Abstract: Transfer learning algorithms focus on reusing related domain data to help solving learning tasks in the target domain. In this paper, we study the problem of inductive transfer learning. Most of the existing algorithms in inductive transfer learning might suffer from the problem of sample selection bias when the number of target domain data is too small. To address this problem, we propose to utilize domain similarity in a new approach. Through detailed discussion about the similarity of related domains, we define the concept of weak domain similarity. Using this concept to give additional constraints on the target classifiers, we develop a simple but effective approach to leverage the useful knowledge from the related domain, so that related domain data can be directly used in the training process. In this way, we are able to make the target classifier less sensitive to the small amount of target training data. Furthermore, we show that a modified SMO method can be applied to optimize the objective function in the algorithm effectively. The new algorithm is referred to as TrSVM, and can be seen as extension of support vector machines for transfer learning. Experiment results on extensive datasets show that TrSVM outperforms support vector machines and the state-of-the-art TrAdaBoost algorithm, and demonstrate the effectiveness of our algorithm.