Building of Informatics, Technology and Science (BITS)

Comparative Customer Segmentation Pipelines for E-Commerce Using K-Means-KNN and UMAP-K-Means-XGBoost

2026-09-08T16:23:21+07:00

The rapid expansion of e-commerce has generated massive volumes of customer data that remain underutilized for supporting Customer Relationship Management (CRM) strategies. Conventional customer segmentation approaches commonly employ a pipeline consisting of K-Means clustering followed by K-Nearest Neighbors (KNN) classification. However, this approach exhibits limitations in handling high-dimensional data and maintaining classification performance on large-scale datasets. This study presents a comparative analysis of two customer segmentation pipelines: the conventional K-Means-KNN pipeline and the proposed Uniform Manifold Approximation and Projection (UMAP)-K-Means-XGBoost pipeline. The experiments were conducted using the E-Commerce Shopper Behavior & Lifestyle dataset, comprising approximately one million customer records and eight selected features representing transactional, psychographic, and financial behavioral characteristics. Clustering performance was evaluated using the Silhouette Score, Davies-Bouldin Index, and Calinski-Harabasz Index, while classification performance was assessed using accuracy, precision, recall, and F1-score. Experimental results demonstrate that incorporating UMAP improves cluster separability by preserving the intrinsic structure of high dimensional data, whereas XGBoost consistently outperforms KNN in downstream classification, achieving an accuracy exceeding 99%. These findings indicate that the UMAP-K-Means-XGBoost pipeline provides a more robust, scalable, and interpretable framework for customer segmentation, thereby offering more reliable decision support for data-driven CRM strategies in e-commerce environments.

Decision Support System for Determining Signature Menus Using the Integration of BMW, MOORA, and Copeland Scores

2026-09-09T10:32:12+07:00

Menu optimisation is a crucial yet challenging challenge for business sustainability because the speciality coffee sector is marked by dynamic consumer tastes and a high degree of product diversity. This study tackles a basic operational issue at Inti Coffee, where a subjective, intuition-based method is currently used to identify "signature offerings"the important menu items that ought to be given priority for promotion, inventory, and resource allocation. This current process, which mainly depends on the owner, operational manager, and lead barista's intuition, is intrinsically vulnerable to individual cognitive biases, personal taste preferences, and inconsistent evaluation, which frequently results in less than ideal menu performance and lost revenue opportunities. The main goal of this research is to develop and deploy an online Decision Support System (DSS) that offers a transparent, data-driven, and organised framework for objectively ranking menu items in order to get around these restrictions. A hybrid multi-criteria decision-making (MCDM) technique is included into the suggested system to guarantee group unanimity and robustness. In order to minimise pairwise comparison inconsistencies and capture the knowledge of the three primary decision-makers, the Best-Worst Method (BWM) is first used to systematically extract the relative relevance weights of six different evaluation criteria. Second, a thorough dataset of 500 real sales transactions is used to assess and rank 51 menu alternatives using the MOORA (Multi-Objective Optimisation on the basis of Ratio Analysis) method. Both financial parameters (total items sold, HPP or cost of goods sold, and profit margin) and operational characteristics (uniqueness of taste score, preparation time, and ingredient lifetime) are included in the evaluation criteria. In order to successfully resolve any potential conflicts between decision-makers, the Copeland Score is finally used to combine the individual preference rankings into a single, collective group score. This study makes three main contributions: first, it develops a novel integrated DSS framework that integrates BWM, MOORA, and Copeland Score into a single unified workflow specifically designed for signature menu selection; second, it involves three different stakeholders (owner, operational manager, and head barista) in the group decision-making process, ensuring that the final recommendation reflects a balanced consensus rather than individual bias; and third, it creates a fully functional web-based system with statistical validation tools that empirically verify the accuracy and dependability of the recommendations using Spearman correlation, RMSE, and overlap ratio against actual customer preferences. The creation of a useful, web-based DSS that turns menu curating from an art to a science and offers a reproducible model for other food and beverage businesses is the main contribution of this research. With a MOORA score of 0.249180 and a Copeland Score of 50, the interim results show that the system is able to identify "Kopi Susu Aren" as the best-performing signature item. A poll involving 130 participants was carried out to verify the system's output against human judgement in the actual world. In addition to a low Root Mean Square Error (RMSE) of 5.5734 and an overlap ratio of 66.7%, the results show a strong positive correlation (Spearman's rho = 0.9238) between the system's ranks and participant feedback, proving the system's high accuracy and practical applicability. In order to provide scalability and usability for continuous operational choices, the DSS is implemented using a combination of PHP Native, Python, MySQL, and a testing dashboard based on Streamlit.

Perbandingan XGBoost dan Random Forest Untuk Prediksi Pembatalan Pesanan Marketplace Berbasis SHAP

2026-09-09T10:45:22+07:00

The high rate of order cancellations in marketplaces is a challenge that can impact operational efficiency, inventory management, and customer service quality. The ability to identify orders potentially canceled early is crucial so businesses can take more appropriate mitigation measures. The purpose of this study is to compare the performance of the Random Forest, and XGBoost algorithms in the marketplace order cancellations prediction process and to interpret the factors that influence predictions using SHAP (SHapley Additive exPlanations)-based explainable machine learning. The study used a marketplace transaction dataset that had undergone data cleaning, feature transformations, and class through the application of the Synthetic Minority Oversampling Technique (SMOTE) method. Next, the model performance is evaluated using the Accuracy, Precision, Recall, and F1-Score metrics. Four model scenarios were tested: Random Forest + SMOTE, Random Forest Tuned + SMOTE, XGBoost + SMOTE, and XGBoost Tuned + SMOTE. Based on the research results, it is indicated that the XGBoost Tuned + SMOTE model produces the best performance with accuracy values reaching 86.62%, precision reaching 51.43%, recall reaching 34.95%, and F1-Score reaching 41.62%. Meanwhile, the XGBoost + SMOTE model produced the highest recall of 37.09% with an F1-Score of 41.48%. Considering that the research objective is to detect potentially canceled orders in imbalance data, the XGBoost + SMOTE model was chosen as the best model because it produced the highest recall of 37,09% with a competitive F1-Score of 41,48%, making it more capable of detecting minority classes compared to other models. SHAP analysis showed that several transaction features contribute more dominantly to the prediction results produced by the model. These findings can be used as a basis for decision-making to reduce the risk of order cancellations and improve the effectiveness of marketplace operations.

Performance Evaluation of Naive Bayes, SVM, LSTM, and IndoBERT for Fashion Review Sentiment Classification

2026-09-09T11:13:29+07:00

The rapid growth of e-commerce has increased the volume of customer reviews, making it difficult for fashion businesses to manually identify customer sentiment, product concerns, and satisfaction drivers. This study aims to evaluate the performance of Naive Bayes, Support Vector Machine (SVM), Long Short-Term Memory (LSTM), and IndoBERT in classifying sentiment polarity in Indonesian fashion product reviews, while also identifying topic-level sentiment patterns using Latent Dirichlet Allocation (LDA). The dataset consisted of 23,487 Shopee Indonesia fashion reviews collected during the 2025 observation period. After data cleaning, duplicate removal, text normalization, tokenization, stopword removal, and lemmatization, 22,961 valid reviews were retained and categorized into positive, neutral, and negative classes using rating-based labeling. To reduce majority-class bias, stratified data splitting and class-weighted learning were applied, while five-fold cross-validation was used to evaluate model stability. The results show that IndoBERT achieved the highest performance with an accuracy of 93.41%, precision of 92.87%, recall of 92.15%, and F1-score of 92.51%, outperforming LSTM with 90.80% accuracy, SVM with 88.90%, and Naive Bayes with 84.60%. The findings indicate that transformer-based contextual representation is more effective in handling noisy Indonesian fashion review text, including informal expressions, abbreviations, and context-dependent sentiment. Topic-based analysis further revealed that positive sentiment was mainly associated with product quality, material comfort, design, and value, while negative sentiment was driven by sizing mismatch, product inconsistency, delivery delay, and customer service issues. This study contributes to sentiment classification research by providing a comparative evaluation of machine learning, deep learning, and transformer-based models, and offers practical insights for improving fashion e-commerce product communication, customer satisfaction, and data-driven digital marketing decisions.

