Search

This study aims to analyze the development of deep learning methods for optimizing complex data classification model performance through a Systematic Literature Review (SLR) approach examining 25 Scopus-indexed scientific articles published between 2024 and 2025. The analysis employs bibliometric techniques using VOSviewer to map keyword networks, temporal trends, and term density patterns. Visualization results identify three primary clusters: (1) LSTM-based classification and intrusion detection systems in cybersecurity applications; (2) CNN optimization and model efficiency for medical imaging and satellite image classification; and (3) artificial intelligence integration with visual classification and evolutionary optimization algorithms. Recent trends demonstrate the dominance of keywords such as "optimization," "effectiveness," and "feature selection," alongside growing interest in hybrid approaches and metaheuristic algorithms. This research provides a comprehensive overview of methodological transformations and application directions of deep learning in complex data classification domains. These findings are expected to serve as strategic references for advancing research and applications in big data-driven artificial intelligence technologies.

This study investigates public sentiment toward the first 100 days of the Prabowo–Gibran administration by analyzing opinions expressed on X (formerly Twitter) using machine learning approaches. A total of 431 valid tweets were collected, preprocessed, and manually labeled into positive and negative categories. The results reveal that 62% of public sentiment was negative, while 38% was positive, indicating widespread public criticism during the administration’s early period. Two algorithms, Naïve Bayes and K-Nearest Neighbor (KNN), were applied to classify sentiment. The Naïve Bayes model achieved superior performance, with an accuracy of 97.22%, compared to KNN’s 62.65%. The probabilistic nature of Naïve Bayes allowed it to manage high-dimensional, imbalanced textual data effectively, while KNN suffered from the “curse of dimensionality” and class bias. These findings demonstrate that Naïve Bayes remains a reliable and computationally efficient model for political sentiment analysis in the Indonesian digital context. Despite its strengths, this study acknowledges limitations in manual labeling and linguistic nuances such as sarcasm and irony. Future research is encouraged to integrate deep learning architectures like LSTM or BERT and adopt aspect-based sentiment analysis to capture more contextual insights from political discourse.

This research introduces a mobile-based system for real-time identification of ornamental rose varieties using the YOLOv8 deep learning algorithm. Motivated by the growing interest in ornamental plants during the COVID-19 pandemic and the high penetration of smartphone users in Indonesia, the study aims to create an efficient and accessible flower recognition tool. A dataset of 813 labeled rose images—red, white, yellow, orange, and pink—was collected from the Roboflow platform and processed using data augmentation techniques to improve model generalization. The YOLOv8 model was trained with 100 epochs, a batch size of 16, and the SGD optimizer, then converted to TensorFlow Lite for mobile deployment through the Flutter framework. Experimental results achieved a mean average precision (mAP50–95) of 0.581, with strong detection performance across most classes. The system successfully operated offline, delivering real-time classification accuracy despite dataset imbalance, particularly in the orange rose class. These findings demonstrate that YOLOv8 can be effectively adapted for mobile horticultural applications, offering practical benefits for flower sorting, crop management, and educational use. Future studies are recommended to expand dataset diversity, enhance environmental testing, and explore cloud-based integration for scalable deployment.

This study examines the effectiveness of classical computer vision methods for modern waste classification by combining Histogram of Oriented Gradients (HOG) for feature extraction with Support Vector Machine (SVM) for classification. The TrashNet dataset, consisting of five categories—cardboard, glass, metal, paper, and plastic—was used as the primary benchmark. To address data limitations and improve generalization, augmentation techniques such as random rotations, horizontal flipping, and brightness adjustments were applied. Hyperparameter optimization was further conducted using GridSearchCV with the RBF kernel to determine the most effective configuration. The optimized model achieved an accuracy of 84.36%, representing a substantial improvement from the 60% baseline. These findings confirm that non-deep learning approaches remain relevant and can serve as computationally efficient alternatives to CNNs, which typically require GPUs and extensive training time. Challenges persist in classifying reflective materials such as glass and metal, where HOG descriptors are less effective. Future work should integrate complementary descriptors, including color and texture-based features, to enhance robustness and scalability. Overall, the study demonstrates that an optimized HOG-SVM pipeline offers a practical, resource-efficient solution for automated waste classification, with strong potential to support sustainable waste management in real-world applications.

The rapid expansion of e-commerce in Indonesia has resulted in a surge of unstructured online reviews, especially on platforms such as Shopee. These reviews offer valuable insights into customer satisfaction, product complaints, and purchasing behavior but remain largely underutilized due to their volume and informal language style. This study applies Support Vector Machine (SVM) with Term Frequency–Inverse Document Frequency (TF-IDF) feature extraction to classify reviews of Android smartphones into positive, negative, and neutral categories. Using a dataset of 300 manually annotated reviews from Samsung, Xiaomi, and Oppo official stores, the model achieved an accuracy of 76.67% and demonstrated stable results through 5-fold cross-validation. The negative class showed the highest performance (F1 = 0.86), while the neutral class performed weakest (F1 = 0.62), reflecting challenges posed by mixed opinions and underrepresented samples. Compared with Naïve Bayes and Logistic Regression, the SVM model consistently outperformed both baselines, confirming its suitability for high-dimensional text data and informal Indonesian expressions. The findings highlight SVM’s potential to support automated sentiment monitoring in e-commerce, enabling businesses to identify emerging issues, improve customer service strategies, and leverage positive reviews for marketing. Future research should consider larger and more balanced datasets, techniques for handling imbalanced classes, and integration with deep learning models such as LSTM or BERT to improve performance and generalization.