The rapid growth of data in various domains has heightened the need for accurate and efficient predictive models, particularly for tabular data. While deep learning has revolutionized fields like computer vision and natural language processing, its effectiveness in tabular data classification remains a topic of debate. This study conducts a comprehensive comparison between deep learning models (NODE, SAINT, TabNet, Tab Transformers) and tree-based models (Random Forest, XGBoost, LightGBM, CatBoost) to determine whether deep learning can outperform traditional methods in this context. The results indicate that tree-based models, particularly LightGBM and CatBoost, consistently achieve the highest accuracy and F1 scores, coupled with efficient execution times, making them more suitable for real-world applications that require quick and accurate predictions. In contrast, deep learning models show varied performance. Although SAINT sometimes achieves comparable accuracy, its processing time makes it less practical. The findings suggest that despite the potential of deep learning, tree-based models remain superior for tabular data classification tasks, particularly when considering a balance of accuracy, speed, and robustness. This study contributes to the ongoing discussion on the role of deep learning in tabular data and highlights the conditions under which traditional models may still be preferred.
Imagine having a personalized nutrition plan that caters to your unique dietary needs based on your age and gender. Such a system could revolutionize the way we approach health and wellness. A key component of this vision is the accurate classification of age and gender from facial images, which can be leveraged to provide tailored nutritional recommendations. In this paper, we explored the use of convolutional neural networks and their preprocessing techniques to enhance the performance of age and gender classification models. Our age classification aimed to identify the age group according to the regulation of the Indonesian Republic's Health Ministry in 2014 about the guidelines for balanced nutrition, which includes the following categories: 10-12 years old, 13-15 years old, 16-18 years old, 19-29 years old, and 30-49 years old. We utilized the ResNet50 and Inception-v3 models, which were fine-tuned on the UTKFace dataset, a collection of more than 20,000 face images with corresponding age and gender labels. However, the UTKFace dataset suffers from a data imbalance problem. To address this issue, we proposed innovative data augmentation methods to create a more balanced dataset. Our experimental results demonstrated that our proposed augmented methods could significantly improve the classification performances of the models, leading to more accurate age and gender predictions. This advancement in facial attribute classification could pave the way for developing personalized nutrition systems that cater to individual needs and preferences, ultimately improving health outcomes and quality of life.
The automotive industry in Indonesia has gradually experienced rapid development over time, leading to intense competition among automotive companies. In such a competitive environment, the quality inspection process is a critical aspect of the automotive manufacturing industry. One significant issue in the quality inspection process is human error. The Fourth Industrial Revolution offers solutions through computer vision technology to improve production quality. YOLOv8 is a real-time object detection algorithm that offers fast inference and good background reduction, but it is less effective for small and low-contrast objects. Conversely, Faster R-CNN has high confidence scores but slow inference capabilities. Company XYZ has a visual inspection system using YOLOv8, with a mAP50@95 value of 74%. However, the confidence scores produced often do not meet the company's standard threshold, which requires a confidence score above 90%. Therefore, a study was conducted to combine the inference results of YOLOv8 and Faster R-CNN using the weighted boxes fusion method to enhance the inference results of YOLOv8. The study results showed an increase in mAP when combining the inferences of the two models, compared to when each model performed inference individually – there was a 3.8% increase compared to YOLOv8's performance and a 5.5% increase compared to Faster R-CNN's performance.
Early identification of brain tumors using magnetic resonance imaging helps doctors make quick and informed decisions about treatment. Although recent deep learning approaches achieve high accuracy, many rely on complex architectures that increase computational cost and limit interpretability. In order to overcome these constraints, this work proposes a system for four-class brain tumor classification utilizing a public MRI dataset of 3,264 images that is built on EfficientNetV2-B3 and an Efficient Channel Attention (ECA) module used after feature extraction and Grad-CAM. The ECA module enhances cross-channel feature representation with minimal computational overhead. Experimental results indicate consistent performance gains over the baseline model, with accuracy increasing from 97.58% to 99.09% and macro-averaged F1-score from 97.51% to 99.08%. Despite the strong baseline, these gains are achieved without increasing architectural complexity. Grad-CAM visualizations support model interpretability by highlighting tumor-relevant regions that contribute most to the classification decisions. Overall, the proposed framework provides a balanced trade-off between classification accuracy, computational efficiency, and interpretability within the evaluated setting.
Vocational high schools (SMK) aim to produce work-ready graduates. However, the open unemployment rate (TPT) for SMK graduates remains high at 9.01%, indicating a significant competency gap. This study designs a model to predict graduates workforce competitiveness using the Naive Bayes algorithm combined with the Synthetic Minority Over-sampling Technique (SMOTE). SMOTE is employed to address the class imbalance between capable and incapable graduates. The study follows the Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology, utilizing academic scores and tracer study datasets. Evaluation results demonstrate that applying SMOTE with a 70:30 train-test split successfully increased model accuracy to 97%. Notably, the model effectively detects the minority class with a Recall of 90%. Furthermore, cross-validation yielded an average accuracy of 97.66%, demonstrating stable performance. Finally, the model was implemented as a web-based dashboard to serve as an early warning system for schools.
