Abstract: For causality analysis between dynamic time series, a topology recognition method based on recursive sparse online learning is proposed. The sequence sparse topology identification standard was minimized by composite target iteration to promote sparse update. The tracking performance of input variability was greatly improved by the estimation criterion based on the least square method, and the historical samples were considered in the instantaneous loss function. The logarithmic regret limit was obtained through static regret analysis. The numerical results of real data and synthetic data show the effectiveness of the proposed algorithm in static and dynamic scenes.