A Student Performance Prediction Using RNNs Models with variety of optimization Techniques in Deep Learning

The goal of learning analytics is to assess students performance over time. While virtual learning environments enable educators to intervene quickly, distance can make it challenging to evaluate students’ success. Many studies have developed prediction models using data from Massive Open Online Courses (MOOCs), but these models were limited to classifying students into binary groups based on the courses they had completed. The paper tackles an important gap in predicting student performance by introducing a daily multiclass model based on Recurrent Neural Networks (RNNs), specifically leveraging Gated Recurrent Units (GRUs) and Long Short-Term Memory (LSTM) networks. To validate the GRU model, it is compared against two baseline models: Artificial Neural Networks (ANNs) and LSTM networks. The results show that the GRU model achieves an impressive accuracy of nearly 90%, outperforming the LSTM model, which reaches 88% accuracy. This highlights the potential of GRUs to better capture temporal dependencies and patterns in student performance data, making them a strong candidate for educational forecasting. The study suggests that GRU-based models could serve as a powerful tool for educators and institutions to predict and address student performance issues proactively. This demonstrates how early student performance in MOOCs can be predicted using the design and latent dependency maintenance capabilities of the GRU time series model. Along with this, the paper evaluates the accuracy, loss, and training time of several stochastic gradient descent algorithms, such as Adam, AdaGrad, RMSProp, and Stochastic Gradient Descent (SGD) with momentum. It also assesses the loss that each algorithm must endure in order to produce an optimal solution. When comparing all optimizers utilizing the pre-trained GRU model, the highest accuracies of 97.58% and 97.66% are obtained by Adam and SGD with momentum, respectively.