Online music personalized recommendation algorithm based on KKBox large-scale real-world dataset and integrated behavior sequence transformer

To address the user selection overload and decision making difficulties caused by the explosive growth of online music platforms, this paper proposes and validates a deep learning recommendation model that incorporates the Behavior Sequence Transformer, aiming to significantly improve music recommendation accuracy, diversity, and novelty. We use the BST Behavior Sequence Transformer to perform multi head self attention modeling on user listening sequences and compare it with a LightGBM boosted tree and logistic regression baseline model. Experiments use AUC, accuracy, and RMSE as core metrics, and use an 80/20 train-test split and early stopping to control overfitting. The BST model achieved an AUC of 0.7662 and an accuracy of 0.7551 on the test set, significantly outperforming LightGBM and logistic regression . Feature importance analysis shows that user genre cross features, contextual entry points, and sequence position encoding contribute most to the model. The recommendation algorithm based on the Behavior Sequence Transformer effectively captures the dynamic evolution of user interests, alleviates information silos, and possesses good scalability and engineering potential. Future research into multimodal fusion and reinforcement learning can further enhance diversity and interpretability. **************** ACKNOWLEDGEMENTS**************** This work was supported by Jiangxi Provincial College Students' Innovation and Entrepreneurship Program (S202510421047).