[Objective] Aimed at less labeled data and more unlabeled samples in micro-blog sentiment classification, a novel method is proposed. [Methods] Active learning is introduced into co-training, the method selects the most valuable ones from low confidence samples, then labels and adds them into training dataset, trains classifiers again. [Results] Experimental results show that classifiers have better performance in this way, and the accuracy is improved obviously. Especially when labeled data reaches 40%, the accuracy increases by about 5%. [Limitations] In the collaborative process, random feature subspace generation can not build two strong classifiers, so hypothesis are not fulfilled. [Conclusions] This method solves the defects of co-training after introducing active learning; the performance and accuracy of classifiers are enhanced.