Don’t make this mistake when clustering time series data

That title seems awfully clickbaity to me.

How about something with some meat on its bones, like "Sliding windows on time series subsequences yields meaningless clusters"?

Keogh has produced a bunch of interesting work. Toward Parameter Free Data Mining is a gem: https://www.cs.ucr.edu/~eamonn/SIGKDD_2004_long.pdf Of course SAX is also super interesting: https://www.cs.ucr.edu/~eamonn/SAX.pdf

Just to give some context, the paper referenced by the post[0] is a bit old. Eamonn Keogh is (to my knowledge) the current head of the research group that developed the Matrix Profile, which purports to be useful in clustering[1]. Actually on their main introduction to the MP[2] they use a chart to explain motif discovery and chains that is very similar to the one on section 6 of [0].

To be fair, I don’t remember any of the papers of the group explicitly discussing subsequence clustering, but one hopes they have advanced on that front.

[0] https://www.cs.ucr.edu/~eamonn/meaningless.pdf

[1] https://www.cs.ucr.edu/~eamonn/MatrixProfile.html

[2] https://www.cs.ucr.edu/~eamonn/Matrix_Profile_Tutorial_Part1...

My intuition is that clustering subsequences of time series data should easily find diurnal/seasonal patterns - am I thinking about this wrong?