As, I am working with feature selection for dementia patients using HFD, I want to select the most suitable K max to my study. How to select the K max.
I attached four articles as you can see below. These articles propose certain techniques on how to choose the Kmax value. You can try their techniques for your problem.
These articles suggest k from 6 to 60. I think, there is a reason to set some fixed Kmax, say, 6, and work with it. If this feature will be useful for diagnostics, it will work in quite wide range of Kmax. If you will find it useful (after some feature selection methods applying), you can go to "fine tuning".
You can calculated the Fractal Dimension with K=1:Kmax and plot fractal Dimension Vs K . this figure will reach a plateau . In this figure the fractal dimension doesn't vary significantly for values greater than a specific K.That value is a suitable value for k in calculation of fractal dimension
In case of Higuchi's Fractal dimension calculation, we can take only that value as K_max for which a perfect fall down of the data points on the linear line that illustrates the power-law scaling between log L(k) vs log (k).
My apologies if I am being dumb, but if we have to calculate FD of a timeseries using Higuchis FD. I would take some values of k and draw graph and take the values using the plateau.
In a machine learning problem I have a data-set with so many timeseries in it, on what thing am I supposed to apply this plateau test?
for machine learning you will need to add a test for when rollover begins to occur (the error in the slope will increase beyond some tolerance you will set).