flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Attila BernĂ¡th <bernath.at...@gmail.com>
Subject Re: Spargel pagerank with sinks
Date Fri, 19 Sep 2014 08:36:28 GMT
Dear Stephan,

Thank you for your answer.

> With "sinks" in the graph, you mean vertices with no out-links?
Yes, I do.

> There might be a simple trick, by adding to each vertex an edge-to-self (put
> an entry in the diagonal of the adjacency matrix).
> I have not thought through the implications 100%.
> @ssc Can you elaborate on this?
I don't think that this works.

> What would always work is that you gather statistics about how much
> probability is accumulated in the sinks and redistribute it across the other
> nodes.
> The iteration aggregators allow you to do this. They can sum up the
> probability in the message sender function (when there is no outgoing edge),
> and re-add it to the non-sink nodes (by accessing the aggregate from the
> previous iteration).
Thank you for the idea. I was trying to use an aggregator, but I
thought that the aggregate from the previous iteration is of no use.

I will try this.



> Have a look at the function "registerAggregator()" on the
> "VertexCentricIteration", and the Functions "getIterationAggregator()" and
> "getPreviousIterationAggregate()" on the VertexUpdateFunction and the
> MessagingFunction.
> Stephan
> On Thu, Sep 18, 2014 at 5:01 PM, Attila BernĂ¡th <bernath.athos@gmail.com>
> wrote:
>> Dear All,
>> I wonder how to write the pagerank program in the spargel API if there
>> might be sinks in the graph.
>> What is the nicest way to solve this?
>> Thank you for your answer.
>> Attila

View raw message