Predicting academic performance for students' university: case study from Saint Cloud State University.

Area of Science:

• Educational Data Mining
• Machine Learning in Education
• Academic Performance Prediction

Background:

• Student performance prediction is crucial for academic success and early intervention.
• Identifying at-risk students aids in preventing failure, major changes, or dropout.
• Existing models often lack the sophistication to capture complex academic and demographic influences.

Purpose of the Study:

• To enhance student Grade Point Average (GPA) prediction accuracy.
• To implement and evaluate a Long Short-Term Memory (LSTM) model for performance forecasting.
• To identify key features influencing student academic outcomes.

Main Methods:

• Utilized a comprehensive dataset of 29,455 students from Saint Cloud State University (2016-2024).
• Preprocessed data by handling missing values, encoding variables, and normalizing features.
• Employed a permutation-based method for feature importance and fine-tuned LSTM hyperparameters via experimental validation.

Main Results:

• The LSTM model demonstrated superior performance over traditional models (LR, KNN, DT, RF, SVR) and other deep learning models (RNN, CNN).
• Achieved a 99% R² score, with a Mean Absolute Percentage Error (MAPE) of 9.54%, Mean Absolute Error (MAE) of 0.0059, and Root Mean Square Error (RMSE) of 0.0001.
• Experiments were successfully conducted at both college and department levels, validating the model's generalizability.

Conclusions:

• The proposed LSTM model offers a powerful and accurate approach for predicting student academic performance.
• Feature importance analysis provides insights into critical factors affecting student success.
• This predictive capability can significantly assist educational institutions in student support and intervention strategies.