Indonesian Journal of Data and Science

[RETRACTED] Comparison of Parameter Estimation Methods in Weibull Distribution

2025-04-17T22:24:52+00:00

The main objective of this study is to compare the parameter estimation methods for Weibull distribution. We consider maximum likelihood and Bayes estimation methods for the scale and shape parameters of Weibull distribution. While computing the Bayes estimates for a Weibull distribution, the continuous conjugate joint prior distribution of the shape and scale parameters does not exist and the closed form expressions of the Bayes estimators cannot be obtained. In this study, we assume that the scale and shape parameters have the exponential prior and they are independently distributed. We use the Lindley approximation and the Markov Chain Monte Carlo (MCMC) method to obtain the approximate Bayes estimators. In simulation study we compare the effectiveness of the parameter estimation methods with Monte Carlo simulations.

Improving Part-of-Speech Tagging with Relative Positional Encoding in Transformer Models and Basic Rules

2025-04-24T16:09:39+00:00

Introduction: Part-of-speech (POS) tagging plays a pivotal role in natural language processing (NLP) tasks such as semantic parsing and machine translation. However, challenges with ambiguous and unknown words, along with limitations of absolute positional encoding in transformers, often affect tagging accuracy. This study proposes an enhanced POS tagging model integrating relative positional encoding and a rule-based correction module. Methods: The model utilizes a transformer-based architecture equipped with relative positional encoding to better capture token dependencies. Word embeddings, POS tag embeddings, and relative position embeddings are combined and processed through a multi-head attention mechanism. Following the initial classification by the transformer, a corrective rule-based module is applied to refine misclassified tokens. The approach was evaluated using the Groningen Meaning Bank (GMB) dataset, comprising over 1.3 million tokens. Results: The transformer model achieved an accuracy of 98.50% prior to rule-based corrections. After applying the rule-based module, overall accuracy increased to 99.68%, outperforming a comparable model using absolute positional encoding (98.60%). Additional evaluation metrics, including a precision of 0.92, recall of 0.89, and F1-score of 0.90, further validate the model’s effectiveness. Conclusions: Incorporating relative positional encoding significantly enhances the transformer’s contextual understanding and performance in POS tagging. The addition of a rule-based correction module improves classification accuracy, especially for linguistically ambiguous tokens. The proposed hybrid model demonstrates robust performance and adaptability, offering a promising direction for future multilingual POS tagging systems.

Application of Mamdani Fuzzy Logic in Identifying Postpartum Depression Risk

2025-04-24T16:02:00+00:00

Introduction: Postpartum depression (PPD) is a common psychological disorder affecting mothers after childbirth, often underdiagnosed due to the subjective nature of its symptoms. Early detection is crucial to prevent adverse effects on maternal and child health. This study aims to develop an early detection system for PPD risk using Mamdani fuzzy logic, which is well-suited to handle vague and imprecise symptom data. Methods: A fuzzy inference system was designed using the Mamdani method to classify PPD risk into Low, Medium, and High categories. The system was built upon a dataset of 1503 questionnaire responses sourced from Kaggle. Subjective symptoms such as sadness, irritability, sleep disturbances, and bonding difficulties were mapped into fuzzy membership functions. A total of 243 fuzzy rules were defined to reflect realistic combinations of symptoms. The system was implemented and validated in both Python and LabVIEW environments. Results: Experimental validation using 10 test inputs showed consistent results between the two platforms, with a deviation of less than ±1%. This consistency confirms the reliability of the fuzzy logic model in interpreting subjective symptom data. The system demonstrated strong potential for classifying PPD risk based on nuanced input variables. Conclusions: The Mamdani fuzzy logic system offers a reliable and flexible tool for assessing postpartum depression risk. By effectively interpreting ambiguous symptoms, it supports healthcare professionals in identifying at-risk individuals for early intervention. Future enhancements should include expanding the dataset and refining the rule base for broader applicability and improved accuracy.

Performance Comparasion of DenseNet-121 and MobileNetV2 for Cacao Fruit Disease Image Classification

