[mirtoolbox] understanding mirnovelty
- From: Hendrik Schreiber <hs@xxxxxxxxxxxx>
- To: mirtoolbox@xxxxxxxxxxxxx
- Date: Sat, 19 Feb 2011 13:17:12 +0100
Hi,
I'm trying to understand how the options TimeLag and Horizontal are of any use
when computing mirnovelty and how they really relate to Foote's original paper.
I understand how the timelag and horizontal simatrices differ (nicely
illustrated in the manual), but shouldn't they nevertheless deliver the same
results when it comes to novelty?
And if not, does one of the scores correspond directly to Foote's scores? He
neither seems to use the timelag nor the horizontal incarnation... but it seems
that his results should nevertheless be the same as the horizontal version,
right?
I've been trying to compute the novelty scores with Java code, but never get
the same results (which may very well be my own fault).
Also, assuming that the kernel with length 64 has to cover the simatrix
completely to produce a novelty score (no padding), shouldn't I always get N
minus 64 non-NaN-novelty scores for N features? Because that's not the case.
Thanks for any insights,
-hendrik
PS: The manual falsely states the default kernel width for mirnovelty as 128.
It's actually 64.
Other related posts:
