Re: weighted roles and distance in description logic

Cristiano Longo wrote:
> Yes of course it is a good starting point. But i think
> that it is not enough, because using this raw values
> (a) we cannot retrieve significant information using
> standard inferences(like subsumption, satisfiability,
> etc...)
> (b) it is hard to compare two distances that uses
> different unit measueres(in our exaple miles with km).
> So i was thinking to define "a priori" some distance
> classes(like for example "Near" and "Far") and, in a
> preprocessing phase, assert that a distance belongs to
>  one(or more) of such classes.

Ciao Cristiano,

If I got right you're trying to cache previously used paths from two 
different nodes in a graph no matter how far they are.

I can't see how you can store this kind of information in a generic way 
without supporting numbers AND units AND languages.

Because RDF uses named triplets (via predicate) you start by giving 
names to numbers and the nightmare begins, but in graphs you can give 
weights to the connections and IMHO this is a good addition to RDF. It'd 
simplify lots of graph algorithms (such as shortest path) and be 
optional so you can omit when there is no sense.

Imagine you want a recommendation on LinkedIn for a job and you have 
three shared connections with the guy offering the job, which one would 
you use? If you can give weights on your "friendship" you can find it by 
using "best friend" when available and see what's the "best path" 
between you and him/her. This comparison must be two-way as it won't 
help if the man-in-the-middle doesn't like you.

With current RDF (AFAIK) you have to separate in classes: BestFriend, 
Friend, Known, Enemy, WorstEnemy (etc) and give weights in your code.

With weighted graphs you can have one class Friend with a weight (-2 to 
2) and use standard djikstra to find the shortest path.

At the end, for the results only, it doesn't matter if you convert the 5 
classes into weights in your code or it stays in the data description, 
but RDF is a tool with meta data which should retain the data 
information in the data itself and not rely on specific algorithms to 
interpret it (AFAIK again).

my 2 cents,

Received on Monday, 18 February 2008 11:24:42 UTC