Personalized recommendation system for network popular music based on audio feature extraction

Abstract

With the rise of popular music on the Internet, users need a powerful recommendation system to increase their subscription volume for music consumption. Research was done using deep learning techniques in streaming media platforms to convert audio signals into Mel spectrograms for extracting audio features. Under the computation of the Mel filter, feature extraction of audio files can be input into a binary network with platform user behavior, thereby constructing a music recommendation system. Finally, the relationship learning module is utilized to integrate user behavior and audio features, combined with a music encoder to complete personalized audio feature processing, thereby achieving the construction of a music personalized recommendation system. Through dataset testing, it was found that the initial dataset had a sparsity ratio of 0.14% and a recall rate of 30.16%. In the partitioning and data analysis of the Subset dataset, it was found that the user group recommendation algorithm had the highest recommendation accuracy of 24% for Subset100 subset, and the highest recall rate was 33.5% when the number of recommendations was 50, indicating the necessity of data recommendation. In the testing of the personalized recommendation system, the training accuracy of the system was as high as 93%. Finally, different recommendation methods were compared between the Subset500 dataset and the Mirex dataset, and the research method based on audio features had the highest recommendation accuracy of 97.5%, recall rate and F1 value of 42.1% and 58.81%, respectively, and a running time of 3.2 seconds, thus proving the excellent performance. The personalized recommendation system not only improves the predicted rating results of music in practical applications but also provides reliable technology for the future dissemination and development of popular music on the Internet.