0

I am working with k-means and K-medoids. With K-means execution appear the following info:

Within cluster sum of squares by cluster:
[1] 12636160  7631152 10226254
(between_SS / total_SS =  79.2 %)

Is between_SS/ total_SS a rate that shows the general throughput from the algorithm?

And with pam:

Objective function:
build     swap 
211.6604 210.5670

How do you interpret these results?

Roy
  • 19
  • 4

1 Answers1

2

If by "throughput" and "efficiency" you mean anything about processing speed, then no. These are all measures of how successful the clustering algorithm was at finding a good grouping (or perhaps how well these points can be grouped).

k-means
The meaning of between_SS (between clusters sum of squares) and total_SS (total sum of squares) has been explained in this previous Cross Validated question and its answers. The ratio of between_SS to total_SS is some measure of how well the points clustered.

PAM
From ?pam help page

the algorithm first looks for a good initial set of medoids (this is called the build phase). Then it finds a local minimum for the objective function, that is, a solution such that there is no single switch of an observation with a medoid that will decrease the objective (this is called the swap phase).

The values listed are the values of the objective function (sum of distances of points to their medoid) at the two stages. Again, this is a measure of how well the points clustered.

For more details, see the pam help page ?pam ,
the pam.object help page ?pam.object,
the Wikipedia Page on k-medoids or
the original paper Kaufman, L. and Rousseeuw, P.J. (1987), Clustering by Means of Medoids

G5W
  • 36,531
  • 10
  • 47
  • 80
  • Thanks, and is there any way to enhance algorithm performance? There are some errors when I compare it with the original data. The estimations are not very accurated, thus classification by cluster does not offer a good solution at this case. – Roy Jun 10 '17 at 20:26
  • Classification by clustering is rarely (probably never) a good idea. If you have classes in your original data, why not just use one of the many supervised classification algorithms? I suggest that you look at this [Previous Post](https://stackoverflow.com/q/5064928/4752675) which describes the difference. – G5W Jun 10 '17 at 20:32