Data stream classification is one of important research tasks in the field of data mining. Most existing data stream classification algorithms require the labeled data for training. However, there are few labeled data in data streams in real applications. To solve this problem, the labeled data can be obtained by manual labeling, but it is very expensive and time consuming. Considering the unlabeled data are huge and full of information, a data stream ensemble classification algorithm based on Tri-training for labeled and unlabeled data is proposed in this paper. The proposed algorithm divides data stream into chunks by sliding windows and trains base classifiers with Tri-training on the first coming k chunks with labeled and unlabeled data. Then the classifiers are iteratively updated by weighted voting until all unlabeled data are labeled. Meanwhile, the k+1 data chunk is predicted by using the ensemble model of k Tri-training classifiers and the classifier with higher classification error is discarded, which reconstructs a new classifier on current data chunk to update the model. Experiments on 10 UCI data sets show that the proposed algorithm can significantly improve the class ification accuracy of data stream even with 80% unlabeled data in comparison with traditional algorithms.