The main goal of transfer learning is to reuse related domain data to learn models for the target domain. In existing instance-transfer learning algorithms, the relevance of instances is estimated mainly according to small amount of labeled instances, and the generalization ability of these algorithms needs to be improved. To make the relevance estimation more reliable, we propose to use unlabeled target domain instances as additional training data. These instances would serve as new domain knowledge sources to help determining the relevance of related domain instances. Under the universal framework of boosting, we introduce local smoothness regularizer, and obtain new empirical loss function, where unlabeled instances are included. Gradient decent method is used to iteratively optimize the loss function, and we finally obtain a new instance-transfer learning algorithm. Experiment results on text datasets show that the new algorithm outperforms competitive algorithms.