Abstract: Aiming at the low transfer learning performance in cross-media heterogeneous transfer learning oriented to semi-paired problem, a novel hybrid Laplacian eigenmap based on balanced heterogeneous distance in cross-media heterogeneous transfer learning is proposed in this paper. The proposed method takes full advantage of the abundant semantic information in the massive unpaired samples and a few paired samples to learn the mapping matrices from the original feature spaces of different media domains to the latent common feature space. Moreover, by constructing mixed graph Laplacian matrix in cross-media transfer learning, it not only maintains the manifold structure of the samples from the same media domain, but also maintains the manifold structure of the samples from different media domains, which promotes the model performance in the target media domain. Extensive experiments are conducted on two common datasets: the NUS-WIDE and LabelMe. The experimental results show that the accuracy and robustness of the model can be increased by using a large number of unpaired data and a few of paired data simultaneously.