A Hybrid Probabilistic Local Robust Expectile Graph Attention Network for Student Academic Performance Prediction with Mental Health Indicators

Abstract

Student academic performance prediction has become important one for educational institutions with support systems. Student academic performance is influenced by both cognitive abilities and psychological well-being. Mental health affects the student academic performance and impacts the physical and mental development, social interaction and future careers. Different machine learning and deep learning techniques are employed to compute relationship between mental health conditions and academic performance. But, the inaccurate prediction was carried out by existing methods as mental health changes frequently over time.
In addition, the existing methods failed to identify the most relevant psychological indicators. In order to address these issues, a new intelligent hybrid predictive model called Probabilistic Local Robust Expectile Graph Attention Network (PLRExGAN) is introduced. The designed PLRExGAN model integrates the mental health indicators with academic and behavioral features to accurately predict student performance. PLRExGAN Model uses one input layer, three hidden layers and one output layer for student performance analysis. PLRExGAN Model uses deep learning concepts for performing four processes, namely data pre-processing, feature selection, classification and fine-tuning. Hybrid Graph Attention Network uses the attention mechanism for finding the student academic performance prediction. PLRExGAN Model considers the number of student data points as input at input layer. Probabilistic Local Outlier Factor Data Handling Process is employed in hidden layer 1 for identifying the outlier data points through evaluating the density of data points in their local neighborhoods. Then, Robust Expectile Regressive Feature Selection process is employed in hidden layer 2 to select the relevant features based on mental health. With the selected features, an Extended Jaro–Winkler Similarity Classification is carried out in the hidden layer 3 based on mental health (i.e., memory capability) for performing student academic performance analysis. Then, meta-heuristic elephant herd optimization is carried out for performing hyperparameter fine-tuning to attain the accurate student academic performance results with minimum error.
Finally, output layer displays the student academic performance results. Experimental analysis of PLRExGAN Model is carried out with the performance metrics, namely student academic performance prediction accuracy, student academic performance prediction time, precision, recall, f1-score, specificity, confusion matrix and ROC-AUC with respect to number of student data points.