Efficient Mining of Event Periodicity in Data Series

This paper investigates the problem of efficiently discovering periodicity of a certain event in data series. To that end, the current work argues firstly that the periodicity of an event in data series may be formalized as the distribution period, the structure period, or the both. Along this line, a partition method, $$\pi (n)$$ , is proposed to divide the data series into length-equal and position-continuous segments. Based on the results of implementing $$\pi (n)$$ on a data series, we propose two new concepts of distribution periodicity and structure periodicity. Then, a cross-entropy-based method, namely CEPD, is proposed to mine the periodicity of data series. The experimental results show that CEPD can be used to mine feasible event periodicity in data series, especially, with very low level of time consumption and high capability of noise resilience.