[igraph] Re: Calculating measures for a bunch of ego-networks

[Magnus Torfason]

Hi again,

I just wanted to give an update. It turns out that although 
graph.neighborhood() is very slow, neighborhood() is very fast, and I've 
managed to get quite a ways using a combination of
  - neighborhood()
  - get.adjedgelist()
and then using the V() and E() lookup functions, which are also very 
fast when using known indexes (that I can look up from the lists 
returned by neighborhood() and get.adjedgelist().

One piece that I haven't figured out yet is how I can get a list of edge 
IDs, given a pair-wise list of nodes. I tried the following, but none of 
them got me where I wanted to go:

  - E(g, P=pair.wise.vector)
This errored out because not all the edges existed. The problem is for 
each pair, I don't know beforehand if a tie exists between them.

  - are.connected(g, v1, v2)
This would allow me to ask E() only for edges that actually exist - but 
this function is not vectorized, meaning that I need to loop thousands 
of times to check a certain set of vertices.

  - E(g)[ left.side.vector %--% right.side.vector ]
This actually does more than E(g, P=...) because it checks all 
combinations. That is actually cool, because that is what I need. 
However, this actually seems to take orders of longer than doing the 
are.connected() loop.

Any thoughts on a good way to solve my particular problem?

And any thoughts on which general low-level edge/vertex lookup functions 
are useful for calculating custom network measures of any sort? My idea 
for criteria are:
 - Very fast
 - Returns a set of edge/vertex IDs fulfilling some reasonable criteria
 - Returns them in a single list of length vcount(g) and therefore
 - Do not need to be called in a loop
 - Do not fail easily (see comment on E(g, P=...))

neighborhood() and get.adjedgelist() are clearly such functions, neat 
combinations of E()/V() and particular parameters might also be in the set.



On 5/5/2010 12:29 PM, Magnus Torfason wrote:
> Hi all,
>
> I have a (relatively sparse) network of 100K nodes in igraph. Each node
> has a bunch of attributes, and the edges have some attributes as well.
>
> I've calculated some standard measures on the nodes (degree, constraint,
> etc) and this has been pretty fast.
>
> However, now I need to calculate some non-standard measures for each
> node, that might theoretically depend on the attributes of:
> - The attributes of the focal vertex
> - The attributes of the edges of the focal vertex
> - The attributes of the vertices in the immediate neighborhood
> - The ties between the neighbors
>
> So this can be accomplished by looping through the vertices of the
> graph, extracting the graph.neighborhood(g, 1, i) for each, and
> manipulating it. However, this takes a long time. Are there any better
> ways for doing stuff like this?
>
> I think I'm kind of hoping that there is a way of writing a function
> that takes a vertex ID, and that this can be "applied" to all the
> vertices of the graph, but in an efficient manner that does not require
> generating 100K subgraphs.
>
> I guess an alternative way of phrasing the question would be: If I
> wanted to implement the constraint measure in R code in an efficient
> manner, could I do it?
>
> Thanks for any help and ideas!
>
> Best,
> Magnus