price data with market sentiment information. This study concludes that combining deep learning techniques with measurable sentiment indicators provides a more comprehensive and effective framework for Bitcoin price prediction.

Comparison of XGBoost and Random Forest for Prediction of Male Fertility Status Based on Semen Analysis Parameters

2026-07-04T21:15:20+07:00

Male infertility is a major reproductive health problem that contributes to approximately half of infertility cases among couples of reproductive age. Accurate evaluation of male fertility status commonly relies on semen analysis, including parameters such as semen volume, sperm concentration, motility, morphology, and vitality. However, manual interpretation of these parameters remains time-consuming and is highly dependent on clinical expertise. This study aims to compare the performance of the XGBoost and Random Forest algorithms in predicting male fertility status based on semen analysis parameters. The study employed a secondary dataset consisting of 1,000 semen analysis records with 11 predictor variables and one target variable representing fertility status. Data preprocessing included categorical encoding, data cleaning, and an 80:20 train–test split before model development and evaluation using accuracy, precision, recall, F1-score, and ROC-AUC. Experimental results showed that XGBoost outperformed Random Forest, achieving an accuracy of 99.00%, precision of 100.00%, recall of 96.88%, F1-score of 98.41%, and ROC-AUC of 99.86%, while Random Forest achieved an accuracy of 97.50%. Feature importance analysis identified Total Motility, Vitality, and Progressive Motility as the most influential predictors of male fertility status. The main contribution of this study is the direct empirical comparison of Random Forest and XGBoost under identical experimental settings using comprehensive semen analysis parameters, providing evidence on the relative effectiveness of ensemble learning algorithms for male fertility status prediction. Although the proposed model demonstrates excellent predictive performance, it was developed using secondary data and is intended to support, rather than replace, clinical decision-making. Future studies should validate the model using larger multicenter clinical datasets to improve its generalizability and practical applicability.

Effect of Word Embedding on Indonesian Social Media Hate Speech Classification Using Hybrid CNN-SimpleRNN

2026-07-04T21:26:56+07:00

The rapid rise in the number of people using social media platforms, specifically platform X, poses great difficulties for spotting any possible harm from hate speech. In particular, due to the high informality of discourse of Indonesian internet users, expressed through slang, abbreviation, and distorted spelling, automatic moderation becomes even more complicated. The current study attempts at classifying hate speech on social media platform X, proposing a hybrid model of CNN combined with a Bidirectional SimpleRNN architecture, alongside a comparative study on word embedding approaches. The combination between CNN and SimpleRNN is chosen because it allows for exploiting both CNN's capability to extract local spatial features by finding toxic n-grams and the strength of Bidirectional SimpleRNN for capturing long contextual dependency in the sequence of text data. Given that the problem of OOV is highly significant, TF-IDF, FastText, and Word2Vec have been rigorously tested not only separately but also combined in different ways. Compared to the baseline configuration using standalone TF-IDF (which achieved 84.56% accuracy), the results show that the use of the hybrid TF-IDF + FastText provided the best performance with average accuracy of 86.49%, average precision of 86.32%, recall of 86.80% and F1 score of 86.55%. Conversely, the combination of multiple dense semantic vectors (Word2Vec and FastText) led to semantic drift and feature ambiguity; this created feature overlap and computational noise that obscured classification decision boundaries, resulting in redundancy and poor performance. It shows that the combination of lexico-statistical significance and semantic subword context greatly contributes to achieving better results in the Indonesian language setting and is very resistant to slang and OOV words found in digital settings. This study contributes to the field of natural language processing by providing a lightweight, highly accurate, and computationally efficient lexico-semantic framework tailored for moderating highly unstructured Indonesian social media text.

