Review

The problem of finding similar time sequences has vexed us for quite a while. The central issue is that similarity lies in the eye of the beholder -- any formal definition of similarity is likely to be suitable only for certain very specific contexts. For the most part, the approach taken by various authors has been to permit "transformations" (from some allowed class, such as scaling, shifting, time-warping, and so on), and then performing a distance measurement. While a rich enough class of transformations can accomplish a great deal, permitting a large class of transformations means that an expensive search is required to find the sequence of transformations that minimizes the distance between two given time series.

I like this paper because it proposes a nice appealing definition of a landmark, and proposes a notion of landmark similarity that is very easy to work with in terms of tractability, yet can do a decent job of capturing similarity in many contexts.