Time series data is of interest to most science and engineering disciplines and analysis techniques have been developed for hundreds of years. There have, however, in recent years been new developments in data mining techniques, such as frequent pattern mining, that take a different perspective of data. Traditional techniques were not meant for such pattern-oriented approaches. There is, as a result, a significant need for research that extends traditional time-series analysis, in particular clustering, to the requirements of the new data mining algorithms.
Time series clustering is an important component in the application of data mining techniques to time series data (Roddick, Spiliopoulou, 2002) and is founded on the following research areas:
Data Mining: Besides the traditional topics of classification and clustering, data mining addresses new goals, such as frequent pattern mining, association rule mining, outlier analysis, and data exploration (Tan, Steinbach, and Kumar 2006).
Time Series Data: Traditional goals include forecasting, trend analysis, pattern recognition, filter design, compression, Fourier analysis, and chaotic time series analysis. More recently frequent pattern techniques, indexing, clustering, classification, and outlier analysis have gained in importance.
Clustering: Data partitioning techniques such as k-means have the goal of identifying objects that are representative of the entire data set. Density-based clustering techniques rather focus on a description of clusters, and some algorithms identify the most common object. Hierarchical techniques define clusters at multiple levels of granularity. A survey of clustering that includes its application to time series data is provided in (Gan, Ma, and Wu, 2007).
Data Streams: Many applications, such as communication networks, produce a stream of data (Muthukrishnan, 2003). For real-valued attributes such a stream is amenable to time series data mining techniques.
Time series clustering draws from all of these areas. It builds on a wide range of clustering techniques that have been developed for other data, and adapts them while critically assessing their limitations in the time series setting.Top
Main Thrust Of The Chapter
Many specialized tasks have been defined on time series data. This chapter addresses one of the most universal data mining tasks, clustering, and highlights the special aspects of applying clustering to time series data. Clustering techniques overlap with frequent pattern mining techniques, since both try to identify typical representatives.