Fusi Cross-Attention CNN–Transformer untuk Klasifikasi Multi-Kelas Acute Lymphoblastic Leukemia

2026-07-04T21:42:24+07:00

Acute Lymphoblastic Leukemia (ALL) is a hematological malignancy that requires a quick and accurate initial examination. Peripheral Blood Smear (PBS) imaging can be used as a source of cell morphology information, but image-based classification still faces challenges due to variations in the shape, color, and structure of blood cells. This study proposes an ALL multi-class classification model based on CNN–Transformer cross-attention fusion with two image inputs, namely the original PBS image and the segmented image that is already available in the dataset. The main contribution of this study lies in the integration of the local features of the CNN and the global features of the Transformer through the cross-attention mechanism, as well as the evaluation of the model components through baseline comparison and ablation studies. The dataset used consisted of 3,256 pairs of PBS images in four classes, namely Benign, Early, Pre, and Pro. The test results showed that the model obtained an accuracy of 0.9980 and a macro F1-score of 0.9975 on the test data. Nonetheless, this very high performance needs to be interpreted with caution as the research has not involved external validation based on different institutions or direct assessments by pathologists. Therefore, the proposed model is more appropriately positioned as a potential computational approach to the classification of PBS images, rather than as a final clinical diagnostic system. Advanced evaluation of external datasets, patient-based allocation schemes, and expert validation are required to assess the generalization and clinical relevance of the model.

Random Forest, LSTM, and IndoBERT Comparison for TikTok App Sentiment Analysis

2026-07-04T21:52:00+07:00

The rapid growth of social media platforms like TikTok has generated a massive volume of user reviews on the Google Play Store, serving as a critical indicator of application service quality. However, the unstructured nature of Indonesian social media text and the significant imbalance between sentiment classes pose substantial challenges for automated classification systems. Addressing this class imbalance is highly crucial for application developers, as critical negative and neutral feedback containing essential feature complaints is easily marginalized by the overwhelming majority of positive reviews, leading to biased operational insights. This research conducted a comprehensive comparative study of three distinct computational paradigms: Random Forest, Long Short-Term Memory (LSTM), and IndoBERT, to identify the most effective model for sentiment analysis. A dataset of 5,000 TikTok reviews was meticulously processed using a negation-aware preprocessing pipeline to preserve semantic integrity. To address class imbalance, architecture-specific techniques were deployed, including SMOTE for Random Forest, Class Weighting for LSTM, and Random OverSampling for IndoBERT. The experimental results demonstrate that IndoBERT significantly outperforms other models, achieving the highest global accuracy of 81% and a Macro F1-Score of 0.56. While Random Forest and LSTM yielded lower accuracies of 75% and 71%, respectively, they exhibited stability in predicting the majority class but struggled with the inherent ambiguity of neutral sentiments. The study concludes that IndoBERT’s bidirectional self-attention mechanism provides superior contextual understanding of Indonesian slang and non-formal syntax. This research contributes a robust framework for application developers to monitor public opinion objectively. Furthermore, the findings highlight that despite advanced balancing techniques, the "neutrality bottleneck" remains a challenge, suggesting that future research should explore aspect-based sentiment analysis to enhance classification granularity in the Indonesian NLP domain.

Analisis Sentimen dan Evolusi Topik terhadap Program Makan Bergizi Gratis Menggunakan IndoBERT dan cDTM

2026-07-06T23:35:01+07:00

This study aims to analyze public sentiment and the development of discussion topics related to the MBG program. Sentiment analysis was conducted using the IndoBERT model, while evolution topic analysis used the Continuous-Time Dynamic Topic Model (cDTM). The evaluation results showed that the IndoBERT model was able to classify sentiment with an accuracy value of 92.5% and an F1-score of 0.924. Integration between IndoBERT and cDTM showed a dominance of negative sentiment, especially in topics related to program implementation, while positive sentiment appeared more often in topics related to health and nutrition. The integration of sentiment and temporal topic analysis provides a more comprehensive understanding of the dynamics of public opinion regarding the MBG program.

Hybrid Two-Stage CatBoost and Multilayer Perceptron Model for Sleep Disorder Classification