Komparasi Algoritma Machine Learning pada Prediksi Kemenangan Tim VALORANT Berbasis Delta Analysis

2026-09-09T11:23:31+07:00

The development of the Valorant esports ecosystem through the VALORANT Champions Tour (VCT) creates a need for objective data analysis. Current analysis is often trapped in subjective bias, risking inaccurate predictions. To address this problem, this research compares various team win prediction methods in VCT 2025. The selection of the VCT 2025 dataset provides novelty as it represents a franchise league structure with much more evenly distributed game meta and tactical complexity than previous seasons. The compared algorithms include Naïve Bayes, Decision Tree, XGBoost, CatBoost, and Stacking Classifier. The main contribution of this research is the application of a differential feature extraction technique (Delta Analysis) to objectively measure the performance gap between teams. Comparative experimental results on 1,507 match observations show that the Stacking Classifier achieved the most optimal accuracy (75.72%) compared to XGBoost (75.19%), Naïve Bayes (74.26%), CatBoost (73.59%), and Decision Tree (72.86%). These findings prove that Delta Analysis-based meta-learning synergy can provide more stable analytical generalization in handling modern esports data complexity.

Superioritas Optimizer AdaGrad pada Arsitektur Spasio-Temporal PolyFace-LSTM untuk Prediksi Kepribadian Berbasis Video

2026-09-09T11:48:51+07:00

Conventional personality assessment generally relies on psychometric questionnaires, such as the Big Five Inventory (BFI). However, this method is susceptible to self-report bias and requires a relatively long evaluation time. On the other hand, the development of automated video-based personality analysis systems faces high computational challenges in simultaneously integrating facial spatial features and the temporal dynamics of micro-expressions. This study aims to propose a solution in the form of a hybrid spatio-temporal neural network architecture utilizing transfer learning techniques to predict personality more objectively. The proposed system integrates the Multi-Task Cascaded Convolutional Networks (MTCNN) algorithm for automatic face detection, the PolyFace model as a spatial feature extractor, and the Long Short-Term Memory (LSTM) algorithm to model temporal relationships between frames. Experiments were conducted by comparing three optimization algorithms, namely Adam, AdaGrad, and SGD, using the ChaLearn LAP 2017 dataset. The results show that the AdaGrad optimizer achieved the best generalization and prediction performance during the validation phase, with an average accuracy of 89.13%, outperforming SGD (88.34%) and Adam (88.33%). Theoretically and mathematically, the superiority of AdaGrad in this architecture is attributed to its ability to dynamically adjust the learning rate based on the accumulation of historical gradients. This mechanism makes AdaGrad considerably more stable in extracting complex spatio-temporal micro-expression features without becoming trapped in local minima. The highest performance achieved by AdaGrad was observed in the Agreeableness dimension, reaching an accuracy of 90.14%. The main contribution of this study is the development of an efficient (low-latency), precise spatio-temporal personality prediction model that avoids local minima, making it suitable for implementation in real-world automated psychological detection systems.

Implementasi Teknik SMOTE Menggunakan Random Forest dan XGBoost pada Klasifikasi Tingkat Kualitas Udara

2026-09-09T12:13:55+07:00

Air pollution is a global environmental issue with significant impacts on public health and environmental sustainability. The problem is compounded by the generally imbalanced class distribution in air quality data; consequently, the highest-risk category "Hazardous" often constitutes the minority class, making it the most difficult to detect using conventional classification models. This study implements a machine learning approach using Random Forest and XGBoost algorithms combined with the Synthetic Minority Oversampling Technique (SMOTE) to classify air quality levels. The dataset comprises 5,000 samples featuring nine environmental and demographic variables. Air quality is categorized into four classes Good, Moderate, Poor, and Hazardous with an initial imbalanced distribution (Good: 40%, Moderate: 30%, Poor: 20%, Hazardous: 10%). SMOTE was applied exclusively to the training data to balance the class distribution. Results indicate that XGBoost combined with SMOTE achieved the best performance, yielding an accuracy of 0.952, an F1-Score of 0.952, and an average cross-validation score of 0.969. This represents a 0.022 improvement over a previous study that utilized a Decision Tree model without SMOTE (achieving 0.930 accuracy). For the "Hazardous" minority class, the XGBoost-SMOTE combination improved recall from 0.80 (Random Forest without SMOTE) to 0.87, while also achieving more balanced precision and F1-Score values (0.87). CO levels and proximity to industrial areas emerged as the dominant features, contrasting with PM2.5 in the earlier study. These findings confirm that combining ensemble methods with SMOTE effectively addresses class imbalance. Evaluation was conducted using precision, recall, and F1-Score metrics, alongside 5-fold stratified cross-validation to ensure model stability across all classes including the "Hazardous" minority class, which is most critical for public health. These results suggest that combining ensemble algorithms with data balancing techniques can serve as a practical reference for developing machine learning-based air quality monitoring systems.

Optimasi Hyperparameter Grid Search untuk Multi-Algoritma Klasifikasi Tingkat Hidrasi pada Daily Water Intake

2026-09-09T12:25:04+07:00

Water is a vital element for sustaining human physiological functions, ranging from metabolic processes to cognitive and physical performance. Insufficient daily water intake can trigger various health problems that significantly reduce individual well being. Machine learning provides a data driven approach to classify hydration levels more objectively based on individual and environmental attributes. This study evaluates and compares the performance of three classification algorithms Random Forest, Naive Bayes, and K-Nearest Neighbor in predicting hydration status (Good or Poor) using the Daily Water Intake public dataset comprising 30,000 samples. Class imbalance in the dataset (79.7% Good, 20.3% Poor) was addressed through SMOTE, while Stratified K-Fold Cross Validation with K=10 was employed for model evaluation. Hyperparameter optimization was performed using Grid Search, and model performance was assessed through accuracy, precision, recall, F1-Score, and AUC-ROC metrics. Results demonstrate that Random Forest achieved the highest accuracy of 99.50% and AUC-ROC of 0.9999 after optimization thanks to its ensemble mechanism, which captures non-linear relationships among features, outperforming Naive Bayes, which is constrained by its feature-independence assumption. KNN showed the most notable improvement post optimization with an accuracy gain of 0.62% (from 97.62% to 98.24%), while Naive Bayes remained unchanged at 83.53% as its optimal parameter matched the default value. These findings offer evidence based guidance for algorithm selection in hydration classification systems.

