[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
## Re: [igraph] Compare clusterings

**From**: |
Stijn van Dongen |

**Subject**: |
Re: [igraph] Compare clusterings |

**Date**: |
Fri, 23 Jul 2010 17:34:38 +0100 |

**User-agent**: |
Mutt/1.4.2.2i |

Hi,
On Fri, Jul 23, 2010 at 05:19:50PM +0100, Tamas Nepusz wrote:
>* > Hello! I installed igraph as an R package. If i run two clustering*
>* > algorithms on the same graph and i have the two membership vector*
>* > then, how can i compare the two partitions? I would like to know*
>* > which points changed their cluster? Thank you Albert*
>* Well, the problem is not as simple as it seems; consider the following*
>* two membership vectors:*
>* *
>* c(1,1,1,1,2,1,2,2,2,2)*
>* c(2,2,2,2,2,1,1,1,1,1)*
<snip>
Tamas is correct of course, and if two clusterings are very different
it will be generally impossible to establish such a simple
transformational pattern of points changing clusters.
A less ambitious but informative and doable approach is to consider the
contingency table
between the two clusterings -- simply a table containing the sizes of the
intersections of all possible pairings of clusters.
Two clusters (say c1 and d1) that have a large intersection (say larger than
either of the two set differences) can serve as anchors between the two
clusterings; you could say an abstract set S is represented by c1 in the first
clustering and d1 in the second clustering, and relative to S you could make
statements of transformations such as you suggest. But the number of anchors
you are able to find is not clear in advance. It could be a lot or a few, but
the scenario that not both of your clusterings are fully covered by such
anchors is pretty realistic.
I have not used igraph a lot yet, so I don't know what would be a good way
to create such a contingency table.
best,
Stijn
--
Stijn van Dongen >8< -o) O< forename pronunciation: [Stan]
EMBL-EBI /\\ Tel: +44-(0)1223-492675
Hinxton, Cambridge, CB10 1SD, UK _\_/ http://micans.org/stijn