Emotion-driven music recommendation system based on fully convolutional recurrent attention networks and collaborative filtering

With the rapid growth of online music platforms, personalized music recommendation has become a critical task. However, existing methods often struggle to capture the intricate emotional and contextual characteristics of music, resulting in suboptimal user experiences. To address these limitations, we propose a hybrid recommendation system that integrates collaborative filtering, music attribute modeling, and a FCRA. The FCRA extracts key emotional features, including tempo, pitch variation, spectral contrast, harmonic progression, and rhythm patterns, which correspond to affective states such as happiness, sadness, calmness, and intensity. The hybrid design leverages both user interaction history and music content to enhance recommendation accuracy. Experimental results on the MSD and Last.Fm 360k datasets show that our model achieves a Precision@10 of 0.902 and 0.885, Recall@10 of 0.832 and 0.815, and NDCG@10 of 0.855 and 0.84, respectively, outperforming all baseline methods. Furthermore, the system performs robustly in cold-start scenarios, highlighting its adaptability across different recommendation contexts. This research provides a novel approach to addressing the challenges of emotion-driven music recommendation, contributing to improved user satisfaction and better alignment with individual preferences and moods.