2026-07-07T00:11:52+07:00

Sleep disorders are a health problem that significantly impacts quality of life and potentially increases the risk of various chronic diseases. Conventional sleep disorder diagnosis generally requires expensive and complex examinations, so an alternative, non-invasive data-driven approach is needed. This study proposes a sleep disorder classification approach based on tabular medical data using a Hybrid Two-Stage architecture. The proposed approach integrates the CatBoost algorithm as a binary screening stage to distinguish healthy individuals from individuals with sleep disorders, and a Multilayer Perceptron (MLP) as a latent feature extractor, combined with CatBoost to classify sleep disorder subtypes, namely insomnia and sleep apnea. The datasets used were obtained from two public data sources and evaluated using a stratified k-fold cross-validation scheme. Class imbalance was addressed using the SMOTE-ENN technique, while hyperparameter optimization was applied as part of the model training pipeline. Performance evaluation was conducted using accuracy, Macro-F1, and Matthews Correlation Coefficient (MCC) metrics. Experimental results show that the Hybrid Two-Stage architecture achieves an accuracy of 94.1%, a Macro-F1 of 0.90, and an MCC of 0.88, and exhibits stable performance across a wide range of fold variations. These results demonstrate that the hybrid two-stage approach is effective in improving the performance of sleep disorder classification based on medical tabular data. The main contribution of this study is the development of a two-stage hybrid classification framework that explicitly separates healthy-disorder screening and sleep disorder subtype classification, while integrating SMOTE-ENN, Optuna-based hyperparameter optimization, and MLP-derived latent feature representation to improve classification stability on imbalanced medical tabular data.

Analisis Pengaruh Chi-Square Feature Selection terhadap Kinerja Random Forest dan XGBoost dalam Prediksi Konversi Pengunjung Website

2026-07-07T12:49:49+07:00

The increasing number of visitors to e-commerce websites is not always accompanied by a corresponding increase in purchase transactions, making it difficult for companies to identify visitors with high conversion potential. In addition, using all available attributes may increase model complexity without necessarily improving predictive performance. This study analyzes the impact of Chi-Square Feature Selection on the performance of Random Forest and Extreme Gradient Boosting (XGBoost) in predicting website visitor conversion. The study uses the Online Shoppers Purchasing Intention dataset consisting of 12,330 instances with 17 predictor attributes and one target attribute. The research process includes exploratory data analysis, preprocessing, Chi-Square-based feature selection, classification model development using Random Forest and XGBoost, and evaluation using Accuracy, Precision, Recall, F1-Score, Matthews Correlation Coefficient (MCC), and Area Under the Curve (AUC-ROC). Four experimental scenarios were evaluated: all features (baseline), Top-15, Top-10, and Top-5 selected features. The results show that the baseline model using all features achieved the best overall performance. The Random Forest baseline model obtained an Accuracy of 90.05%, Precision of 73.94%, F1-Score of 63.13%, and MCC of 0.5835, while the XGBoost baseline model achieved the highest AUC-ROC of 0.9271. Furthermore, PageValues, BounceRates, ExitRates, ProductRelated_Duration, and ProductRelated were identified as the most influential features affecting visitor conversion. The main contribution of this study is providing empirical evidence that Chi-Square Feature Selection is more effective in reducing feature complexity and identifying relevant attributes than improving classification performance on the Online Shoppers Purchasing Intention dataset, offering practical guidance for feature selection strategies in machine learning-based website conversion prediction

Evaluasi Komparatif Algoritma Decision Tree, Random Forest, dan XGBoost untuk Software Defect Prediction Menggunakan Dataset NASA Software Metrics

2026-07-07T13:32:06+07:00

Software Defect Prediction (SDP) has become an important approach for identifying software modules that are likely to contain defects during the early stages of software development. However, the performance of prediction algorithms remains highly dependent on dataset characteristics, and no single algorithm has consistently demonstrated superior performance across different datasets. In addition, the use of synthetic datasets in SDP research still requires empirical validation to ensure that their characteristics remain representative of benchmark datasets such as the NASA Metric Data Program (NASA MDP). Therefore, this study aims to compare the performance of Decision Tree, Random Forest, and XGBoost using the Playground Series Season 3 Episode 23 dataset, a synthetic dataset developed based on the characteristics of the NASA MDP dataset. Prior to model training, the dataset underwent preprocessing, including missing value imputation, label encoding, and feature standardization. Model performance was evaluated using 10-fold stratified cross-validation with Accuracy, F1-Score, and AUC-ROC as the primary evaluation metrics. The experimental results indicate that ensemble learning methods achieved competitive performance compared with the single-classifier approach. Random Forest achieved the highest Accuracy of 0.8144, while XGBoost obtained the highest AUC-ROC score of 0.7929, indicating strong capability in distinguishing between defective and non-defective software modules on the evaluated dataset. Furthermore, feature importance analysis identified Lines of Code (LOC), Cyclomatic Complexity, Halstead Volume, IOCode, and branchCount as the most influential factors affecting software defect prediction. Based on the experimental results obtained from the selected dataset, the ensemble learning approach demonstrated competitive predictive performance and may be considered a promising alternative for developing Software Defect Prediction models. Nevertheless, the selection of the most appropriate algorithm should remain dependent on dataset characteristics and specific implementation requirements

