Abstract
Course recommendation systems serve as a critical component of online education platforms, playing a vital role in enhancing learning efficiency and personalized experiences. However, existing recommendation approaches, including recent sequential models such as BERT4Rec and LightSANs, primarily concentrate on temporal-domain modeling of user behaviors while neglecting the potential of frequency-domain analysis. This leads to incomplete characterization of user behavior patterns, particularly presenting challenges in capturing stable long-term interests from sparse and noisy interaction data. To address these limitations, this study proposes a novel hybrid attention network for Massive Open Online Courses Course recommendation, designed to jointly model both frequency-domain and temporal-domain features. The model employs Fast Fourier Transform to extract frequency-domain characteristics from user behavior sequences while utilizing a self-attention mechanism to capture temporal dynamics, thereby enabling collaborative modeling of dual-domain features. Experimental results on the public MooCCube dataset demonstrate that the proposed model achieves Hit Ratio@10, MRR@10, and NDCG@10 scores of 0.4534, 0.2018, and 0.2618, respectively, outperforming current mainstream recommendation algorithms. Ablation studies further validate the effectiveness of dual-domain fusion, with approximately 10% and 5% performance improvements in NDCG@10 and Hit@10 compared to single-domain approaches. This research provides a novel technical pathway for overcoming performance bottlenecks in personalized course recommendation.