The segmentation of Time Series

In 2003, Keogh, E., Chu, S., Hart, D., Pazzani and M. Segmenting had written Time Series: A Survey and Novel Approach. in Data Mining in Time Series Databases published by World Scientific Publishing Company. The article talks about how to convert time series data with high frequency noise into some linear regression segmentations. If the original time series data has N timestamps, the linear regression result has K segments(K is far smaller than N), the storage and transmission space required of the data would be minimized.

The article summarized three typical method of segmenting time series: sliding window, top-down and bottom-up.

• Sliding window method is to consider the time series as a data stream, and segment the series from early time to later time data. Every time the method steps forward a timestamp, adds the current value to exist window, or closes the current window and establishes a new window for incoming data. The method could be used as online processing method.
• Top-down method is to recursively cut the current data segment into 2 sub segments, until some error threshold is reached.
• Bottom-up method is to construct a lot of very small pieces segments, and iteratively combine the two segments with minimal error according to original time series.

The results show that top-down and bottom-up methods approximate the original time series better than sliding window due to the sliding window method is lack of global view about data. But top-down and bottom-up cannot be used to perform online time series segmentation.

The article provide a new algorithm SWAB based on bottom-up and sliding window which provide a semi-global view of the time series data. The algorithm establishes a buffer for several sliding windows, and uses bottom-up method to combine them. The SWAB even achieves a better approximation than the bottom-up method.