Optimisasi Menu Balita Berdasarkan Gizi Harian Menggunakan XGBoost, Genetic Algorithm dan Particle Swarm Optimization

2026-07-08T23:50:16+07:00

Toddler menu planning requires a systematic approach, as menu recommendations must account for nutritional adequacy, the child’s nutritional status, allergy restrictions, and feasible menu combinations. This study develops an allergy-aware decision-support workflow by integrating XGBoost as a menu-class classifier and Genetic Algorithm and Particle Swarm Optimization for daily toddler menu optimization. This approach is chosen because conventional methods that directly apply population-level dietary references values do not adequately capture individual differences in nutritional status and usually do not treat food allergy as a hard constraint during menu generation. Combined of XGBoost and metaheuristics have instantly map non-linear patterns of a toddler’s physical condition while efficiently exploring millions of complex food combinations a task computationally unachievable by conventional manual methods. In the proposed workflow, the XGBoost output in the form of menu class (catch-up nutrition, balanced nutrition, or energy control) is mapped into personalized nutrient targets for energy, protein, fat, and carbohydrates. These targets become the main inputs to the fitness function and guide the allergy-safe candidate filtering step before optimization. The XGBoost model achieved an average accuracy of 96,4% and a weighted F1-score of 96,38% in cross-validation testing, along with an 87,18% accuracy in the hold-out test. Across six experimental categories, both GA and PSO successfully generated menus closely matching the nutritional targets, with RDI scores generally ranging from 94,6% to 99,7%. GA achieved lower nutritional deviation in the catch-up and balanced nutrition scenarios, whereas PSO performed better in energy-control scenarios and consistently required significantly less computation time (averaging 17,74 seconds compared to 79,78 seconds for GA). These findings indicate that GA is preferable when nutritional precision is prioritized, whereas PSO is more suitable for real-time, time-constrained system deployments. The main contributions of this study are the creation of an end-to-end computational framework that adapts to clinical conditions and toddler allergy safe, and the provision of empirical evidence for selecting metaheuristic algorithm for personalized nutrition recommendation systems.

Perbandingan Kinerja Model ARIMA dan LSTM pada Multi-Horizon Forecasting Harga Emas dengan Evaluasi Mean Directional Accuracy

2026-07-11T13:42:29+07:00

Precious metals, particularly gold, represent one of the most sought-after value-preserving investment instruments, yet their dynamic price fluctuations present significant challenges, making gold a difficult-to-predict yet crucial asset for investment decision-making. This study aims to forecast gold prices by comparing the performance of Autoregressive Integrated Moving Average (ARIMA) and Long Short-Term Memory (LSTM) forecasting models through three testing scenarios: single-step, multi-step forecasting, and rolling forecasting. This study utilizes 40 years of historical gold price data obtained from the public source Kaggle. The ARIMA model was implemented on stationary data, while LSTM was optimized with additional lag, volatility, and momentum features. Experimental results indicate that in the single-step scenario, both models produced equivalent accuracy with a MAPE below 1%. In the multi-step scenario, LSTM significantly outperformed ARIMA with a MAPE of 1.89% compared to 3.54%. In the rolling scenario, LSTM again performed better with a MAPE of 1.83% versus 3.52% for ARIMA. Conversely, ARIMA consistently recorded higher Mean Directional Accuracy (MDA) values across all scenarios, reaching 57.30% in the rolling forecast compared to LSTM's 46.07%, indicating ARIMA's advantage in identifying trend direction. This study concludes that the LSTM approach is more optimal for achieving numerical prediction precision over medium-term horizons, while the statistical ARIMA method is more reliable for accurately projecting market movement direction.