Comparative Evaluation of Machine Learning Models with Explainability and Fairness for Diabetes Risk Prediction

2026-09-20T12:44:12+07:00

Diabetes mellitus remains a major global health challenge, making early risk prediction essential for timely intervention and prevention. This study proposes a trustworthy machine learning framework for diabetes risk prediction by integrating predictive performance, explainability, and fairness evaluation. Six classification algorithms Logistic Regression, Decision Tree, Random Forest, XGBoost, LightGBM, and CatBoost were comparatively evaluated, with LightGBM selected as the best baseline model. Hyperparameter optimization was subsequently performed using Optuna with macro F1-score as the optimization objective to better address the imbalanced multiclass nature of the dataset. Model performance was assessed using Accuracy, Balanced Accuracy, Precision, Recall, F1-score, and Receiver Operating Characteristic–Area Under the Curve (ROC-AUC). Model explainability was analyzed using SHapley Additive exPlanations (SHAP), while fairness was evaluated using Fairlearn based on gender and race through Demographic Parity Difference and Equalized Odds Difference.Experimental results show that hyperparameter optimization increased Balanced Accuracy from 0.3755 to 0.4891 and macro F1-score from 0.3797 to 0.4333, indicating improved recognition of minority classes. Although overall Accuracy decreased from 0.8357 to 0.7023, this trade-off reflects a more balanced classification across diabetes categories, which is preferable for imbalanced clinical datasets where identifying minority cases is essential for early risk detection. SHAP analysis identified Body Mass Index (BMI), Age, Physical Health Days, Mental Health Days, and Income Level as the most influential predictors. Fairness evaluation further demonstrated low demographic disparities across gender and race. These findings demonstrate that integrating predictive performance, explainability, and fairness enables the development of a more transparent, equitable, and clinically reliable diabetes risk prediction framework.

Klasifikasi Status Gizi Bayi sebagai Upaya Deteksi Dini Masalah Gizi Menggunakan Algoritma Naive Bayes

2026-09-09T13:05:14+07:00

The classification of infant nutritional status remains a significant challenge in the utilization of health data, particularly in developing predictive models that are accurate and reliable for supporting decision-making at the primary healthcare level. This study aims to implement the Naïve Bayes algorithm to classify infant nutritional status based on anthropometric indicators, namely weight-for-age (W/A) and height-for-age (H/A). The data used in this study were obtained from Posyandu activity reports at Citra Medika Public Health Center covering the period from 2023 to 2025, with a total of 10,654 infant records. The class distribution in the dataset includes normal, well-nourished, undernourished, severely undernourished, overnourished, and at risk of overnutrition categories. Since the class proportions are relatively balanced, no oversampling technique was applied. The research process involved several stages, including data cleaning, category normalization, class label transformation, and dataset splitting using stratified sampling with a composition of 80% training data and 20% testing data. The evaluation results indicate that the Naïve Bayes model achieved an accuracy of 82%. The precision values were 0.92 for the normal class, 0.39 for the malnutrition class, and 0.14 for the overweight class. The recall values were 0.89 for the normal class, 0.63 for the malnutrition class, and 0.08 for the overweight class. Meanwhile, the F1-scores were 0.83 for the normal class, 0.48 for the malnutrition class, and 0.11 for the overweight class. These findings suggest that the model demonstrates fairly good performance in classifying infant nutritional status based on anthropometric data. Therefore, the Naïve Bayes algorithm can be considered effective for classifying infant nutritional conditions using W/A and H/A indicators. This study is expected to contribute to the development of data-driven systems to enhance analytical quality and support decision-making in primary healthcare services.

Analisis Disparitas Data Kebencanaan Indonesia sebagai Dasar Mitigasi Berbasis K-Means

2026-09-09T13:18:33+07:00

Indonesia experiences a high frequency of disasters; however, differences in characteristics between national and global datasets create challenges for consistent disaster analysis. Previous studies generally rely on a single data source, limiting the ability to capture disparities in data representation comprehensively. This study aims to analyze disparities in disaster data representation and identify regional disaster patterns using a data mining approach. A Disaster Severity Index was constructed using Min-Max normalization, and the K-Means algorithm was applied to cluster regions based on disaster index and event frequency. The optimal number of clusters was determined using the Elbow Method and validated using the Silhouette Score, while a global dataset was used for comparison. The results indicate that the optimal model is achieved at K=4 with a Silhouette Score of 0.610883, indicating good cluster separation. Most regions exhibit moderate disaster characteristics, while a small number show extreme patterns with disproportionate relationships between frequency and severity. Differences between national and global datasets suggest variations in reporting mechanisms and data coverage. These findings demonstrate that relying on a single data source may lead to biased interpretations, highlighting the importance of multi-source data integration to improve the accuracy of disaster analysis.

Analisis Perbandingan Random Forest dan XGBoost dalam Prediksi Pembatalan Pesanan Shopee Menggunakan Interpretasi SHAP

2026-09-09T13:33:18+07:00

Order cancellation is a significant issue for e-commerce platforms because it can result in revenue loss, increased operational costs, and reduced transaction management efficiency. This study aims to compare the performance of Random Forest and Extreme Gradient Boosting (XGBoost) in predicting customer order cancellations on Shopee and to interpret the factors contributing to the prediction outcomes using an Explainable Artificial Intelligence approach based on SHapley Additive exPlanations (SHAP). The study employed the Shopee Consumer Behaviour Cancellation Order Analysis dataset obtained from Kaggle. The research process consisted of data preprocessing, dataset partitioning using the Stratified Train-Test Split method with an 80:20 ratio, implementation of both algorithms, model evaluation using Accuracy, Precision, Recall, F1-Score, and Receiver Operating Characteristic–Area Under Curve (ROC-AUC), followed by model interpretation using SHAP. The evaluation results indicate that Random Forest achieved better performance across most evaluation metrics, while XGBoost obtained a slightly higher ROC-AUC value. SHAP analysis identified Total Discount, Buyer-Paid Shipping Cost, Estimated Shipping Cost, and Estimated Shipping Fee Deduction as the variables contributing most to order cancellation predictions.These findings indicate that combining predictive algorithms with an Explainable Artificial Intelligence approach not only supports the classification of order cancellations but also provides insights into the factors influencing prediction outcomes, which can serve as a consideration for reducing order cancellations on e-commerce platforms.

Comparative IndoBERT, Logistic Regression, SVM, Random Forest, and Naïve Bayes for TikTok Deepfake Misinformation

2026-09-09T18:45:18+07:00

Generative AI has made deepfake content, especially face-swap videos and cloned voices, easier to produce and spread, and short-video platforms such as TikTok have become a major channel for this material. Most existing work still focuses on detecting the manipulated audio or video itself, leaving audience reactions in the comment section relatively unexplored. The problem this study addresses is that, without reading these comments, moderation systems have no way to know whether audiences are being misled by AI-generated content; the objective is to determine whether a fine-tuned IndoBERT model can flag such misinformation-leaning comments more reliably than conventional machine-learning classifiers. This paper takes a comparative approach to that gap, testing whether a fine-tuned IndoBERT model can flag potential misinformation in Indonesian-language TikTok comments more reliably than four TF-IDF-based classifiers: Logistic Regression, Random Forest, Multinomial Naive Bayes, and Linear SVM. Starting from 1,345 comments scraped across 15 TikTok videos, a multi-stage cleaning process left 1,071 usable comments, manually sorted into three labels: Misinformation, Skeptical, and Other. IndoBERT was fine-tuned with a weighted cross-entropy loss to offset class imbalance and evaluated through a train-validation-test split and Stratified 5-Fold Cross Validation. The fine-tuned model reached 82.41% accuracy and 81.58% F1-Macro, beating every machine-learning baseline by at least 13.84 points in F1-Macro, with five-fold results holding steady at an average F1-Macro of 80.17%. These results suggest IndoBERT is a stronger option than conventional machine learning for flagging misinformation-leaning comments on AI-generated content, offering a practical foundation for automated moderation on social platforms.

