Analyzing Periodically Occurring Patterns in Time Series

Abstract In this paper, we address the problem of identifying periodically occurring patterns in a time series. The domain of data-center management is the primary focus. Data here comprises of request latencies, resource utilization of servers, data center workload etc. to name a few. Although periodicity detection has been researched, the past work does not address the challenges presented by such data-sets. The major challenges include time scaling, time shifting, amplitude scaling, amplitude shifting and noise. We propose an innovative solution to cater to the new challenges. In this paper, we address the problem of identifying the shape of the periodically occurring pattern and the timeseries regions which exhibit periodic behavior. We also present a crisp definition of a periodic pattern in the face of such challenges. In addition, we present experimental evaluation of the proposed technique on various data-sets to evaluate its robustness