The clustering behavior of sliding windows
1–10 of 22 posts
Re: The clustering behavior of sliding windows
#2Interesting proofs, though I would love more of the examples that show the dataset and the unexpected results, as well as describing what may lead to these traps and how to avoid them
Re: The clustering behavior of sliding windows
#3Of further potential interest: This paper cites an earlier paper by Keogh and Lin with the provocative title "Clustering of time-series subsequences is meaningless", available online at https://www.cs.ucr.edu/~eamonn/meaningless.pdf
Re: The clustering behavior of sliding windows
#4Well, I thought this was going to be about home design and sliding windows.
Re: The clustering behavior of sliding windows
#5Interesting proofs, though I would love more of the examples that show the dataset and the unexpected results, as well as describing what may lead to these traps and how to avoid them
... if I get it right the whole idea of clustering sliding windows is wrong, the question of "what you should do instead?" is an interesting one.
I'd imagine two answers are: (1) for time series which are somewhat periodic you might cut out individual days or weeks and try to cluster them for each other, (2) for time series which are intermittent you might create some definition of an "event" (an earthquake, or a particle passing through a detector) and then cluster events and maybe (3) for something episodic such as "heart rate during a workout" you would cluster episodes.
Re: The clustering behavior of sliding windows
#6You'd think this would be more well known. In fourier analysis or other frequency/wavelet decomposition techniques, the data you get out depends a ton on the windowing function that you use.
For example:
https://en.wikipedia.org/wiki/Sinc_function
https://en.wikipedia.org/wiki/Window_function#Examples_of_wi...
Re: The clustering behavior of sliding windows
#7very interesting, but I wonder if it would hold for clustering with other distance metrics such as dynamic time warping (DTW).
Re: The clustering behavior of sliding windows
#8The problematic projection into vectors is done using a window size of a constant number of samples? Isn't that very weird?
Unless the samples are nicely periodic, that makes the contents of a vector, and the position of a value in a vector, dependent only on the window length and the order of time values, and not the time values themselves. Since the ordering in the vector, and thus the meaning of that dimension in vector space, is likely to be effectively random, why would we expect clustering in that vector space to mean anything? It's a weird thing to do on the face of it.
Re: The clustering behavior of sliding windows
#9The problematic projection into vectors is done using a window size of a constant number of samples ? Isn't that very weird? Unless the samples are nicely periodic, that makes the contents of a vector, and the position of a value in a vector, dependent only on the window length and the order of time values, and not the time values themselves. Since the ordering in the vector, and thus the meaning of that dimension in…
My understanding is that people would mostly try this approach when the samples are periodic/regular.
Re: The clustering behavior of sliding windows
#10The problematic projection into vectors is done using a window size of a constant number of samples ? Isn't that very weird? Unless the samples are nicely periodic, that makes the contents of a vector, and the position of a value in a vector, dependent only on the window length and the order of time values, and not the time values themselves. Since the ordering in the vector, and thus the meaning of that dimension in…
Imagine you know that there are some repeating patterns in your dataset, e.g. heartbeats in an EKG. If you take two windows that happen to contain a heartbeat at the same position, the corresponding vectors will be close, and if the positions are different, the vectors should become more dissimilar the more the alignment is off.
Then if you create a number of clusters equal to the window size, you might expect that each cluster will correspond to one of the possible positions of the heartbeat within the window. But somehow that's not what happens...