A robust periodicity mining method from incomplete and noisy observations based on relative entropy

Abstract

A long-standing problem in data mining is the accurate estimation of the period of a periodic process. To mine periodicity in an event, we have to face real-world challenges of inherently uncertain periodic behaviors and imperfect data collection. This paper presents a method for periodicity detection from incomplete and noisy observations. Firstly, we review previous works and point out its defects for specific periodic patterns. Secondly, a novel relative entropy based measure is proposed, and its validity is proved in a probabilistic framework. Experiments on simulated event sequences of various configurations and real-world human movement datasets show the effectiveness of our method. And performance results demonstrate that our scheme outperforms previous works generally, especially when the time interval ratio is quite small or large.