Perbandingan Random Forest dan SVM dalam Analisis Sentimen Opini Masyarakat Terhadap Kenaikan Harga Pertamax

2026-09-09T18:55:12+07:00

The increase in Pertamax prices triggered various responses from the public, many of which were expressed through social media X. Due to the large number of opinions that emerged, sentiment analysis proved to be an appropriate way to automatically recognize public opinion trends. This study was conducted with the aim of determining how effective the "Random Forest" and "Support Vector Machine" (SVM) algorithms are in the process of grouping public sentiment regarding the increase in Pertamax prices. The research data was obtained by collecting information from platform X using keywords that match the topic of Pertamax. After filtering and removing duplicates, 5,006 records were obtained which were used as the dataset in this study. The steps of this study include data pre-processing, sentiment classification process, feature extraction using the Term Frequency Inverse Document Frequency (TF-IDF) technique, Using the SMOTE method to balance the number of classes and divide the dataset into categories, then classifying with two algorithms, namely Random Forest and Support Vector Machine (SVM). Confusion matrix is used to assess model performance with various metrics such as accuracy, precision, recall, and F1 score. This study shows that the data is dominated by negative sentiment, amounting to 51.64%, then dominated by positive sentiment at 35.60%, and neutral sentiment reached 12.76%. Before using SMOTE, Random Forest had an accuracy of 73% while SVM reached 75%. After using SMOTE, the accuracy of the Random Forest model increased to 85% and the accuracy of the SVM increased to 89%. This research contributes to improving the quality of sentiment analysis and serves as a reference for future studies; furthermore, it provides stakeholders with an understanding of public opinion trends regarding the price hike of Pertamax, thereby assisting them in decision-making.

An Interval-Informed Hybrid CNN-LSTM for Ten-Category ECG Beat and Event Classification

2026-09-09T19:11:01+07:00

Automated electrocardiogram (ECG) classification remains challenging because waveform-based deep-learning models capture rich morphology, whereas clinically relevant conduction and timing intervals are not always represented explicitly. This study investigates whether a compact interval-informed representation can retain useful discriminative information for ten ECG beat and event categories. MIT-BIH Arrhythmia Database recordings were filtered and divided into seven-second segments. RR, PR, and QT intervals and QRS width were summarized using minimum, maximum, mean, median, skewness, and kurtosis, yielding 24 features. To prevent information leakage, the original samples were partitioned before oversampling; feature scaling and SMOTE were applied only to training data, while validation and held-out test data retained their original distributions. Leakage-controlled five-fold cross-validation yielded 92.15% ± 0.43% accuracy and 86.78% ± 0.68% macro-F1. On the untouched 826-sample test set, the model achieved 92.62% accuracy, 87.49% macro-F1, and 92.68% weighted F1. The most frequent bidirectional errors occurred between Normal and Premature Ventricular Contraction beats. These results support the use of interval-statistical features as a compact representation for multicategory ECG analysis, while showing that rare-class performance and fiducial reliability remain limiting factors for broader clinical use.

Komparasi Random Forest dan XGBoost Berbasis SMOTE dalam Penanganan Ketidakseimbangan Data untuk Prediksi Risiko Stroke

2026-09-22T12:42:04+07:00

Class imbalance is one of the main obstacles in building machine learning models for stroke prediction, as this condition tends to make models learn the characteristics of the majority class more dominantly than those of the minority class. This study aims to examine the comparison and evaluate the performance between the Random Forest and XGBoost algorithms in predicting stroke risk, by applying SMOTE as a data balancing technique. The study uses the Kaggle Stroke Prediction Dataset, consisting of 5,110 patient records. The research stages include data cleaning, categorical feature representation using TF-IDF, numerical feature normalization through Standard Scaling, data splitting using the Stratified Train-Test Split technique, application of SMOTE on the training data, and model development using the Random Forest and XGBoost algorithms. Model performance was evaluated using several metrics, namely Accuracy, Precision, Recall, F1-Score, ROC-AUC, Confusion Matrix, and Feature Importance. The test results show that XGBoost achieved a Recall of 0.1600, an F1-Score of 0.1702, and an ROC-AUC of 0.7877—all higher than those of Random Forest, which achieved a Recall of 0.0800, an F1-Score of 0.1111, and an ROC-AUC of 0.7068. Although these results represent an improvement over Random Forest, the still relatively low Recall and F1-Score values indicate that the model’s ability to detect stroke cases is not yet optimal and therefore requires further development. Furthermore, SMOTE was applied to balance the class distribution in the training data so that the model’s learning process would give greater attention to the minority class. This study contributes by implementing a consistent preprocessing pipeline, including data cleaning, Standard Scaling, TF-IDF-based categorical feature representation, Stratified Train-Test Split, and SMOTE on the training data, enabling a more objective comparison of Random Forest and XGBoost through the use of comprehensive evaluation metrics on an imbalanced stroke prediction dataset.

Analisis Komparasi Kinerja SVM, Random Forest, dan XGBoost untuk Diagnosis Penyakit Batu Empedu

2026-09-22T13:10:13+07:00

Gallstone disease is a hepatobiliary disorder that requires accurate diagnosis to support timely management and reduce the risk of complications. The availability of clinical and laboratory data provides an opportunity to apply machine learning as a supporting approach for disease classification. This study aims to compare the performance of Support Vector Machine (SVM), Random Forest, and XGBoost algorithms for gallstone disease classification using the Gallstone Dataset from the UCI Machine Learning Repository. The contribution of this study is to provide a comparative evaluation of SVM, Random Forest, and XGBoost on the Gallstone Dataset by consistently applying feature selection, feature standardization, class imbalance handling, and cross-validation to obtain a more objective comparison of the three algorithms. The dataset consists of 319 samples; therefore, model evaluation needs to consider the limited sample size and the potential risk of overfitting, which may affect model generalization. The research stages include feature selection using the ANOVA F-Value, handling class imbalance using the Synthetic Minority Over-sampling Technique (SMOTE), feature standardization, model development, and validation using Stratified 5-Fold Cross Validation. Evaluation was conducted using accuracy, precision, recall, F1-score, and Area Under the Curve (AUC). The results show that XGBoost achieved the best performance, with an accuracy of 79.00%, precision of 80.95%, recall of 75.32%, F1-score of 78.03%, and AUC of 85.04%. Random Forest achieved an accuracy of 78.06%, while SVM achieved 74.92%. Feature importance analysis indicated that C-Reactive Protein (CRP), Vitamin D, Hyperlipidemia, Bone Mass, and Aspartate Aminotransferase (AST) were among the attributes that contributed significantly to the classification. The results indicate that XGBoost outperformed SVM and Random Forest on the dataset used and has potential as a supporting approach for developing a gallstone disease classification system; however, evaluation using larger and more diverse datasets is required to confirm its generalization capability.

Analisis Komparasi Random Forest dan XGBoost untuk Prediksi Produksi Kelapa Sawit Berbasis SHAP

2026-09-22T13:31:27+07:00

Palm oil production is influenced by various factors, such as plant age, land area, the number of productive trees, and fertilizer usage. Production fluctuations make it difficult for companies to formulate optimal plans for production, distribution, and fertilization strategies. Unlike previous studies, this research utilizes a dataset obtained directly from the company to compare the performance of Random Forest and XGBoost algorithms in predicting palm oil production, while also analyzing which variables most significantly influence prediction results using SHAP (SHapley Additive exPlanations). The study employs historical palm oil production data from PT XYZ covering the period from January 2020 to December 2025 comprising 588 data points using a multivariate time series approach. The research stages include data preprocessing, the creation of sliding windows with sequence lengths of 3 and 6 months, model training, hyperparameter tuning, evaluation using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and the Coefficient of Determination (R²), as well as model interpretation using SHAP. The results indicate that both models achieved their best performance with a sequence length of 3 months. The XGBoost model outperformed Random Forest, achieving an MAE of 553,293, an RMSE of 1,496,990, and an R² of 0.877, whereas Random Forest recorded an MAE of 1,283,547, an RMSE of 1,792,539, and an R² of 0.824. SHAP analysis revealed that month was the most influential factor in predicting palm oil production, followed by lag_1, lag_3, and land area. This research contributes a comparison of Random Forest and XGBoost on primary palm oil production data obtained directly from the company, as well as a SHAP-based interpretation that can serve as a practical reference for PT XYZ in production planning and corporate strategy. These findings demonstrate that the combination of XGBoost and SHAP can generate accurate predictions while providing insights into the factors influencing production, thereby supporting decision-making in palm oil production management.

