Abstract:Existing classification algorithms for data stream are mainly based on supervised learning,while manual labeling instances arriving continuously at a high speed requires much effort.A low-cost learning algorithm for stream data classification named 2SDC is proposed to solve the problem mentioned above.With few labeled instances and a large number of unlabeled instances,2SDC trains the classification model and then updates it.The proposed algorithm can also detect the potential concept drift of the data stream and adjust the classification model to the current concept.Experimental results show that the accuracy of 2SDC is comparable to that of state-of-the-art supervised algorithm.