The invention provides an online active learning method suitable for a label-free unbalanced data stream, which comprises the following steps that: the time sequence of the label-free data stream is input into a linear classifier for prediction, and the category of the data stream has a highly unbalanced problem, the number of positive samples is rare; According to the proposed asymmetric access strategy, the linear classifier dynamically determines a sample needing to be labeled for the unbalanced data; According to the proposed asymmetric updating strategy, the linear classifier updates thelinear classifier by using the labeled data of the error prediction, and the learning efficiency is improved by using the second-order information of the sample; According to the online active learning method suitable for the label-free unbalanced data flow, a new asymmetric strategy is provided by using second-order information of samples; The asymmetric strategy simultaneously considers the marking of the samples and the updating of the model, so that the problem of class imbalance of the samples can be better solved, and the classification performance of the active learning model based on the streaming data is improved.