2025-04-24T16:00:11+00:00

Introduction: Cacao fruit diseases significantly impact cocoa yield and quality in Indonesia. Early detection is critical, yet traditional methods are often time-consuming and error-prone due to the visual similarity of disease symptoms. This study aims to compare the performance of two Convolutional Neural Network (CNN) architectures—DenseNet-121 and MobileNetV2—for automated image classification of cacao fruit diseases. Methods: The dataset comprises 8000 images augmented from 2000 original photos taken in cocoa plantations in Bali, categorized into four classes: fruit rot, fruit-sucking pests, pod borers, and healthy fruits. Both CNN models were trained using the Adam optimizer, a learning rate of 0.001, and a dropout rate of 0.4. The input images were resized to 224×224 pixels. Evaluation metrics included accuracy, precision, recall, and F1-score. Results: DenseNet-121 outperformed MobileNetV2 across all metrics. DenseNet-121 achieved an accuracy of 94.50%, precision of 94.75%, recall of 94.25%, and F1-score of 94.50%. In comparison, MobileNetV2 reached an accuracy of 93.88%. Although MobileNetV2 offered faster training time and lower model complexity, DenseNet-121 demonstrated superior feature extraction and stability, supported by its deeper architecture and greater parameter capacity. Conclusions: DenseNet-121 is more effective than MobileNetV2 in classifying cacao fruit diseases, providing higher accuracy and robustness. Despite requiring more computational resources, it is better suited for developing a web-based cocoa disease detection tool to assist farmers in timely and accurate disease identification.

Implementation of Ensemble Deep Learning for Brain MRI Classification in Tumor Detection

2025-04-24T15:57:06+00:00

Introduction: Brain tumor detection from MRI images is critical for early diagnosis and treatment planning. While individual deep learning models have shown high accuracy in medical image classification, combining multiple models can potentially enhance performance. This study aims to develop an ensemble deep learning framework using ResNet18 and DenseNet121 to improve the accuracy of brain tumor classification. Methods: A dataset of 7,023 brain MRI images categorized into four classes—glioma, meningioma, no tumor, and pituitary tumor—was used. Pre-processing included resizing to 224×224 pixels, normalization, and augmentation (random flipping and rotation). ResNet18 and DenseNet121 models were fine-tuned separately using the Adam optimizer with a learning rate of 0.001. The ensemble method was implemented by averaging the softmax outputs of both models to generate final predictions. Results: When evaluated individually, ResNet18 and DenseNet121 achieved validation accuracies of 97.72% and 97.79%, respectively. The ensemble model significantly outperformed both, reaching a validation accuracy of 99.36%. This result demonstrates that integrating both architectures effectively reduces misclassification and enhances overall robustness. Confusion matrix analysis confirmed high classification accuracy across all four tumor categories. Conclusions: The proposed ensemble deep learning approach successfully leverages the strengths of ResNet18 and DenseNet121, achieving superior classification accuracy for brain tumor detection in MRI images. This method holds promise as a reliable tool in clinical diagnostic workflows. Future research should focus on integrating additional architectures, advanced augmentation strategies, and hyperparameter optimization to further improve performance

An Optimization Strategy for Reducing CO₂ in Livestock Farming with IoT Integration and Decision Support System Approach Using Linear Programming

2025-04-24T15:58:48+00:00

Introduction: Livestock waste mismanagement contributes significantly to CO₂ emissions, adversely affecting animal health and environmental sustainability. This study aims to develop an optimization strategy for reducing CO₂ levels in livestock environments through the integration of Internet of Things (IoT) technology and a Decision Support System (DSS) using Linear Programming. Methods: IoT sensors were deployed to monitor environmental parameters such as CO₂ levels, temperature, and humidity in real time. A Linear Programming (LP) model was formulated to determine the optimal frequency of two CO₂-reducing actions: spraying Effective Microorganisms (EM4) and performing waste dredging. The objective was to maximize CO₂ reduction under cost and time constraints. The model iteratively updated its parameters based on sensor data feedback, ensuring dynamic and adaptive optimization. Results: Simulation results indicated that the LP model successfully identified optimal actions within predefined constraints. The optimal strategy was spraying EM4 eight times over eight days, achieving a CO₂ reduction of 800 ppm with a total cost of Rp 400,000—within the Rp 500,000 budget limit and 8-hour duration constraint. Validation through simulation confirmed the model’s accuracy, with prediction deviations consistently falling within an acceptable threshold (±20 ppm). Conclusions: The integration of IoT with an LP-based DSS offers a practical and efficient solution for CO₂ reduction in livestock farming. This system enhances decision-making for environmental management, demonstrating potential for scalable application in sustainable agriculture. Future work should incorporate more environmental variables and broader validation to improve model generalizability and precision.

Implementation of Support Vector Machine Algorithm for Classification of Study Period and Graduation Predicate of Students

2025-04-24T15:55:04+00:00

Introduction: Accurately predicting the duration of study and graduation predicates in higher education is essential for improving academic outcomes and decision-making. This study aims to classify students’ study period and graduation predicates in the Information Systems program at UTDI using the Support Vector Machine (SVM) algorithm. Methods: A dataset of 500 student records containing academic and demographic variables—including GPA, age, semesters, and graduation predicates—was processed through data cleaning, normalization, and feature selection. Study duration was categorized into three classes: short (≤4 years), medium (4–6 years), and long (>6 years). An SVM with a linear kernel was applied, and the model was evaluated using accuracy, precision, recall, and F1-score. Results: The SVM model achieved perfect classification for study duration, with 100% accuracy, precision, recall, and F1-score across all categories. For graduation predicate classification, the model attained 95.18% accuracy. While it performed well overall, it faced some difficulty distinguishing between "Cum Laude" and "Very Satisfactory" due to overlapping GPA ranges. The analysis identified GPA as the most influential feature in both classification tasks, while age and the number of semesters played supporting roles. Conclusions: The SVM model demonstrates strong capability in classifying study duration and graduation predicates, offering valuable insights for academic management. Although performance was high, especially for study period prediction, further refinement is suggested to enhance classification in overlapping categories. Future work may benefit from larger, more balanced datasets and exploration of advanced models to increase prediction reliability.

Comparison of ResNet-50 and DenseNet-121 Architectures in Classifying Diabetic Retinopathy

2025-04-24T16:03:27+00:00

Introduction: Diabetic Retinopathy (DR) is a vision-threatening complication of diabetes that requires early and accurate diagnosis. Deep learning offers promising solutions for automating DR classification from retinal images. This study compares the performance of two convolutional neural network (CNN) architectures—ResNet-50 and DenseNet-121—for classifying DR severity levels. Methods: A dataset of 2,000 pre-processed and augmented retinal images was used, categorized into four classes: normal, mild, moderate, and severe. Both models were trained using two approaches: standard train-test split and Stratified K-Fold Cross Validation (k=5). Data augmentation techniques such as flipping, rotation, zooming, and translation were applied to enhance model generalization. The models were trained using the Adam optimizer with a learning rate of 0.001, dropout of 0.2, and learning rate adjustment via ReduceLROnPlateau. Performance was evaluated using accuracy, precision, recall, and F1-score. Results: ResNet-50 outperformed DenseNet-121 across all evaluation metrics. Without K-Fold, ResNet-50 achieved 84% accuracy compared to DenseNet-121’s 80%; with K-Fold, ResNet-50 scored 83% and DenseNet-121 81%. ResNet-50 also demonstrated better balance in class-wise classification, with higher recall and F1-score, especially for moderate and severe DR classes. Confusion matrices confirmed fewer misclassifications with ResNet-50. Conclusions: ResNet-50 provides superior accuracy and robustness in classifying DR severity levels compared to DenseNet-121. While K-Fold Cross Validation enhances model stability, it slightly reduces overall accuracy. These findings support the use of ResNet-50 in developing reliable deep learning-based screening tools for early DR detection in clinical practice

Evaluating Machine Learning Approaches: A Comparative Study of Random Forest and Neural Networks in Grade Classification

2025-04-24T15:50:16+00:00

Introduction: Accurate grade classification in education is essential for early intervention and performance assessment. This study presents a comparative analysis of Random Forest and Neural Networks in classifying student grades using a dataset of 2,392 high school students. The aim is to evaluate both models’ predictive performance and interpretability in an educational data mining context. Methods: The dataset, containing academic and demographic features, was pre-processed by handling missing values, encoding categorical variables, and scaling numerical features. Grades were categorized into five classes: A, B, C, D, and F. Both models were implemented using Python and evaluated with metrics including accuracy, precision, recall, and F1-score. Hyperparameter tuning was performed via Grid Search with cross-validation to optimize performance. Results: The Random Forest model achieved a baseline accuracy of 70.2%, outperforming Neural Networks at 69.1%. After tuning, Random Forest improved to 71.45% accuracy, while Neural Networks reached 70.49%. Both models demonstrated strong precision and recall in identifying failing students (class F), with F1-scores of 0.90 and 0.89, respectively. However, classification of mid-range grades (A to D) remained challenging due to class overlap. Feature importance analysis highlighted interpretability advantages in the Random Forest model. Conclusions: Both models are effective for grade classification, with Random Forest offering slightly better accuracy and interpretability. Neural Networks, while slightly less accurate, capture nonlinear relationships effectively post-tuning. The results suggest that model selection should be guided by context-specific needs, balancing performance with transparency. Future work may include ensemble techniques and expanded feature sets to improve classification robustness.

