hama-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Thomas Jungblut <thomas.jungb...@googlemail.com>
Subject Re: Multiple aggregators & PageRank example
Date Tue, 22 May 2012 10:11:57 GMT
It's not a big deal to extend it for multiple aggregators.
I open a new ticket for 0.5.0 then.
Do you need the sum of each probability in every superstep?
Otherwise it may be more simple to initialize the leafs to 1.0 vertexvalue
and don't let them update in the compute logic.
Also the aggregator is observing the vertex value, would it be more
convenient if the whole vertex object is available when aggregating?
Currently it is done more implicitly.

2012/5/22 SWP <contact-semweb@unister-gmbh.de>

> Hi,
>
> looking at the example PageRank.java (revision 1341332), it appears that:
> - within PageRankVertex.compute, we have numEdges=1 for a vertex without
> outgoing edges, because each is assigned a self-loop in line 114,
> pageJob.set("hama.graph.self.ref", "true");
> - the corresponding "random surfer" will leave a vertex without outgoing
> edges (apart from the self-loop) with a probability of
> (1-DAMPING_FACTOR) = 0.15, assuming the default settings are used
>
> I would like to change this behaviour such that:
> - vertices never have loops
> - the random surfer leaves a vertex without outgoing edges with
> probability 1
>
> However this implies that I will have to sum up globally the probability
> mass over all vertices without outgoing edges and redistribute it
> uniformly over the graph.  This should happen in addition to the 0.15
> restart probability.
>
> Now, aggregators seem to be the concept in Hama to support this kind of
> global communication.
> But there can only be one aggregator, and it is already in use to
> facilitate the stopping criterion.
>
> Is there a way to support more than one aggregator?  Should I write a
> new aggregator class that somehow supports both tasks?  And am I getting
> things right anyway? ;-)
>
> Clemens Gröpl
>
>
> --
>
> Semantic Web Project, IT
>
> Unister GmbH
> Barfußgäßchen 11 | 04109 Leipzig
>
> Telefon: +49 (0)341 49288 4496
> contact-semweb@unister-gmbh.de <mailto:%20contact-semweb@unister-gmbh.de>
> www.unister.de <http://www.unister.de>
>
> Vertretungsberechtigter Geschäftsführer: Thomas Wagner
> Amtsgericht Leipzig, HRB: 19056
>
>


-- 
Thomas Jungblut
Berlin <thomas.jungblut@gmail.com>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message