igraph-help
[Top][All Lists]

## Re: [igraph] Building a graph from a genetic distance matrix and finding

 From: Fabio Daolio Subject: Re: [igraph] Building a graph from a genetic distance matrix and finding community structure Date: Tue, 28 Jan 2014 11:35:56 +0100

```Just my two pennies.

1) you could also try to find communities by means of the Markov Cluster
Algorithm (mcl) with the option "-pi", which applies a pre-inflation of the
network weights in order to make them less homogeneous; this way you might not
need to filter out the links.

2) if you need to compare different partitioning of the same network, you could
use the normalized mutual information measure (nmi) or the adjusted rand index;
both are already available in igraph for R via compare.communities().

Best,
--
Fabio

On Jan 28, 2014, at 11:22 , Tamás Nepusz wrote:

>> 1) A (mostly) complete graph with genetic distances between all populations
>> (a small number of the distances are 0) clusters much differently than a
>> graph with a reduced edge set. Does anyone know if techniques exist to
>> identify an optimal cutoff distance value for reducing the number of edges?
>
> I don’t think that there is a generally accepted (standard, out-of-the-box)
> method for that. You can try looking at the distribution of weights; if it
> shows two distinct peaks (one larger peak for small weights and one smaller
> peak for large weights), then the cutoff can be selected halfway between the
> mode of the two peaks. A somewhat more formal method that I sometimes use is
> to calculate the median weight and drop weights smaller than the median minus
> two or three standard deviations. (Using the median instead of the mean
> should give you some protection against the effect of outliers).
>
>> 2) There are a large number of modularity metrics in igraph.
> There is only one modularity metric, but there are multiple methods that try
> to optimize modularity.
>
>> I have experimented with both the leading.eigenvector.community and the
>> fast.greedy.community and they seem to give fairly similar results. Are
>> these appropriate metrics for my particular problem?
>
> Probably yes. Since your graph is pretty small, I would also try
> optimal.modularity since it is guaranteed to find the partition with the
> largest possible modularity (although it is clearly unsuitable for any graph
> that has more than a handful of vertices). The only catch is that
> optimal.modularity does not support weights in older versions of igraph and I
> don’t remember whether we have released a version that supports weights
> already or whether it is still in the development version. If you have the
> most recent version of igraph and it does not support a weights=… argument
> for optimal.modularity, you could try installing one of the nightly builds
> from http://igraph.org/nightly.
>
> All the best,
> T.
>
>
> _______________________________________________
> igraph-help mailing list