Komparasi CNN dan SVM untuk Klasifikasi Penyakit Daun Jagung Berbasis CLAHE dan Explainable AI

2026-09-22T13:50:20+07:00

Corn productivity losses from leaf diseases remain a serious problem, while manual observation is subjective and accurate models are hard to trust when their decisions cannot be traced. The contribution of this research lies in controlled testing that isolates the contribution of CLAHE from architectural influences, establishing preprocessing parameters based on empirical testing, and directly comparing both methods under identical data and treatment conditions. This study compares Convolutional Neural Network (CNN) and Support Vector Machine (SVM) for corn leaf disease classification under equivalent conditions, using CLAHE preprocessing and LIME. Testing used 4,188 images in four classes. The CNN based on EfficientNet-B0 reached 94.98%, outperforming the SVM based on GLCM and color histogram features at 92.46%. This 2.52-point margin is narrower than assumed, and the additional SVM errors concentrate in one class. A hybrid combining EfficientNet-B0 features with an SVM classifier reached 95.33%, indicating the CNN advantage stems from its feature representation rather than its decision layer. The SVM also predicted one hundred sixty times faster without a graphics accelerator. An ablation test shows the CLAHE benefit is method-dependent: the SVM gained 1.19 points, the CNN did not. LIME shows most influential regions fall on lesion areas, though some appear in the background. These findings keep the SVM relevant when computational resources are limited.

A Comparative Evaluation of Random Forest, SVM, KNN, Logistic Regression, and Gradient Boosting for Stress Prediction

2026-09-22T14:05:07+07:00

Student stress has become a critical concern in contemporary education, necessitating early detection systems to enable timely intervention. This study presents a comprehensive comparative evaluation of five machine learning algorithms Logistic Regression, K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Random Forest, and Gradient Boosting for classifying student stress levels into three categories (low, moderate, high). Using a dataset of 14,003 students with 14 academic, demographic, and psychosocial features, the research addresses the significant class imbalance (high stress: 50.7%, moderate: 29.1%, low: 20.2%) through systematic multi-metric assessment. Random Forest with 300 trees and class_weight='balanced' demonstrated superior performance across all metrics: accuracy (93.47%), precision (93.97%), recall (93.47%), F1-score (93.43%), and ROC-AUC (0.992). Feature importance analysis identified AssignmentCompletion, Attendance, and ExamScore as the most influential predictors. The contributions of this research include: (1) a comprehensive comparative evaluation of five foundational ML algorithms on a large-scale imbalanced educational dataset; (2) the explicit demonstration of Random Forest's effectiveness in handling class imbalance through built-in class weighting; (3) the identification of key academic predictors of student stress through feature importance analysis; and (4) the provision of a scalable evidence-based framework for early stress detection with practical implementation guidance for educational institutions.

Optimizing CNN Architecture Selection For Small-Scale Imbalanced Chili Leaf Disease Classification

2026-09-22T14:34:26+07:00

Chili leaf disease reduces yield and farmer income in Indonesia, yet automated diagnosis is typically evaluated using a single architecture or artificially balanced datasets that do not reflect field conditions. This study compared eight CNN architectures (ResNet50, ResNet101, EfficientNetB0, EfficientNetB3, MobileNetV2, MobileNetV3Large, VGG16, and DenseNet121) for six-class chili leaf disease classification on an original dataset of 1,847 images, using on-the-fly augmentation only on the training subset. Each architecture was trained with a frozen feature-extraction transfer-learning protocol and class weighting across three random seeds (24 runs total) to assess robustness. MobileNetV3Large achieved the highest mean accuracy (98.80%) and was selected as the final model, reaching 98.92% accuracy, a macro-average F1-score of 98.09%, and a weighted-average F1-score of 98.90%. A Friedman test confirmed that performance differences among the eight architectures were statistically significant ( = 19.432, = 0.0069), with a Kendall's W of 0.9253 indicating strong ranking agreement across seeds. Grad-CAM visualization showed that the final model's predictions were consistently driven by symptomatic leaf regions rather than background artifacts. These results indicate that a lightweight, lower-parameter architecture can match or exceed deeper models under frozen feature extraction, offering a favorable performance-to-size trade-off for agricultural extension services. This study contributes a statistically validated benchmarking protocol enabling fair comparison across CNN design philosophies on a single, non-synthetically-balanced dataset, a setting most prior chili leaf disease studies have not provided.

Implementasi Algoritma BERTopic Berbasis IndoBERT untuk Pemodelan dan Analisis Trending Topik Berita Online pada Dashboard Website

2026-09-22T14:55:33+07:00

The growing volume of online news makes it difficult for the Sumatera Ekspres editorial team to manually track popular topics in a timely manner. This study implements the BERTopic algorithm based on IndoBERT to automatically model trending topics of online news, following the CRISP-DM stages of business understanding, data understanding, data preparation, modeling, evaluation, and deployment. Data were collected through web scraping from eight national news portals and YouTube, then processed through case folding, cleaning, tokenization, and stopword removal before being modeled using IndoBERT as the embedding, UMAP for dimensionality reduction, HDBSCAN for clustering, and c-TF-IDF for keyword extraction, complemented by a trend_score metric to identify topics that are genuinely gaining momentum. The main contribution of this study lies in integrating more heterogeneous multi-source data than prior work and introducing the trend_score metric as a more accurate indicator of momentum, realized as a ready-to-use dashboard for the editorial team. Testing across four data periods (1, 7, 14, and 30 days) shows that the model consistently produces topics of good coherence and diversity quality. The results were implemented into a website-based dashboard built with FastAPI and SQLite that presents trending topic visualizations, word cloud, and news headline recommendations, which has been tested and confirmed to function as required, making it suitable as a decision-support tool for the editorial team to monitor popular issues quickly, efficiently, and in a data-driven manner.

Implementation of Machine Learning for Clustering Lethal Diseases Using 21st-Century Global Mortality Data

2026-09-22T15:41:41+07:00

Entering the 21st century, the increasing global prevalence of lethal diseases demands a data-driven surveillance system capable of accurately classifying risk levels. Conventional epidemiological monitoring is often insufficient for identifying patterns in the massive volume and complexity of mortality data. This study conducts a comparative analysis between K-Means and Fuzzy C-Means (FCM) algorithms in clustering fatal diseases based on aggregated global death counts using the Global Health Statistics dataset containing 100,000 records. Optimal cluster numbers were determined independently: the Elbow Method yielded K=4 for K-Means, while the Fuzzy Performance Index suggested K=7 for FCM. Performance evaluation using the Silhouette Score under natural configurations showed FCM slightly outperforming K-Means (0.5574 vs. 0.5411), whereas K-Means achieved the highest score under a controlled K=7 setting (0.6014). This research provides a significant contribution by establishing the first direct comparative framework for global multi-disease mortality clustering. The findings indicate that cluster cardinality exerts a stronger influence on clustering quality than the choice of algorithm itself. For policymakers, the results offer a choice between a simple four-tier K-Means risk stratification and the finer granularity of a seven-tier FCM model, providing a scientifically robust, data-driven foundation for evidence-based health intervention and resource allocation strategies.

