Abstract: Due to the different distribution of features between the source and target domains in a multi-label transfer learning problem, source domain data cannot exert any effect. To resolve this problem, here we propose novel multi-label transfer learning via the maximum mean discrepancy. The proposed algorithm decomposes a relational matrix to learn a common subspace. Furthermore, we incorporate the empirical maximum mean discrepancy into the objective function of matrix factorization to minimize the probability distance between different domains. Experimental results from multi-label classification demonstrate that the proposed approach achieves better performance than other similar algorithms in terms of accuracy and efficiency.