The continuous development of cloud computing and deep learning technologies has opened up new possibilities for the efficient dissemination of digital music. This paper expands the relevance of query content through data preprocessing and pseudo-correlation feedback techniques. By combining a query likelihood model with a deep ranking learning method based on Pointwise, user preference-based music ranking is achieved. A deep learning-based composite model (ResGRU) is established to extract user implicit behavior data and preferred audio features, enabling precise recommendations for queried music. Research shows that the comprehensive ranking method achieves metric values greater than 0.800 at music popularity levels of 30%, 60%, 90%, and 100%. The conversion rates for both the same network and different networks exceed 87% and 85%, respectively. The ResGRU composite model outperforms five comparison models across six metrics in two datasets, and the best music recommendation results are achieved when the optimal convolutional kernel size is set to 6.