Komparasi Algoritma Support Vector Machine, Random Forest, dan XGBoost: Prediksi Penyakit Huntington Berdasarkan Data Ekspresi Gen

2026-09-22T15:57:56+07:00

Huntington's disease is a progressive neurodegenerative disorder caused by mutations in the Huntingtin (HTT) gene, and early diagnosis remains an ongoing challenge. Leveraging gene expression data through machine learning holds potential for improving the prediction accuracy of this disease; however, the characteristics of microarray data very high feature dimensionality combined with a limited sample size (small-n large-p) together with class imbalance, increase the risk of overfitting and data leakage. This study compares the performance of Support Vector Machine (SVM), Random Forest, and Extreme Gradient Boosting (XGBoost) in predicting Huntington's disease status using the GSE135589 microarray dataset. The research pipeline includes data standardization, feature selection (SelectKBest based on the ANOVA F-test), class balancing (SMOTE), and hyperparameter tuning via GridSearchCV with Stratified 5-Fold Cross-Validation, in which the number of features (k) is also tuned as a nested hyperparameter alongside other model parameters to prevent information leakage between the training and test data. All preprocessing steps were applied exclusively to the training data (fit-on-train, transform-on-test) to ensure performance estimates free from optimistic bias. The results show that the linear-kernel SVM achieved the best performance, with an accuracy of 88.89%, precision of 89.66%, recall of 96.30%, F1-score of 92.86%, and ROC-AUC of 90.95%, outperforming Random Forest and XGBoost on the independent test set. These findings indicate that regularized linear approaches outperform tree-based models under conditions where feature dimensionality far exceeds sample size. The main novelty of this study lies in tuning the number of features (k) as a nested hyperparameter within the cross-validation scheme rather than as a separate step prior to model training, as is common in most comparable studies, thereby producing a stricter, leakage-free generalization estimate, while also serving as the first comprehensive benchmark comparing the three algorithms specifically within the domain of Huntington's disease gene expression data.

Perbandingan Embedding IndoBERT dan FastText dalam Deteksi Cyberbullying Berbahasa Indonesia

2026-09-22T16:12:21+07:00

The escalation of cyberbullying on social media necessitates automated content moderation systems capable of deciphering the characteristics of non-standard digital language. This study compares the effectiveness of contextual word embeddings (IndoBERT) and static subword embeddings (FastText) in multi-class text classification (Safe, Mild, Moderate, Severe, and Very Dangerous) for Indonesian-language social media comments. The research focuses on analyzing embedding architectures: IndoBERT is integrated with a Dense Layer, while FastText is evaluated using Logistic Regression and Random Forest (15 trees). A dataset of 16,643 samples was extracted from social media platforms and processed through text cleaning and dictionary-based slang normalization stages. Test results indicate that the bidirectional contextual architecture of IndoBERT achieved the highest performance, with an accuracy of 82.15% and a Macro F1-Score of 74.97%. Regarding FastText representations, the linear Logistic Regression approach (78.70% accuracy) proved more stable than the Random Forest model (65.50% accuracy). These findings demonstrate that contextual embeddings are more adaptive than static embeddings at distinguishing sarcasm and mutated lexicons in digital text. As a research contribution, this study not only offers a precise five-level severity classification architecture for informal Indonesian but also provides a prototype automated moderation system (CyberGuard AI) ready for implementation in real-time risk mitigation.

Comparison of Global Horizontal Irradiance Prediction Using LSTM and M-LSTM Across Seasonal Conditions

2026-09-22T16:36:18+07:00

Short-term prediction of Global Horizontal Irradiance (GHI) is challenging because solar radiation exhibits dynamic temporal patterns and rapidly responds to changing atmospheric conditions. This study compares univariate Long Short-Term Memory (LSTM) and multivariate LSTM models using four and six predictor variables (M-LSTM-4 and M-LSTM-6) for 30-minute-ahead GHI forecasting using Automatic Solar Radiation Station data from Malang, East Java. One-minute observations were aggregated into 30-minute intervals and evaluated using sliding windows of 12, 24, and 36 timesteps. Model performance was assessed using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the coefficient of determination (R²) for full-year, wet-season, and dry-season datasets. For the full-year dataset, M-LSTM-6 with a 12-timestep window achieved the best performance, with an MAE of 37.700 W/m², an RMSE of 79.340 W/m², and an R² of 0.9283. In contrast, M-LSTM-4 provided the best performance for both seasonal datasets using a 36-timestep window, achieving an MAE of 36.551 W/m² and an R² of 0.9378 during the wet season, and an MAE of 39.181 W/m² and an R² of 0.9454 during the dry season. M-LSTM-6 showed substantial performance degradation with longer historical windows, particularly during the dry season. This result suggests that combining a larger number of predictors with longer historical sequences increases the amount of input information processed by the model, which may be less effective when the additional historical information does not provide proportional predictive benefit under rapidly varying irradiance conditions. The results demonstrate that the effectiveness of multivariate information depends on predictor relevance, historical window length, and seasonal conditions. The main contribution of this study is the identification of different optimal combinations of predictor complexity and historical window length between full-year and seasonal GHI forecasting. Overall, the findings indicate that input-variable complexity and window length should be jointly optimized for local short-term GHI forecasting.

Multi-Run Statistical Evaluation of Retinex-Net, SNR-Aware, and RetinexFormer for Nighttime Vehicle Re-Identification

2026-09-22T16:50:40+07:00

Nighttime vehicle re-identification (Re-ID) is challenging because low illumination, sensor noise, and motion blur can obscure identity-related visual features in surveillance images. Comparative studies of low-light enhancement (LLE) as a preprocessing step for Re-ID often rely on a single training run, making it difficult to distinguish genuine method effects from training stochasticity. This study evaluates Retinex-Net, SNR-Aware, and RetinexFormer as LLE preprocessing methods for TransReID with a ViT-Base/16 backbone on a stratified nighttime subset of the DN-348 dataset. The subset contains 8,255 images from 150 training identities and 120 shared gallery/query identities, with nighttime degradation simulated using gamma darkening, Poisson-Gaussian noise, and Gaussian blur. Each scenario was trained using three independent random seeds, with performance summarized across runs using mean Average Precision (mAP) and Cumulative Matching Characteristic (CMC) at Rank-1, Rank-5, and Rank-10. Additional statistical validation was performed on a representative run using bootstrap confidence intervals and paired Wilcoxon signed-rank tests with Holm-Bonferroni correction. Image quality was additionally assessed using PSNR, SSIM, and LPIPS, while SIE and JPM were examined through a factorial ablation. RetinexFormer achieved the highest mean mAP of 48.90 ± 0.11%, followed by SNR-Aware at 48.69 ± 0.21% and the baseline at 48.51 ± 0.33%. However, neither RetinexFormer nor SNR-Aware showed a statistically significant improvement over the baseline, whereas Retinex-Net performed significantly worse. RetinexFormer also produced the best image-quality metrics, indicating that stronger visual restoration does not necessarily yield better downstream Re-ID. These findings demonstrate the importance of multi-run and statistically validated evaluation when assessing LLE methods for nighttime vehicle Re-ID. Given that none of the evaluated LLE methods yielded a statistically significant mAP improvement over the no-enhancement baseline, these results indicate that LLE should not be adopted by default as a preprocessing step for nighttime vehicle Re-ID and should only be deployed after scenario-specific multi-run statistical validation.

