Re: [foaf-dev] Tinkerpop

  On 5/10/2011 6:00 AM, [email protected] wrote:
> Interesting.
>
> I think this may be different to RDF Graphs in that you can annotate
> the edges.  I saw A knows B (since=2007) in one slide.  You cant
> easily do that in RDF without reification.
>
> So that makes it more a "network" than a "graph".  (Similar to a
> Geometry vs a Topology)
>
> There's a whole other field of problem you can solve with this
> technology, e.g. travelling salesman, Dijkstra's algorithm etc.
>
     Tinkerpop looks pretty cool but my immediate reaction is that the 
flexibility of accessing different graph stores through an A.P.I. means 
that it's not really scalable.  My take is that neo4j and friends have a 
hard time dealing with the DBpedia graph never mind Freebase or anything 
bigger.

     Not all algorithms on graphs will parallelize,  but I'd be much 
more impressed with some kind of system that can do parallel traversals 
that are spread out on a set of database nodes.  Without something like 
that,  Tinkerpop is syntactic sugar for algorithms that aren't that hard 
to code in straight Java,  PHP or whatever you do.

     That said,  I've been interested lately in running network analysis 
algorithms on semantic graphs.

     One of the first things that comes up is that you have to put some 
kind of weights on the edges to increase the dynamic range of the 
metrics you get.  For instance,  if you want to estimate relatedness by 
the number of edges that separate two nodes and you've got a graph with 
"six degrees of separation" you have a very small number of discrete 
distance values,  and the fact that some nodes are going to end up close 
by accident dooms the enterprise,  unless you've got some way to 
characterize the strength of relatedness,  in which case you might find 
errors get averaged out.