Journal of Sustainable Software Engineering and Information Systems

Improving the Classification Accuracy of Parang Batik Motifs with High Visual Similarity Through the Integration of GLCM and MobileNetV2

2026-02-14T09:24:29+07:00

Background: Despite its high aesthetic value, automatic classification of Parang Surakarta batik is difficult due to the extreme textural similarities between sub-motifs. Standard CNN architectures, including MobileNetV2often fail to detect the subtle textural details that distinguish each variation of the motif.

Aims: This study develops a hybrid classification model that combines manual and automated spatial texture features to improve identification accuracy on motifs with high visual similarity.

Methods: Using a dataset that has been expanded to 120 original images (40 per class) which is then augmented to a total of 1,200 images to ensure stronger model generalization. This methodology hybrid GLCM-MobileNetV2architecture through transfer learning techniques. Features from both methods are combined through feature fusion before being classified using a Dense layer.

Result: The hybrid GLCM-MobileNetV2model achieved an accuracy of 99%. This performance outperformed the pure MobileNetV2 method (66.67%) and GLCM-SVM (85%), demonstrating that texture features provide significant discriminatory power against similar repetitive patterns.

Conclusion: The integration of GLCM and MobileNetV2 is highly effective for classifying visually similar batik motifs, achieving a superior accuracy of 99% compared to the pure MobileNetV2 (66.67%). This hybrid approach provides a robust and efficient solution for digital cultural preservation on mobile devices.

Software and Information Systems for Sustainability: A Systematic Literature Review of Models, Applications, and Evaluation Metrics

2026-02-17T15:10:15+07:00

Background: The existing literature on sustainable software and information systems is fragmented, with research often siloed into specific models, applications, or evaluation metrics without a cohesive overview. This fragmentation hinders the development of a unified understanding necessary for researchers, practitioners, and policymakers to effectively implement sustainability principles.

Aims: This study aims to systematically analyze and synthesize research on software and information systems for sustainability. Its scope is to identify the dominant models, primary application domains, and key evaluation metrics used in the field to establish a consolidated understanding and guide future efforts.

Methods: A Systematic Literature Review (SLR) following PRISMA screened 314 Scopus documents (2017–2026) to 25 articles, analyzed using Biblioshiny, VOSviewer, and thematic synthesis.

Result: The analysis reveals a field in a consolidative phase, dominated by systematic review-based research (76%) focused on theoretical synthesis. While geographically diverse, the research centers on "sustainability" and "information systems" as core themes. A critical gap exists between conceptual frameworks and practical application, evidenced by a scarcity of empirical studies (only 4% quantitative) and the absence of standardized evaluation metrics.

Conclusion: This review concludes that while significant progress has been made in mapping the conceptual landscape, the field of software and information systems for sustainability must now prioritize empirical validation, the development of AI-driven systems, and the establishment of uniform measurement standards to bridge the gap between theoretical promise and tangible real-world outcomes.

Application of K-Means, Random Forest, and Linear Regression to Improve the Accuracy of Kaggle E-Commerce Shopping Behavior Analysis

2026-04-04T16:06:43+07:00

Background: Analyzing customer shopping behavior is essential for improving e-commerce marketing strategies. The use of machine learning enables the identification of purchasing patterns and enhances predictive accuracy in understanding customer preferences.

Aims: This study aims to improve model accuracy and classification performance in analyzing customer shopping behavior using the shopping_behavior_updated.csv dataset from Kaggle. The scope includes the application of both supervised and unsupervised learning techniques for segmentation, classification, and prediction tasks.

Methods: Three machine learning algorithms were applied: K-Means for customer segmentation, Random Forest Classifier for product category prediction, and Linear Regression for estimating purchase amounts. The research involved systematic preprocessing steps, including data cleaning, encoding, scaling, and feature engineering to enhance data quality and model interpretability.

Result: The results show that the optimized Random Forest model achieved 100% accuracy. K-Means clustering produced five distinct customer segments with an inertia value of 41,928.17 and a silhouette score of 0.065. However, the Linear Regression model demonstrated poor performance with an R² value of -0.02.

Conclusion: The findings indicate that integrating supervised and unsupervised learning methods is effective in identifying customer purchasing patterns and can contribute to improving e-commerce marketing strategies, although not all predictive models yield optimal performance.

Comparative Analysis of Machine Learning Algorithms for Diabetes Prediction with Feature and Hyperparameter Optimization

2026-04-03T13:41:14+07:00

Background: Diabetes is a chronic disease with increasing global prevalence, making early detection essential. Machine learning has shown strong potential in improving prediction accuracy; however, robust validation and systematic optimization are still required.

Aims: This study tries to compare different machine learning methods to predict diabetes using a. reproducible and methodologically sound framework.

Methods: The Pima Indian Diabetes dataset (768 samples, 8 clinical features) was used. Six algorithms were evaluated: Logistic Regression, K-Nearest Neighbors, Decision Tree, Random Forest, Support Vector Machine, and Gradient Boosting. Hyperparameter tuning was done with GridSearchCV, and the models were checked using stratified 5-fold cross-validation. The performance of the model was assessed using several metrics including accuracy, precision, recall, F1-score, and AUC-ROC.

Results: The results show that ensemble methods outperform traditional models. Random Forest achieved the highest The model performed with an accuracy of 98% plus or minus 1.8% and an AUC-ROC of 0.999 plus or minus 0.02, then Gradient Boosting achieved 91% plus or minus 2.1%. Logistic Regression and KNN had lower performance with accuracy scores of 79% plus or minus 2.3% and 77% plus or minus 2.5%, respectively. The analysis of which features are most important found that glucose levels, BMI, and age are the top factors that have the biggest influence.

Conclusion: The study demonstrates that ensemble methods combined with hyperparameter optimization and robust validation significantly improve diabetes prediction performance and can support clinical decision-making.

Butterfly Classification Accuracy Analysis Using EfficientNet with Adam and AdamW Optimizer

2026-04-03T16:58:00+07:00

Background: Accurate butterfly species classification plays an important role in biodiversity monitoring and environmental conservation. However, image-based classification remains challenging due to similarities in wing color patterns and shapes among species.

Aims: This study aims to evaluate the performance of butterfly image classification using transfer learning based on the EfficientNet architecture with different optimization strategies.

Methods: EfficientNet-B2 and EfficientNet-B3 were implemented as the main models. Image preprocessing techniques, including resizing, normalization, and data augmentation, were applied to improve model performance. The models were optimized using Adam and AdamW optimizers with different learning rates. Performance evaluation was conducted using accuracy, precision, recall, and F1-score.

Results: The results show that EfficientNet-B2 optimized with the Adam optimizer (learning rate 5 × 10⁻⁴) achieved the best performance, with a validation accuracy of 92.62% and an F1-score of 92.61%. In comparison, EfficientNet-B3 optimized using AdamW (learning rate 6 × 10⁻⁴) produced slightly lower performance, indicating the influence of optimizer selection and learning rate configuration on model convergence and generalization.

Conclusion: EfficientNet-B2 combined with the Adam optimizer provides a stable and effective approach for butterfly species classification. These findings highlight the importance of optimization strategies in improving model performance and support the development of automated systems for biodiversity monitoring.