A ROBUST PSO-LSTM FRAMEWORK FOR PREDICTING ACADEMIC PERFORMANCE WITH EXPERIMENTAL VALIDATION

Accurately predicting students' academic performance is crucial for identifying students with learning difficulties and tailoring educational interventions effectively. This study proposed a PSO-LSTM framework that combines Particle Swarm Optimization (PSO) with Long Short-Term Memory (LSTM) to improve prediction accuracy by optimizing LSTM hyperparameters. The model was evaluated on a dataset of 722 students from the Computer Science program at Wenzhou Business College, covering six academic years (2018–2024). Two experimental scenarios were designed: semester-based GPA prediction and year-based CGPA prediction. The results indicated that PSO-LSTM outperformed traditional machine learning models (Decision Tree, Random Forest, SVM, KNN) and standard deep learning models (ANN, LSTM) in both experimental settings. In predicting the cumulative grade point average (CGPA), the model's accuracy increased from 87.59% in DSTU5 to 95.17% in DSTU7, and further to 96.55% after applying SMOTE. This method effectively alleviates bias caused by class imbalance, particularly by improving recall for identifying the minority (at-risk) class, thereby significantly enhancing fairness and robustness (e.g., DSTU5 increased by 8.0%). These findings highlight PSO-LSTM as a highly effective framework for predicting academic performance, with strong potential for early intervention strategies and personalized learning recommendations.