Discovering Multiple Time Lags of Temporal Dependencies from Fluctuating Events

As one of the key features of temporal dependency, time lag plays an important role in analyzing sequential data and predicting the developing trend. Huge number of temporal mining approaches have been successfully applied in many applications, like finance, environmental science and health-care. However, these approaches cannot effectively deal with a more realistic scenario, where more than one types of time lags are existed in sequences and all of them are fluctuating due to the inevitable noise. In this paper, we study the problem of discovering multiple time lags of temporal dependencies from event sequences considering the randomness property of the hidden time lags. We first present a parametric model as well as an EM-based solution for solving this problem. Then two approximate approaches are proposed for efficiently finding diverse types of time lags without significant loss of accuracy. Extensive empirical studies on both synthetic and real datasets demonstrate the efficiency and effectiveness of our proposed approaches.