Time series data accounts for a large fraction of the data stored in financial, medical and scientific databases. Recently there has been an explosion of interest in data mining time series, with researchers attempting to index, cluster, classify and mine association rules from increasing massive sources of data. In this tutorial, I will give a complete overview of the state of the art in time series data mining. I will explain why the unique structure of time series presents difficulties for classic data mining algorithms, and how this structure can potentially be leveraged off.
As with any computer science problem, representation is the key. I will therefore review every representation proposed for mining time series, including wavelets, Fourier transforms, singular value decomposition, piecewise polynomial models and symbolic mappings. I will present detailed and extensive empirical comparisons of how these representations perform on a variety of data mining tasks.
Dr Keogh obtained his Ph.D. from the University of California, Irvine in 2001. His thesis is entitled "Similarity Search in Massive Time series Databases". Dr Keogh has published more than a dozen papers on mining time series data, and has won three best paper awards for his work (including SIGMOD 2001). His research interests include Machine Learning, Data Mining, Multimedia Indexing and Information Retrieval. Beginning fall 2001, he will be on the faculty at University of California, Riverside.
Questions and comments to:
clifton@computer.org.
Last modified:
September 22, 2001.