Churn Prediction in Credit Customers Using Random Forest and XGBoost Methods

2025-04-24T15:51:40+00:00

Introduction: Customer churn in the credit card industry presents a significant challenge for financial institutions, potentially resulting in substantial revenue loss. This study aims to develop predictive models for identifying credit card customers likely to churn, thereby enabling proactive retention strategies. Methods: A dataset of 5,000 credit card customer records was used, including 800 churn and 4,200 non-churn instances, reflecting a class imbalance addressed using the Synthetic Minority Over-sampling Technique (SMOTE). Two machine learning models—Random Forest and XGBoost—were implemented. Data pre-processing involved feature scaling, categorical encoding, and class balancing. Key predictive features included age, marital status, education level, transaction count, and total transaction value. Both models underwent hyperparameter tuning to optimize performance. Results: The Random Forest model achieved a baseline accuracy of 95%, improving to 96% after tuning, with an F1-score of 88% for the churn class. XGBoost demonstrated consistent accuracy of 96% before and after tuning but outperformed in minority class detection with an F1-score of 87%, precision of 86%, and recall of 89%. Analysis revealed that customers aged 40–55 were more likely to churn, influenced by behavioral and demographic factors. Conclusions: Both Random Forest and XGBoost models showed excellent performance in churn prediction. However, XGBoost proved more effective in identifying minority class instances, making it the preferred model for credit customer churn prediction. These findings support the integration of predictive analytics in customer retention strategies within the banking sector.

Comparative Analysis of Gradient-Based Optimizers in Feedforward Neural Networks for Titanic Survival Prediction

2025-07-08T03:37:00+00:00

Introduction: Feedforward Neural Networks (FNNs), or Multilayer Perceptrons (MLPs), are widely recognized for their capacity to model complex nonlinear relationships. This study aims to evaluate the performance of various gradient-based optimization algorithms in training FNNs for Titanic survival prediction, a binary classification task on structured tabular data. Methods: The Titanic dataset consisting of 891 passenger records was pre-processed via feature selection, encoding, and normalization. Three FNN architectures—small ([64, 32, 16]), medium ([128, 64, 32]), and large ([256, 128, 64])—were trained using eight gradient-based optimizers: BGD, SGD, Mini-Batch GD, NAG, Heavy Ball, Adam, RMSprop, and Nadam. Regularization techniques such as dropout and L2 penalty, along with batch normalization and Leaky ReLU activation, were applied. Training was conducted with and without a dynamic learning rate scheduler, and model performance was evaluated using accuracy, precision, recall, F1-score, and cross-entropy loss. Results: The Adam optimizer combined with the medium architecture achieved the highest accuracy of 82.68% and an F1-score of 0.77 when using a learning rate scheduler. RMSprop and Nadam also performed competitively. Models without learning rate schedulers generally showed reduced performance and slower convergence. Smaller architectures trained faster but yielded lower accuracy, while larger architectures offered marginal gains at the cost of computational efficiency. Conclusions: Adam demonstrated superior performance among the tested optimizers, especially when coupled with learning rate scheduling. These findings highlight the importance of optimizer choice and learning rate adaptation in enhancing FNN performance on tabular datasets. Future research should explore additional architectures and optimization strategies for broader generalizability

Use of Machine Learning in Power Consumption Optimization of Computing Devices

2025-04-24T15:45:22+00:00

Introduction: The high-power consumption of computing devices poses both economic and environmental challenges in the digital era. This study aims to optimize power usage using machine learning to maintain device performance while reducing energy costs and carbon emissions. Methods: The Random Forest algorithm was selected for its robustness in handling non-linear interactions among features. A dataset containing historical power consumption, workload metrics, environmental conditions, and hardware configurations was collected from sensors and logs. Data pre-processing included cleaning, normalization, and feature selection. The model was trained and evaluated using accuracy, precision, recall, F1-score, MAE, and RMSE metrics. Hyperparameter tuning via grid search, random search, and Bayesian optimization was applied to enhance model performance. The model was deployed on real devices to test energy optimization under varied workloads. Results: The Random Forest model achieved 92% accuracy and an RMSE of 0.15. Tuning reduced RMSE by 10% and improved F1-score from 0.875 to 0.905. Implementation on computing devices led to average power savings of 15–20% across workload scenarios without notable performance degradation (<5%). The model also projected annual carbon emission reductions of up to 5 tons of CO₂ and operational savings of $50,000 when scaled to 1,000 servers. Conclusions: Machine learning, particularly Random Forest, proves effective in optimizing power consumption on computing devices. The proposed approach not only ensures computational efficiency but also promotes environmental sustainability. These findings support further exploration of ML-based solutions for green technology initiatives in IT infrastructure.

