I would say 'yes' though somehow I suspect it will not be the most
effective technique to apply to this problem.

For example, suggesting friends of friends as new contacts, something
as simple as that, might well be the best thing.

I don't think you need to know the strength of the relationship
necessarily, though more information is always better. There are
similarity measures like the Tanimoto coefficient (yep, these and
others are already implemented) which are just based on binary yes/no
relationships between "items" -- here, users.

Once you have a similarity metric, you can cluster, sure. That alone,
I think, doesn't show you who is not connected as friends who should
be. That is where CF helps in the sense that it is trying to guess at
who *else* should be friends, rather than reporting on who has similar
groups of friends.

Should be easy to try! and you can even use RecommenderEvaluator to
score your implementations. Anyone know of a good social network data
set? I could tinker with it to see if anything interesting comes out.
This would also be a chance for me to really use and refine the code
I've added to support the binary yes/no rating types that I wrote in
recently.


On Tue, Jul 15, 2008 at 9:01 AM, Robin Anil <[EMAIL PROTECTED]> wrote:
> If the strength of relationship between 2 parties is given. Then essentially
> you can create a graph of people with connections. Normal, clustering of
> nodes can help find similar users. I dont know how cf will help in that.
>
> Robin
>

Reply via email to