Abstract: Online learning is the effective way to solve the sample's non-recurrence in data streams classification, and how to deal with the problem of sample deficiency is the critical point for improving online learning efficiency. According to the mean square error decomposition theory of the model's parameter estimation and the idea of cluster, the new samples are constructed by linear synthesis with the class center and the sample, which can improve the distribution information of sample and reduce the lower bound of parameter value. The online incremental learning is executed and the class center point is continuously updated. Through theory analysis and simulation experiment, it is suggested that the provided schema is feasible and has superiority over other algorithm.

Key words: online learning, data streams classification, cluster, incremental learning