A model that incorporates semantic information can alleviate the sparse data and cold start problems in a music recommendation system. Current models learn the global relations between users and music by minimizing prediction score error. However

they ignore the fine-grained differences between the implicit features of users and music. Current models also extract semantic features by unsupervised learning that is irrelevant to recommending

leading to inaccurate recommendations. We propose a metric-ranking-learning recommendation model that incorporates content representation (CAMRL). This model is a probabilistic graphical model that optimizes a personalized ranking and explores user music preferences through metric learning at both global and fine-grained levels. To solve the cold start recommendation problem

a supervised learning strategy in relation to the recommendation task is proposed to train the model of content semantic feature exaction. The results of trials using the KKBOX and MIGU datasets show that the proposed model significantly improves cold start music recommendations when compared with other algorithms; it is also more robust when using sparse datasets.