[ 
https://issues.apache.org/jira/browse/FLINK-1707?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15442548#comment-15442548
 ] 

Josep RubiĆ³ edited comment on FLINK-1707 at 8/28/16 1:48 AM:
-------------------------------------------------------------

Hi [~vkalavri],

I've pushed a new version of AP with following changes:

https://github.com/joseprupi/flink/blob/vertexcentric/flink-libraries/flink-gelly/src/main/java/org/apache/flink/graph/library/AffinityPropagation.java

- I've changed the model to the vertex centric avoiding having the values in 
the the vertices

- I've added the option of no damping. When a 0 factor of damping is passed to 
the constructor the convergence condition is no changes on the exemplars on a 
certain number of iterations, avoiding to have the old values in vertices. This 
number of iterations is the last parameter of the constructor. If a damping 
factor different to 0 is used it keeps working as before, having to hold the 
old sent values in the vertex. 

To do:

- I have not changed the initialization of the graph yet. I posted a question 
in dev thread with no much luck. Maybe I will implement an initialization with 
a similarity matrix for now and will see how I can do it using gelly 
functionality later

- I will try to change where are the weight values to be in the edges instead 
of vertices (I've created a new version of the design document with the 
diagrams too). This way vertices will only have the old values in case the 
damping factor has to be used.

Thanks!!


was (Author: joseprupi):
Hi [~vkalavri]

I've pushed a new version of AP with following changes:

https://github.com/joseprupi/flink/blob/vertexcentric/flink-libraries/flink-gelly/src/main/java/org/apache/flink/graph/library/AffinityPropagation.java

- I've changed the model to the vertex centric avoiding having the values in 
the the vertices

- I've added the option of no damping. When a 0 factor of damping is passed to 
the constructor the convergence condition is no changes on the exemplars on a 
certain number of iterations, avoiding to have the old values in vertices. This 
number of iterations is the last parameter of the constructor. If a damping 
factor different to 0 is used it keeps working as before, having to hold the 
old sent values in the vertex. 

To do:

- I have not changed the initialization of the graph yet. I posted a question 
in dev thread with no much luck. Maybe I will implement an initialization with 
a similarity matrix for now and will see how I can do it using gelly 
functionality later

- I will try to change where are the weight values to be in the edges instead 
of vertices (I've created a new version of the design document with the 
diagrams too). This way vertices will only have the old values in case the 
damping factor has to be used.


> Add an Affinity Propagation Library Method
> ------------------------------------------
>
>                 Key: FLINK-1707
>                 URL: https://issues.apache.org/jira/browse/FLINK-1707
>             Project: Flink
>          Issue Type: New Feature
>          Components: Gelly
>            Reporter: Vasia Kalavri
>            Assignee: Josep RubiĆ³
>            Priority: Minor
>              Labels: requires-design-doc
>         Attachments: Binary_Affinity_Propagation_in_Flink_design_doc.pdf
>
>
> This issue proposes adding the an implementation of the Affinity Propagation 
> algorithm as a Gelly library method and a corresponding example.
> The algorithm is described in paper [1] and a description of a vertex-centric 
> implementation can be found is [2].
> [1]: http://www.psi.toronto.edu/affinitypropagation/FreyDueckScience07.pdf
> [2]: http://event.cwi.nl/grades2014/00-ching-slides.pdf
> Design doc:
> https://docs.google.com/document/d/1QULalzPqMVICi8jRVs3S0n39pell2ZVc7RNemz_SGA4/edit?usp=sharing
> Example spreadsheet:
> https://docs.google.com/spreadsheets/d/1CurZCBP6dPb1IYQQIgUHVjQdyLxK0JDGZwlSXCzBcvA/edit?usp=sharing



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to