Analisis Sentimen Komentar TikTok pada Konten Sound Horeg Menggunakan XLM-RoBERTa, Kamus Bahasa Jawa dan CatBoost

2026-09-22T17:04:10+07:00

Comments on TikTok can provide valuable data for analyzing public sentiment toward social and cultural phenomena. However, informal language, multilingual expressions, code-mixing, and limited linguistic resources for regional languages present challenges for sentiment classification. Sound Horeg was selected as the research subject because this social and cultural phenomenon generates diverse public responses on social media, including support and opposition, making it a relevant context for examining public perceptions through sentiment analysis. This study aims to develop a hybrid sentiment analysis pipeline for classifying TikTok comments related to Sound Horeg by integrating XLM-RoBERTa, a Javanese sentiment lexicon, and CatBoost. The methodological novelty of this study lies in integrating the multilingual semantic representations of XLM-RoBERTa with a Javanese lexical resource in the context of Sound Horeg sentiment analysis, while evaluating the contribution of each component through an ablation study. The dataset was collected from five Sound Horeg videos and consisted of 7,330 comments, which were processed through data cleaning, preprocessing, Lexicon-Based sentiment labeling, XLM-RoBERTa embedding extraction, and Random Undersampling. Lexicon-based labeling produced 960 labeled comments, consisting of 559 positive and 401 negative comments, while Random Undersampling produced 802 balanced samples for the main model. The 768-dimensional XLM-RoBERTa embeddings were used as input features for CatBoost. The main model achieved an accuracy of 89.44%, an F1-Macro score of 89.44%, and an ROC-AUC of 90.25%. The ablation study showed that the Lexicon-only configuration achieved an accuracy of 94.79%, followed by XLM-R only at 93.23% and XLM-R + Lexicon at 91.15%. These results indicate that the simpler configurations, particularly Lexicon-only and XLM-R only, achieved higher performance than the main pipeline, suggesting that the hybrid pipeline cannot yet be considered the most optimal configuration for operational implementation. These findings indicate that XLM-RoBERTa can provide effective semantic representations without using skor leksikals as the main features, while the contribution of Javanese vocabulary remains limited because it was detected in only 1.6% of Group A comments and 0.8% of Group B comments. This finding implies that the selection of components in a sentiment analysis system should consider the balance between predictive performance, processing complexity, and the linguistic characteristics of the analyzed data.

Sistem Rekomendasi Film dengan Collaborative Filtering Berbasis SVD, KNN dan Explainable AI pada Dataset MovieLens

2026-09-26T01:46:08+07:00

The rapid growth of streaming platforms triggers information overload, necessitating effective recommendation systems. However, conventional Collaborative Filtering faces extreme data sparsity, reaching 98.47% in the MovieLens dataset, and operates as a black-box lacking transparent predictive rationalization. This study aims to develop a robust movie recommendation system that accurately predicts sparse data while offering novelty through predictive transparency using Explainable AI (XAI). The methodology applies Top-N Filtering to densify the matrix, followed by optimizing Singular Value Decomposition (SVD) and K-Nearest Neighbors (KNN) algorithms via GridSearchCV. Results demonstrate that matrix densification to 81.20% successfully accelerated algorithmic performance. SVD significantly outperformed KNN, not only in error reduction (RMSE 0.182090) but also in real-world relevance metrics, achieving a Precision@10 of 0.917, Recall@10 of 0.806, and F1-Score@10 of 0.858. The main contribution of this research lies in demystifying the SVD black-box using XAI. Individual-level evaluation using Local Interpretable Model-Agnostic Explanations (LIME) quantitatively proved that factorization model recommendations are driven more by universal trend sentiments (global mean contribution +0.729) than personal history matching. Furthermore, spatial exploration through Principal Component Analysis (PCA) highlighted the complexity of SVD's 100-dimensional latent space that cannot be linearly reduced, confirming the mathematical need for caution in 2D visual interpretations. In conclusion, the synergy of optimized SVD and the XAI framework produces a robust recommendation system that breaks black-box limitations, directly impacting the enhancement of transparency and user trust in digital platforms.

Perbandingan Holt, Brown DES, dan Theta untuk Peramalan Minyak dan Gas dengan Data Terbatas

2026-09-26T02:08:41+07:00

The decline and fluctuations in natural resource production in East Kalimantan, particularly oil, natural gas, and coal, create uncertainty regarding future production conditions, necessitating forecasting methods suited to the characteristics of time-series data. However, few studies have specifically compared Holt’s Linear Trend Method, Brown’s Double Exponential Smoothing (Brown’s DES), and the Theta Model using East Kalimantan natural resource data with a relatively limited number of annual observations. This study aims to compare the performance of these three methods in forecasting oil lifting, natural gas lifting, and coal production. The novelty of this study lies in its comparative evaluation of the three methods across three natural resource indicators with different time-series patterns, as well as its assessment of their ability to use limited historical data to generate five-year projections. The Theta Model was included as a comparative method because it can represent time-series patterns through trend decomposition and extrapolation, making it relevant for evaluating data that exhibit long-term changes. The study uses annual data for the 2016–2024 period obtained from the Statistics Indonesia Office of East Kalimantan Province and the East Kalimantan One Data portal. The use of annual data with a relatively limited number of observations was considered when selecting the methods because trend- and smoothing-based approaches such as Holt, Brown’s DES, and Theta can be applied to time series with limited sample sizes without requiring the estimation of a large number of parameters. The research stages included data collection, model development using the three forecasting methods, accuracy evaluation based on Mean Absolute Percentage Error (MAPE), and five-year forecasting. The results show that the Theta Model performed best for oil and natural gas, with MAPE values of 13.04% and 17.42%, respectively. Meanwhile, Brown’s DES produced the lowest MAPE for coal, at 37.35%. The relatively high MAPE values for coal indicate that the three methods have not fully captured the characteristics of the data's fluctuations, and no single forecasting method consistently outperforms the others across all natural resource indicators; therefore, the selection of a method must take into account the pattern characteristics and levels of variation inherent in each time series. The limited accuracy of coal forecasting highlights the need to explore alternative methods capable of capturing fluctuating patterns.

Model Hibrida PSO-LSTM dan Interpretasi SHAP untuk Peramalan Harga Emas Domestik Multivariat

2026-09-27T23:24:56+07:00

Forecasting volatile domestic gold prices requires a predictive model that is not only accurate but also transparent. Although the Long Short-Term Memory (LSTM) architecture excels at capturing temporal dependencies in time-series data, its performance is highly sensitive to hyperparameter tuning and suffers from a black-box nature. To overcome the limitations of manual optimization and lack of transparency, this study proposes a hybrid multivariate PSO-LSTM forecasting model integrated with the SHapley Additive exPlanations (SHAP) method. Built upon the CRISP-DM framework using daily historical data from 2020 to 2026, the Particle Swarm Optimization (PSO) algorithm is employed to automatically optimize the number of neurons, learning rate, and batch size. Furthermore, the SHAP method with K-Means background reduction is applied to map the contribution of input features. Experimental results demonstrate that the PSO-LSTM model outperforms the baseline LSTM by reducing Root Mean Squared Error (RMSE) by 11.70% and Mean Absolute Error (MAE) by 14.08%, while achieving a highly accurate Mean Absolute Percentage Error (MAPE) of 2.62%. Statistical testing using the Diebold-Mariano test confirms that the predictive superiority of the PSO-LSTM model is statistically significant. SHAP analysis indicates that historical memory of domestic gold prices provides the highest marginal contribution to model predictions, followed by USD/IDR exchange rate adjustments and global gold price sentiment. Overall, this study contributes an integrated financial modeling framework that not only enhances accuracy on highly volatile data but also provides rational feature interpretability as a foundation for hedging decisions by investors and policymakers.

Pendekatan Hybrid Transformer-BiLSTM Berbasis Multi-Task Learning untuk Klasifikasi dan Prioritisasi Pengaduan Pelanggan

2026-09-29T18:45:48+07:00

The rapid growth of e-commerce platforms has generated a massive volume of user reviews that can be utilized to understand customer issues. However, the informal nature of user-generated text, the diversity of complaint types, and the need to identify both complaint categories and their handling priorities make automatic complaint classification a challenging task. Therefore, this study proposes a Hybrid Transformer–BiLSTM model based on Multi-Task Learning (MTL) to simultaneously classify complaint categories and priority levels through a shared feature representation. The dataset was collected by web scraping Google Play Store reviews of the Tokopedia application, resulting in 100,000 raw reviews. After the preprocessing stage, 42,929 reviews were retained for model development. The labeling process employed a rule-based pseudo-labeling approach involving keyword identification, weighted scoring, and ambiguity resolution to generate category and priority labels. The proposed model integrates IndoBERT as the transformer encoder with BiLSTM to capture contextual representations and sequential dependencies in Indonesian text. Experimental results show that the proposed model achieved 96.17% accuracy, with a weighted F1-score of 0.9617 and a macro F1-score of 0.9353 for complaint category classification. For complaint priority classification, the model achieved 80.04% accuracy, with a weighted F1-score of 0.7437 and a macro F1-score of 0.4236. The lower performance on the priority classification task indicates that the model is still affected by class imbalance, resulting in better recognition of majority classes than minority classes. In addition, the quality of the generated labels depends on the predefined rules used in the pseudo-labeling process. Nevertheless, the proposed approach demonstrates strong potential as a foundation for developing an automated complaint classification and prioritization system for e-commerce customer service. The contribution of this study lies in the implementation of a Hybrid IndoBERT–BiLSTM architecture based on Multi-Task Learning for the simultaneous classification of customer complaint categories and priorities, as well as the provision of an Indonesian Tokopedia complaint dataset that has undergone a pseudo-labeling process as a reference for future research.

Perbandingan EfficientNetB0, MobileNetV2, dan DenseNet121 Berbasis Transfer Learning untuk Deteksi AI-Generated Images

2026-10-05T23:30:39+07:00

The rapid advancement of AI-based generative models such as GANs and Diffusion Models has produced high-quality synthetic images that are visually indistinguishable from real photographs. This poses serious challenges in digital security, media verification, and public trust in visual content. However, few studies have compared multiple CNN architectures using a two-phase transfer learning strategy on the same dataset. This study proposes a transfer learning-based approach by comparing three ImageNet-pretrained CNN architectures, namely EfficientNetB0, MobileNetV2, and DenseNet121, for detecting AI-generated images. The dataset used is the AI vs Real Images Dataset from Kaggle, consisting of two classes: AI-generated and real images. Training was conducted in two phases: the first phase with a frozen backbone to adapt pretrained weights, and the second phase involving fine-tuning the last 30 layers of the backbone. Data augmentation including rotation, shifting, zoom, shear, and brightness variation was applied to the training set, while class weighting and label smoothing were employed to address class imbalance. Evaluation was performed using accuracy, precision, recall, F1-score, and AUC-ROC metrics. Experimental results show that DenseNet121 achieved the highest performance with an accuracy of 98.00%, precision 99.10%, recall 98.21%, F1-score 98.65%, and AUC-ROC 99.72%. EfficientNetB0 ranked second with 95.33% accuracy, while MobileNetV2 recorded 94.67% accuracy with the fewest parameters, making it superior in terms of efficiency. Grad-CAM visualizations confirm that the models focus on the main objects as well as textured and color-transition regions that serve as markers of generative artifacts. This study demonstrates that a two-phase transfer learning approach with structured data augmentation is effective for AI image detection tasks.

Analisis Komparatif Metode Logistic Regression dan Support Vector Machine pada Sentimen Aplikasi GuitarTuna

2026-10-05T23:45:38+07:00

This study focuses on identifying sentiment trends among GuitarTuna app users while comparing the performance of Support Vector Machine (SVM) and Logistic Regression (LR) algorithms. A total of 20,339 raw data points were collected from the Google Play Store, the App Store, and questionnaires. Following a rigorous filtering process, which involved eliminating duplicate entries, empty text, and irrelevant responses, 13,615 valid data points were obtained. The analysis proceeded through text preprocessing, feature extraction using Term Frequency-Inverse Document Frequency (TF-IDF) weighting, and classification using both algorithms. Evaluation results indicate that SVM achieved an accuracy of 90.53%, a macro recall of 0.60, and a macro F1-score of 0.62. This performance was descriptively superior to that of LR, which recorded an accuracy of 90.30%, a macro recall of 0.56, and a macro F1-score of 0.59. However, McNemar's test, yielding a p-value of 0.6338, demonstrated that SVM's nominal advantage was not statistically significant. Furthermore, the limited sample size for the neutral class contributed to low prediction precision for both models within that category. Regarding review topics, positive sentiment was largely driven by user satisfaction with tuning precision, ease of use, and educational features. Conversely, negative sentiment was dominated by complaints concerning subscription costs, transaction issues, feature access restrictions, and connectivity problems. This research makes contributions in two key areas. From a practical standpoint, the topic extraction findings provide developers with concrete guidance for prioritizing feature and service improvements. From an academic perspective, the study offers comprehensive empirical evidence regarding the comparative evaluation of machine learning algorithms on cross-platform review datasets characterized by class imbalance.

A Climatology-Relative Framework for Evaluating Rainfall-Condition Categories Using Observed and Random Forest-Predicted Dasarian Rainfall

2026-10-11T22:56:23+07:00

Rainfall variability strongly influences rain-fed agricultural activities in monsoonal regions, particularly across seasonal and spatially varying rainfall regimes. This study develops and applies a climatology-relative framework for evaluating rainfall-condition categories derived from observed and Random Forest (RF)-predicted dasarian rainfall for Besuki Na-Oogst tobacco in ZOM 45–49, Jember, East Java. A 30-year historical rainfall dataset (1993–2022) was used to derive ZOM- and dasarian-specific P25 and P75 relative rainfall-condition thresholds. The same fixed thresholds were then applied to observed and RF-predicted rainfall throughout January–December 2024 to classify rainfall into Dry, Normal, and Wet categories. Categorical correspondence was evaluated using percentage agreement, Cohen's kappa, and a 3 × 3 confusion matrix at the ZOM × dasarian level. Across 180 ZOM × dasarian pairs, observed rainfall produced 50 Dry, 84 Normal, and 46 Wet classifications, whereas RF-predicted rainfall produced 7 Dry, 68 Normal, and 105 Wet classifications. Overall categorical agreement was 38.33% (κ = 0.071), with variation among the five ZOMs. The most pronounced discrepancy occurred in the Dry category: none of the 50 observed Dry cases were reproduced as Dry by the RF predictions, while 36 were classified as Wet. Conversely, all seven RF-predicted Dry cases corresponded to observed Wet conditions. The continuous comparison indicated that the RF predictions generally reproduced the broad seasonal rainfall pattern but tended to exhibit smoother variations and, in several periods, lower rainfall magnitudes than observed. These differences can lead to categorical reversals when predicted values fall across fixed climatological P25 and P75 boundaries. The findings indicate that RF-derived rainfall-condition patterns may provide complementary information for climatology-relative assessment, but should not be interpreted as direct substitutes for observed rainfall conditions or as standalone planting recommendations. The proposed framework complements conventional numerical rainfall-prediction evaluation by assessing the categorical representation of predicted rainfall within a common local historical climatology at the ZOM × dasarian scale.