Optimizing Javanese Numeral Recognition Using YOLOv8 Technology: An Approach for Digital Preservation of Cultural Heritage

2025-04-24T15:43:30+00:00

Introduction: The preservation of Javanese script as part of Indonesia’s cultural heritage is increasingly urgent in the digital era, especially due to declining literacy among younger generations. This study aims to explore the effectiveness of YOLOv8, an advanced object detection algorithm, for recognizing handwritten Javanese numerals to support efforts in cultural digitization and education. Methods: A dataset of 2,790 handwritten Javanese numerals (0–9) was collected from 93 respondents. Each numeral was manually annotated using bounding boxes via the MakeSense.ai platform. The YOLOv8 model was trained using 80% of the data and validated on the remaining 20%. Training was performed in the PyTorch framework with data augmentation techniques to increase robustness. Model performance was evaluated using precision, recall, F1-score, and mean Average Precision (mAP), along with visualization through confidence curves and confusion matrices. Results: The model achieved a high validation precision of 88.3%, recall of 89.1%, and mAP of 0.88 at IoU 0.90. F1-score peaked at a confidence threshold of 0.89, while certain numerals like 'six' and 'nine' achieved near-perfect detection. Visualizations confirmed the model’s ability to accurately classify and localize characters in both training and unseen data. Minor misclassifications occurred between visually similar numerals. Conclusions: YOLOv8 demonstrates high effectiveness in recognizing handwritten Javanese numerals and holds significant potential for digital heritage preservation. Future work should focus on expanding the dataset, improving generalization under varied conditions, and integrating this model into educational tools and augmented reality applications for interactive learning.

A Comperative Study on Efficacy of CNN VGG-16, DenseNet121, ResNet50V2, And EfficientNetB0 in Toraja Carving Classification

2025-04-24T16:06:41+00:00

Introduction: Passura', or Toraja carvings, are an essential element of the cultural heritage of the Toraja people in Indonesia. These carvings feature complex motifs rooted in nature, folklore, and spiritual symbolism. This study aims to evaluate the efficacy of four Convolutional Neural Network (CNN) architectures—VGG-16, DenseNet121, ResNet50V2, and EfficientNetB0—in classifying seven traditional Toraja carving motifs. Methods: A dataset of 700 images was collected and categorized into seven motif classes. The dataset was split into 80% for training and 20% for validation. Each CNN model was trained for 25 epochs with standard pre-processing, including resizing to 224×224 and normalization. Performance evaluation was conducted based on validation accuracy and confusion matrix analysis to assess classification precision and model overfitting. Results: EfficientNetB0 achieved the highest validation accuracy of 98%, although signs of overfitting were observed. ResNet50V2 followed closely with a validation accuracy of 95.33% and demonstrated the most balanced classification results across all motif categories. VGG-16 and DenseNet121 achieved 94.67% and 81.82%, respectively. Confusion matrix analysis confirmed the robustness of ResNet50V2 in correctly identifying complex patterns. Conclusions: The findings indicate that ResNet50V2 provides a reliable balance between accuracy and generalizability for classifying Toraja carvings, making it suitable for digital preservation of cultural heritage. EfficientNetB0, while achieving higher accuracy, may require additional regularization to avoid overfitting. This study contributes to the development of AI-driven cultural documentation and suggests future research with larger and more diverse datasets to improve model robustness

Sugeno Fuzzy Logic for IoT-based Chicken Farm Drinking Water Quality Monitoring

2025-04-24T16:05:07+00:00

Introduction: The quality of drinking water plays a crucial role in the health and productivity of broiler chickens. In Indonesia, many poultry farms still rely on manual water testing using litmus paper, which may yield inaccurate results. This study aims to develop an Internet of Things (IoT)-based system integrated with the Sugeno fuzzy logic method to monitor and assess the quality of drinking water for broiler chickens in real time. Methods: An IoT prototype was developed using an ESP32 microcontroller, pH and turbidity sensors, and a cloud-based mobile application. Water quality data from 1,975 samples were collected over three days from a broiler farm in East Kalimantan using water sourced from a former mining lake. The system applies the Sugeno fuzzy inference system with 15 expert-defined rules to classify water quality into four categories: Very Good, Good, Bad, and Very Bad. Performance was evaluated using a Confusion Matrix. Results: The system achieved a classification accuracy of 96.76%, precision of 97.52%, recall of 98.79%, and F1-score of 98.15%. The results demonstrate the system's effectiveness in identifying water quality, with the majority of predictions falling into the correct class. The system also successfully transmitted real-time data to an Android application for monitoring purposes. Conclusions: The integration of IoT and Sugeno fuzzy logic provides a reliable, accurate, and scalable solution for real-time water quality monitoring in poultry farming. This system enhances decision-making for farmers, supports animal welfare, and can be further developed to include additional environmental parameters for broader livestock health monitoring