Perbandingan Kinerja Naive Bayes dan SVM dalam Analisis Sentimen Program Makanan Bergizi Gratis (MBG) sebagai Pendukung Pengambilan Keputusan

2026-07-11T21:55:14+07:00

The Free Nutritious Meal Program is one of the Indonesian government programs aimed at improving community nutritional quality and reducing stunting rates. The program has generated various publik responses expressed through media social platforms, particularly in YouTube comment sections. This study was conducted to analyze publik sentiment toward the Free Nutritious Meal Program (MBG) and to compare the performance of the Naive Bayes and Support Vector Machine (SVM) algorithms in classifying sentiment from YouTube user comments. The research data were obtained through a YouTube comment scraping process and then processed through several preprocessing stages, including cleaning, case folding, normalization, tokenization, stopword removal, and stemming. Furthermore, feature weighting was performed using the TF-IDF method, and data labeling was carried out using a lexicon-based approach. The sentiment classification process employed the Naive Bayes and Support Vector Machine (SVM) algorithms, while model evaluation was conducted using confusion matrix, accuracy, precision, precision, and f1-score metrics. The results showed that the Support Vector Machine (SVM) algorithm achieved better performance than Naive Bayes. The SVM algorithm obtained an accuracy of 77.4%, precision of 78.4%, precision of 77.4%, and f1-score of 77.6%, whereas the Naive Bayes algorithm achieved an accuracy of 70.5%, precision of 74.4%, precision of 70.5%, and f1-score of 67.7%. The main contribution of this study is the comparative evaluation of Naive Bayes and Support Vector Machine (SVM) for classifying public sentiment from YouTube comments related to the MBG program, providing empirical evidence on the most effective classification approach for supporting social media–based public opinion analysis of government policies.

Symptom-Based Classification of Migraine Severity Using Random Forest and LassoNet

2026-07-17T23:00:25+07:00

This study aims to develop and compare classification models for predicting symptom-based migraine intensity using a public dataset from Kaggle. The research was conducted on multiclass data with an imbalanced class distribution. Therefore, target creation and result interpretation must be carefully designed so that the resulting evaluation remains consistent with the characteristics of the data used. In this study, the Intensity variable was recoded into three operational classes: low, moderate, and high. The features used included symptoms, characteristics of migraine episodes, and `symptom_count`, which represents the number of symptoms in each sample. The two models compared were Random Forest and LassoNet, both of which were tested using Stratified 5-Fold Cross-Validation. Model performance was assessed using the Macro F1-score as the primary metric, Balanced Accuracy as the main supplementary metric, and Accuracy as a complementary metric. The test results showed that Random Forest performed better, with a Macro F1-score of 0.6366, Balanced Accuracy of 0.6181, and Accuracy of 0.6225. Meanwhile, LassoNet achieved a Macro F1-score of 0.2474, a Balanced Accuracy of 0.3333, and an Accuracy of 0.5900. These results indicate that symptom patterns in the dataset can still be utilized to distinguish migraine intensity within a computational classification framework, although the separation between closely related classes is not yet fully robust.

Optimasi Model Evaluasi Kinerja Karyawan Berbasis Rekam Jejak Digital Menggunakan PCA dan Algoritma Machine Learning

2026-07-22T13:50:47+07:00

In the era of digital transformation, organizations face challenges in evaluating employee performance objectively and based on data. Traditional performance appraisal systems often contain subjectivity and limitations in data integration, making them less effective in dynamic work environments. This study aims to develop a performance evaluation model based on digital footprints using machine learning and multivariate analysis. Digital footprints include work activity data (daily working hours, screen time, meetings, and emails), wearable data (physical steps, sleep duration, and stress levels), satisfaction (work-life balance, organizational support), capability (tech skills score, job level, and training), and organizational data (salary, incentives, and overtime). Principal Component Analysis (PCA) is used to reduce data dimensions and identify key performance indicators. Three machine learning algorithms—Decision Tree, Random Forest, and Gradient Boosting—are applied to classify employee performance into Low, Average, Good, and Excellent categories. Model evaluation is performed using accuracy, precision, recall, and F1-score metrics. The results show that the Gradient Boosting model combined with PCA delivers the best performance with an accuracy of 0.887 and an F1-score of 0.884. The application of PCA significantly improved classification model performance by reducing noise and multicollinearity in high-dimensional data. These findings highlight the great potential of leveraging employees' digital behavioral data to build a transparent and adaptive performance evaluation system. This study contributes to the development of intelligent HR management and supports data-driven decision-making in modern organizations.