After reading a bit about it: this seems to implement the Voronoi
Iteration method mentioned in the wikipedia article, not PAM. PAM tries
every possible swap, while the Voronoi Iteration tries only the swaps
within a group. The cost function is the summend distances between
non-mediods and their closest mediods.
On Sat Feb 13, 2021 at 2:47 PM CET, xash wrote:
> This is already very good code! Some improvements other than edist:
>
> Using `each` for rank "0 like in the last line of kM_step should be
> avoided, as it just boxes/unboxes the result unecessarily.
>
> Instead of building up the groups per hand (I. & (= & s)), you can just
> use `/.` to group the indices.
>
> `?` already works on indices, so you can just use `4 ? #` instead of
> `({~ 4 ? #) i. # x`.
>
> I'm not sure how the original algorithm should work, but isn't the
> avarage
> (+/%#) unneccesary as the length of the group is equal among its
> elements?
> So +/ should be enough for comparison.
>
> Tacifying your code then a bit results in:
>
> kM2_step=: 4 : 0
> groups=. (i. <./)@:{"1
> sub=. {"1 {&x
> best=. [ {~ [:(i. <./) [:+/"1 sub
> (y groups x) best/. (i. # x)
> )
>
> kM2=: edist@] kM2_step^:_ (? #)
>
> 4 kM2 data
>
> On Sat Feb 13, 2021 at 10:54 AM CET, Emir U wrote:
> > Hi guys, I'm very new to J (about 6-7 hours experience), I'm coding a
> > bunch of useful algos for clustering, optimisation and stats to give J a
> > proper go: this is my first. I (think) I've coded the k-mediods PAM algo
> > as described here: https://en.wikipedia.org/wiki/K-medoids
> >
> > I'm trying to tune my intuition as to what this code should look like
> > when written properly. i.e. in canonical form. I'd be grateful for your
> > feedback. I've enclosed the code below which is basically the kM_step
> > and kM functions: the rest is testing code. If you run it it'll generate
> > some data, cluster it and plot it for you.
> >
> >
> > NB. UPDATE MEDIODS FROM DATA.
> > kM_step=: 4 : 0
> > d=. (x { "1 y)
> > s=. ([@> (i. <./) each (;/ d )) { x
> > idx=: I. & (= & s)
> > d=. { "1 ({ & y)
> > w=. (i. <./) & (+/%#) & d
> > ; ((w { [) & idx) each x
> > )
> >
> > NB. ITERATE UPDATE STEP TO CONVERGENCE.
> > choose=: 4 : '({~ x ? #) ,y'
> > kM=: 4 : '(kM_step & y) ^:_ x choose (i. #y)'
> >
> > NB. CALCULATE PAIRWISE EUCLIDEAN DISTANCE MATRIX.
> > edist=: 3 : 0
> > x=. +/ |: y^2
> > a=. (,.x) +/ .+ (,:x)
> > b=. y +/ .* (|: y)
> > a - (2*b)
> > )
> >
> > NB. TESTING: GENERATE DATA, FIT & PLOT.
> > load 'plot'
> >
> > data=: ?(50 2$0)
> > data=: data,3+(3*?(50 2$0))
> > data=: data,5+(4*?(50 2$0))
> > D=: edist data
> > M=: (3 kM D) { data
> >
> > pd 'reset'
> > pd 'type dot'
> > pd 'pensize 3.1'
> > pd 'color gray'
> > pd ;/ |: data
> > pd 'show'
> > pd 'color red'
> > pd 'pensize 4'
> > pd ;/ |: M
> >
> > M
> >
> > ----------------------------------------------------------------------
> > For information about J forums see http://www.jsoftware.com/forums.htm
>
> ----------------------------------------------------------------------
> For information about J forums see http://www.jsoftware.com/forums.htm
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm
