Hybrid Model of Fuzzy Logic and Recurrent Neural Network for Dynamic Student Achievement Prediction

Abstract

In this study, a dynamic prediction model of student academic achievement was developed by integrating fuzzy logic and recurrent neural networks (RNN). The dataset consisted of 235 undergraduate students enrolled in the course Learning Strategy and Behavior Analysis during the spring semester of 2024, with data covering grades, online learning behaviors, and interaction records. The fuzzy rule system was constructed to transform uncertain behavioral variables, such as “learning engagement” and “task completion stability,” into interpretable linguistic categories. These were then combined with an RNN structure to capture temporal dependencies in students’ grade trajectories. Model training was conducted with 120 epochs, a batch size of 32, and a learning rate of 0.003. Results demonstrated high predictive accuracy, with mean squared error (MSE) ranging between 0.022 and 0.033 and coefficient of determination (R²) values above 0.94 across validation samples. Personalized interventions derived from prediction outputs, such as increasing video learning time or enhancing peer discussion, led to measurable improvements, with error reductions of up to 0.039 in specific cases.