ENHANCING STUDENT SUCCESS BY DEVELOPING HYBRID NNRW-LSTM PREDICTIVE MULTI-TASK PERFORMANCE PREDICTION IN HIGHER EDUCATION INSTITUTIONS

Abstract

Student success in college is influenced by both academic and behavioral wellbeing. In this paper, a novel hybrid architecture called NNRW-LSTM (Neural Network Random Weights. Long Short-Term Memory) is proposed for multi-task predicting academic and behavioral risk among college students. The model capitalizes on a large-scale data set including academic data, demographic variables, and behavioral indicators and combines static variables and semester-based time-series data into its framework. With a hybrid feature selection pipeline based on Mutual Information, Correlation Analysis, Recursive Feature Elimination, and PCA achieving high-relevance input variables and input dimensionality reduction, the proposed model predicts Cumulative GPA and behavioral warning percentages using a dual-branch neural network. Results show better performance relative to single-task approaches with Mean Absolute Errors 0.7415 and 0.3163 for GPA and behavioral warnings, respectively. The proposed method offers a scalable and interpretable solution for early risk detection and student-focused intervention in college settings.