：Experts can provide authoritative answers for community question answer (CQA). Efficient and accurate expert discovery can help improve the service quality of CQA. The expert discovery accuracy of the supervised learning model is reduced by the noise label data existing in the community user data as well as the unbalanced classification data due to the small number of experts. A semi-supervised expert discovery method based on feature perturbation is proposed to solve the mentioned problems. In this method, a feature perturbation strategy for unlabeled data is constructed, using the Sharpening algorithm to label the pseudo-label of unlabeled data. Based on the ADASYN algorithm, expert sample data is expanded by constructing neighbor samples of expert users to alleviate the imbalance of classification data. A joint loss function is constructed, which trains the classifier by both the labeled and pseudo-labeled data to enhance the generalization performance of the method. The experimental results show that this method is superior to the existing models and methods in several evaluation indexes.