Deep neural network algorithms have made breakthroughs in automatic labeling tasks，but it is still hard to solve the noise data problem in real music dataset． A music auto-tagging algorithm based on deep analysis on labels (DAL) which captures the potential relationship between audio features and music tags was proposed． The algorithm first extracts the audio features through a multi-level convolutional network，and then learn the vector repre- sentation of music tags to reduce the adverse effects of noise data． The experimental results show that the proposed algorithm can achieve higher mean area under receiver operating characteristic curve (AUROCC) and outperform other auto-tagging methods．