Patterns Discovery Based on Time-Series Decomposition

Abstract

Complete or partial periodicity search in time-series databases is an interesting data mining problem. Most previous studies on finding periodic or partial periodic patterns focused on data structures and computing issues. Analysis of long-term or short-term trends over different time windows is a great interest. This paper presents a new approach to discovery of periodic patterns from time-series with trends based on time-series decomposition. First, we decompose time series into three components, seasonal, trend and noise. Second, with an existing partial periodicity search algorithm, we search either partial periodic patterns from trends without seasonal component or partial periodic patterns for seasonal components. Different patterns from any combination of the three decomposed time-series can be found using this approach. Examples show that our approach is more flexible and suitable to mine periodic patterns from time-series with trends than the previous reported methods.