An LTC-LSTM Framework for Early Prediction of Course Completion Time and Student Dropout Risk

Abstract

Ensuring student success in online and blended learning is a primary issue in education. One of the two most important are predicting the length of time that students will take to finish a course and identifying those that may be at risk of not finishing their study. These are especially challenging tasks. The reasons for this are complex, since the behaviour of students is determined by a variety of factors, including financial situation, Academic performance and engagement levels, mental well-being. Furthermore, these factors can vary considerably Predicting is a difficult problem because it is often changing, from one learner to another. Such non-linear and dynamic relationships are often not represented well by traditional predictive models; these are making it difficult for them to be effective in the real-world classroom. To tackle these issues this study, A Liquid Time Constant (LTC) network that is a unified Long Short-Term Memory (LSTM) network is introduced. Model for estimating course completion time and predicting dropping out. Unlike conventional models, the proposed approach introduces an adaptive liquid time constant into the LSTM framework to achieve an adaptive time constant. The network is to simulate continuous time dynamics and accommodate statistically irregular time periods between student-speech by using the network learning activities. The model can include the long-term temporal dependencies and evolving patterns in this design. More accurately describe behavioural patterns. In the data preprocessing step categorical attributes are converted into Label encoding is used for numerical representations, and min–max for numerical features normalization. The following are done to ensure consistent representation of features, stability of the model, and support: Convergence speed during training. The performance of the proposed LTC-LSTM model was tested on the three datasets that we use for educational benchmarking are KDD Cup 2015, Anon, and Dropout. Experimental results demonstrate the model attained accuracies of 92.10%, 84.65% and 94.48% respectively. In addition, LTC-LSTM achieved lower prediction errors than the others and needed less training to achieve higher predictive performance compared to the common approaches such as the Image Convolutional and Bi-directional Temporal methods, this one appears to be more efficient Convolutional Network (IC-BTCN). Conclusively, the results demonstrate the ability of LTC-LSTM to be a strong mathematically sound and efficient modelling approach for students' learning trajectories. By providing accurate, the proposed framework can help institutions with predictions of course completion and dropout risk. Effectively and timely putting into practice interventions that enhance student engagement, retention, and academic achievement based on the data.