Mining Asynchronous Partial Periodic Pattern from Multi-event Time Series Database with Unknown Periods

摘要

Current research on periodic pattern mining focuses on mining asynchronous but simple single-evnet patterns. However, in real-life situation, there are more than one events happening at one time. In this paper, we propose a thoroughly-new algorithm to really solve the problem we would experience in livelihood. Three parameters min_rep, max_dis and total_rep are employed to specify the constraints a significant pattern must satisfy. Min_rep specify the minimum number of repetitions that is required within each segment of non-disrupted pattern occurrences, max_dis specify the maximum allowed disturbance between any two successive valid segments, and total_rep claims the minimum overall repetitions that is needed within a valid subsequence.Our algorithm is composed of two individual parts. One is called 1-pattern mining, and the other is called pattern growth. In the first part, a sliding window method is devised to find the entire potential valid segment matched by 1-patterns. The second part, we make use of the concept of BFS to gain valid subsequences in the overall time series dataset. Finally in experiments, our algorithm is shown efficient and stable with scale-up dataset size.