The study uses the MKR (Multi-task Knowledge-aware Recommendation) model, which integrates knowledge graphs and multilayer perceptron networks, and focuses on mining users’ high-level music preferences from music information and user behaviors. The study first introduces the three core parts of the MKR model: the recommendation task module, the knowledge learning module, and the cross-compression unit. Then, by optimizing the MKR model, the A-MKR (Attention-based MKR) model is proposed, which introduces the attention mechanism to further improve the model performance. In the application of music teaching, the study explores the keyword analysis of music education, digital teaching methods and their effect evaluation. The experimental results show that compared with traditional teaching, the digital teaching model performs better in both field assessment and ability growth, and the error rate is significantly reduced. For example, the digital group scored 1.3 and 9 points higher than the traditional group in field assessment and ability growth, respectively, and the error rate was reduced by 5.8%. This study proves the effectiveness and practical value of knowledge mapping based on the MKR model in the digital transformation of music teaching in colleges and universities. It provides new perspectives and methods for future music teaching.