Software defect prediction can help developers to optimize the distribution of test resources by predicting whether or not a software module is defect-prone. Most defect prediction researches focus on within-project defect prediction which needs sufficient training data from the same project. However, in real software development, a project which needs defect prediction is always new or without any historical data. Therefore cross-project defect prediction becomes a hot topic which uses training data from several projects and performs prediction on another one. The main research challenges in cross-project defect prediction are the variety of distribution from source project to target project and class imbalance problem among datasets. Inspired by search based software engineering, this paper proposes a search based semi-supervised ensemble learning approach S³EL. By adjusting the ratio of distribution in training dataset,several Naïve Bayes classifiers are built as the base learners, then a small amount of labeled target instances and genetic algorithm are used to combine these base classifiers as a final prediction model. S³EL is compared with other up-to-date classical cross-project defect prediction approaches (such as Burak filter, Peters filter, TCA+, CODEP and HYDRA) on AEEEM and Promise dataset. Final results show that S³EL has the best prediction performance in most cases under the F1 measure.