A Time-Frequency Enhanced Attention Network for Student Performance Prediction in Online Learning

Abstract

Accurately predicting student performance remains challenging due to evolving trends, periodic fluctuations, and the inherent sparsity of behavioral sequences in online learning platforms. Existing approaches primarily focus on single-modality modeling and lack collaborative integration of temporal dependencies and frequency-domain characteristics, limiting their ability to capture correlations across critical time intervals. To address this limitation, we propose a Time-Frequency Enhanced Attention Network (TF-EAN), which jointly models the temporal dependency structure and spectral patterns of student learning behavior sequences. In the temporal branch, multi-head self-attention with sinusoidal positional encoding is employed to capture long-range dependencies. In the frequency branch, Fast Fourier Transform (FFT) is applied to extract amplitude and phase components for modeling periodic patterns. Experiments conducted on the Open University Learning Analytics Dataset (OULAD) demonstrate that TF-EAN outperforms traditional machine learning methods and several deep learning baselines, achieving an accuracy of 98.74% and 99.15% on two public datasets, respectively, thereby validating the effectiveness of time-frequency collaborative modeling.