When a task of a certain domain doesn't have enough labels and good features, traditional supervised learning methods usually behave poorly. Transfer learning addresses this problem, which transfers data and knowledge from a related domain to improve the learning performance of the target task. Sometimes, the related task and the target task have the same labels, but have different data distributions and heterogeneous features. In this paper, we propose a general heterogeneous transfer learning framework which combines linear kernel and graph regulation. Linear kernel is used to project the original data of both domains to a Reproducing Kernel Hilbert Space, in which both tasks have the same feature dimensions and close distance of data distributions. Graph regulation is designed to preserve geometric structure of data. We present the algorithms in both unsupervised and supervised way. Experiments on synthetic dataset and real dataset about user web-behavior and personality are performed, and the effectiveness of our method is demonstrated.