Public Research Journal of Engineering, Data Technology and Computer Science

A Comparison of Machine Learning Algorithms in Predicting Students' Academic Performance

Sat, 31 Jan 2026 00:00:00 +0000

Predicting students’ academic performance enables early interventions and data-driven planning in education. We compare five machine-learning algorithms Decision Tree, K-Nearest Neighbor, Naive Bayes, Random Forest, and Support Vector Machine on a publicly available dataset of 1,001 students, evaluated with Accuracy, Precision, Recall, and F1-Score. The Decision Tree achieved the highest performance, with perfect scores on this dataset, while SVM (?82% F1) and Random Forest (?81% F1) were competitive. These results suggest that simple, interpretable models can be highly effective when features are clean and predictive; however, the Decision Tree’s perfection also indicates potential overfitting and warrants further validation on larger, more diverse samples. The study underscores how model choice should reflect dataset characteristics and practical deployment goals in educational settings, informing early-warning systems and targeted support programs.

Comparison of Convolutional Neural Network and Recurrent Neural Network Algorithms for Indonesian Sign Language Recognition

Dani Harmade, Afif Fathin, Nur Jannah Nai'mah Zainal — Sun, 01 Feb 2026 00:00:00 +0000

Effective communication is a fundamental human need; however, for people with hearing impairments in Indonesia, interaction relies heavily on the Indonesian Sign Language System (Sistem Isyarat Bahasa Indonesia – SIBI). Although deep learning has been widely applied in sign language recognition, comprehensive comparative studies focusing specifically on SIBI remain limited, particularly in evaluating the performance gap between different neural network architectures. This study addresses this gap by comparing the effectiveness of Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) in classifying SIBI hand gesture images. An augmented SIBI dataset was trained using the Adam optimizer to improve generalization and recognition performance. The experimental results reveal a significant performance difference between the two models, where CNN achieved a precision, recall, and F1-score of 94%, while RNN obtained a precision of 76% recall of 74%, and F1-score of 73%. These findings demonstrate that CNN is substantially more effective for image-based SIBI recognition because it extracts spatial features more effectively than the sequential processing mechanism of RNN. This research contributes empirical evidence for selecting appropriate deep learning architectures in SIBI recognition systems and offers practical implications for developing more accurate and reliable assistive communication technologies in educational and accessibility contexts.

Classification of Breast Cancer Ultrasound Images Using Convolutional Neural Network

Rifsya Aulia, Dina Pani Safira, Khaury Audilla, Raudhatul Khairiyah — Sat, 31 Jan 2026 00:00:00 +0000

Breast cancer ranks among the primary contributors to female mortality, thereby underscoring the critical importance of early detection. This research employs a deep learning approach based on Convolutional Neural Networks (CNNs) to classify breast cancer using ultrasound imagery, comparing the ResNet50V2 and MobileNetV2 architectures with three optimizers: Adam, RMSprop, and SGDM. The dataset used in this study is the Breast Ultrasound Images (BUSI) dataset, obtained from Kaggle, which comprises three diagnostic categories: benign, malignant, and normal. The research workflow encompassed several stages, including data acquisition, image pre-processing involving normalization and augmentation, and dataset partitioning using the Holdout Split method, with proportions of 70% for training, 15% for validation, and 15% for testing. The experimental findings revealed that the ResNet50V2 architecture combined with the SGDM optimizer achieved the best performance, recording accuracy, precision, recall, and F1-score values of 92%. Meanwhile, MobileNetV2 with RMSprop achieved the highest performance on its architecture with 86% accuracy, 88% precision, 86% recall, and 86% F1-score. These findings prove that CNN architecture selection and optimization algorithms have a significant influence on medical image classification performance.

