Refining Graduation Classification Accuracy with Synergistic Deep Learning Models

Learning Analytics plays an important role in monitoring and improving educational outcomes, but is often challenged by limited dataset sizes, resulting from privacy regulations and curriculum changes. This paper proposes the LATCGAd (Learning Analysis by Transformer with Conditional Generative Adversarial Framework Network and Adaptive Layer Normalization model), a deep learning framework that combines the Transformer architecture and Conditional Generative Adversarial Network (CGAN) to overcome the above problems. The CGAN component generates synthetic data samples, which balance and expands the dataset size, while the Transformer leverages this rich dataset to improve prediction performance. The integration of Adaptive Layer Normalization (AdaLN) in the Transformer also helps stabilize the learning process and minimize overfitting. Experiments on datasets from Hanoi Metropolitan University and Hanoi National University show that the LATCGAd model achieves an accuracy of up to 96.97%, outperforming traditional models such as Decision Tree, SVM and Transformer alone. This result confirms the effectiveness of LATCGAd in educational predictive analysis and its potential for widespread application in the field of learning analytics.