Abstract: On the basis of data mining,a new self-learning fuzzy method is developed to model and predict chaotic time series,by means of interest measure and improved gradient descent method.The proposed method can not only identify the fuzzy model,update its parameters and determine the optimal output fuzzy sets simultaneously,but also resolve the conflicts between convergence speed and oscillation existing in gradient descent method.Simulation results show the effectiveness and accuracy of the proposed method.It can identify the system characteristics quite well and provide a new way to predict the chaotic time series.