Reconstructing A Framework for Predicting University Tuition Payment Trends Using Temporal Fusion Transformer

Abstract

To improve the prediction accuracy and interpretability of complex university tuition payment time series, this paper proposes a reconstructed forecasting framework based on the Temporal Fusion Transformer. The framework encodes multi-source heterogeneous data, including static and dynamic variables such as student majors, economic status, scholarship disbursement plans, and real-time payment reminders, into a low-dimensional representation space. A sigmoid gated network dynamically weights informative variables and suppresses redundant features. LSTM modules and a specialized three-head attention mechanism for time patterns, group synchronization, and policy response are jointly used to model long-term dependencies and capture nonlinear payment fluctuations. Interpretable multi-step forecasts are generated by combining sequence outputs with attention weights. Experimental results show that the model achieves RMSE ≤0.08 and MAE ≤0.07 during peak payment periods, with R2 above 0.9 in five disturbance scenarios, demonstrating strong dynamic adaptability. Interpretability analysis yields a Pearson correlation coefficient of at least 0.68 and an information gain ratio of at least 0.7, confirming robust feature attribution and discriminative capability.