Classification of Corn Leaf Disease Images Using Convolutional Neural Network Algorithm

Irma Fitriani, Rahma Devi, Ariandra Fokker Chaya Sajjana, Muhammad Irfan — Sat, 31 Jan 2026 00:00:00 +0000

Corn leaf diseases can reduce crop yields and cause financial losses, thus requiring accurate and objective classification methods. This study aims to classify four corn leaf conditions, namely Blight, Common Rust, Gray Leaf Spot, and healthy leaves, using a Convolutional Neural Network (CNN) approach based on image processing. A systematic comparative evaluation was conducted on three CNN architectures, namely MobileNetV2, ResNet50V2, and DenseNet201, by examining the effect of architecture-optimizer pairs using Adam and RMSprop to determine the optimal model configuration. The results showed that the proposed approach was effective in classifying corn leaf diseases, with the highest accuracy of 93% achieved by the combination of DenseNet201 and the Adam optimizer. This study contributes by providing a structured comparative analysis of the performance of CNN architectures and optimizers as a reference for the development of more accurate and efficient early detection systems for plant diseases.

Comparative Study of Convolutional Neural Network Architectures and Optimizers for Flower Image Classification

Ekatri Yulisara, Nayla Husna, David Martin, Candrawati Ariesta — Sun, 01 Feb 2026 00:00:00 +0000

This study aims to comparatively evaluate the performance of different Convolutional Neural Network (CNN) architectures and optimization algorithms for flower image classification. Three widely used CNN architectures DenseNet201, InceptionV3, and MobileNetV2 are implemented using transfer learning with pre-trained ImageNet weights and tested with two optimizers, Adam and RMSProp. The experiments are conducted on the Flowers Recognition dataset consisting of five flower classes: daisy, dandelion, rose, sunflower, and tulip. Image normalization and data augmentation are applied to improve model generalization, while performance is evaluated using accuracy, precision, recall, and F1-score. The main contribution of this study lies in a systematic comparison of CNN architectures and optimizers within a unified experimental framework, which is rarely addressed in previous studies. The results show that DenseNet201 combined with the Adam optimizer achieves the highest classification accuracy of 90%, followed by MobileNetV2 with RMSProp, while InceptionV3 yields the lowest accuracy of 85%. These results confirm that the research objective is achieved, demonstrating that both CNN architecture and optimizer selection significantly influence flower image classification performance.

Sentiment Analysis of Public Opinion on the Gaza Conflict Using Machine Learning

Agil Irman Fadri, Nur Futri Ayu Jelita, Diamond Dimas Bagaskara, Raudiatul Zahra — Sun, 01 Feb 2026 00:00:00 +0000

The 2023 escalation of the Gaza conflict triggered widespread public discourse on the X platform, highlighting the importance of sentiment analysis for understanding public opinion on global geopolitical issues. While sentiment analysis has been widely applied to social media data, comparative evaluations of machine learning models on conflict-related datasets remain limited. This study analyzes public sentiment toward the Gaza conflict by comparing the performance of Multi-Layer Perceptron, XGBoost, and Logistic Regression models. A dataset of 2,175 tweets was processed using standard text preprocessing techniques and TF-IDF feature extraction. Model performance was evaluated using multiple train-test split scenarios. The results indicate that Logistic Regression consistently outperformed the other models, achieving the highest accuracy of 73.17% with an 80:20 data split. These findings suggest that simpler linear models may perform more robustly and efficiently than more complex approaches when applied to high-dimensional, noisy social media text data. This study provides practical insights into model selection for sentiment analysis of conflict-related discussions on social media platforms.

Depression Classification in University Students using A Machine Learning Approach Based on Multi-Layer Perceptron

Sun, 01 Feb 2026 00:00:00 +0000

Depression among university students is a critical mental health concern, often exacerbated by academic pressure and social adaptation. While prior studies have utilized Multi-Layer Perceptron (MLP) models to achieve up to 78% accuracy, the effectiveness of these systems remains highly sensitive to architectural design and optimization strategies. To address this gap, this study systematically evaluates the performance of modern MLP architectural variants including DenseNet, ResMLP, and ResNet paired with SGD, Adam, and RMSprop optimizers. Using a dataset of 1,025 student records, the methodology integrates Chi-Square feature selection and Min-Max normalization, followed by an 80:20 Hold-Out validation. Results demonstrate that the ResNet-RMSprop synergy yields a superior accuracy of 83.86%, significantly outperforming traditional MLP benchmarks . By identifying the optimal combination of deep learning structures and optimization algorithms, this research provides a more robust and precise technical foundation for AI-driven early detection systems in academic settings.