Hi Greg,
it seems that it doesn’t matter with the vertex „3“ with no degree.
I removed these vertex in the graph and in a second test of my input file. The ranking order
is still different, and I guess wrong. Furthermore is the sum of all ranks not 1. It depends
on the betaparameter. E.g. a beta of 0.15 on the sg PageRank calculate
(2.0 , 0.38102628032106706)
(4.0 , 0.4547945998174918)
(1.0 , 0.4341925979005684)
The sg and a beta of 0.85 returns:
(2.0 , 97.53826698457634)
(4.0 , 140.49741661507886)
(1.0 , 135.265886297257)
All of these are issues of vertexcentric, sg and gsa implementation. The last one (without
any graph model) works fine.
Do you have any idea what I doing wrong?
Marc
Am 24.07.2017 um 20:56 schrieb Kaepke, Marc <marc.kaepke@hawhamburg.de<mailto:marc.kaepke@hawhamburg.de>>:
Thanks for your explanation.
The vertexcentric, sg and gsa PageRank need a Double as vertex value. A VertexDegree function
generate a vertex with a LongValue as value.
Maybe I can iterate over the graph and remove all edges with a degree of zero?!
Am 24.07.2017 um 16:36 schrieb Greg Hogan <code@greghogan.com<mailto:code@greghogan.com>>:
The current algorithm is unweighted though we should definitely look to add a weighted variant
and consider PersonalizedPageRank as well.
Looking at your results, PageRank scores should sum to 1.0, should be positive unless the
damping factor is 1.0, and use of the convergence threshold will guarantee accurate results
on large graphs.
The PageRank tests compare results from the NetworkX implementation. The missing vertex 3
is trivially fixed by adding the call ".setIncludeZeroDegreeVertices(true)” to the VertexDegrees
function.
On Jul 23, 2017, at 6:38 AM, Kaepke, Marc <marc.kaepke@hawhamburg.de<mailto:marc.kaepke@hawhamburg.de>>
wrote:
Hi Greg,
I do an evaluation between Gelly and GraphX (Spark). Both frameworks implement PageRank and
Gelly provides a lot of variants (*thumbs up*).
During a really small initial test I get for the vertexcentric, scattergather and gsa version
the same ranking result. Just the implementation in 1.3.X (without any graph model) computed
a different result (ranking).
/* vertex centric */
DataSet<Vertex<Double, Double>> pagerankVC = small.run(new PageRank<>(0.5,
10));
System.err.println("VC");
pagerankVC.printToErr();
/* scatter gather */
DataSet<Vertex<Double, Double>> pageRankSG = small
.run(new org.apache.flink.graph.library.PageRank<>(0.5, 10));
System.err.println("SG");
pageRankSG.printToErr();
/* gsa */
DataSet<Vertex<Double, Double>> pageRankGSA = small.run(new GSAPageRank<>(0.5,
10));
System.err.println("GSA");
pageRankGSA.printToErr();
/* without graph model */
DataSet<Result<Double>> pageRankDI = small
.run(new PageRank<>(0.5, 10));
System.err.println("delta iteration");
pageRankDI.printToErr();
My input graph is:
vertices
* id 1, val 0
* id 2, val 0
* id 3, val 0
* id 4, val 0
edges
* src 1, trg 2, val 3
* src 1, trg 1, val 2
* src 2, trg 1, val 3
* src 2, trg 4, val 6
Ranking output
* vertexcentric
* id 4 with 1.16
* id 1 with 1.103
* id 2 with 0.815
* id 3 with 0
* sg and gsa
* id 4 with 2.208
* id 1 with 2.114
* id 2 with 1.546
* id 3 with 0
* new PageRank in Gelly 1.3.X
* id 1 with 0.392
* id 2 with 0.313
* id 4 with 0.294
Do you know why?
Best
Marc
Am 23.07.2017 um 02:22 schrieb Greg Hogan <code@greghogan.com<mailto:code@greghogan.com>>:
Hi Marc,
PageRank and GSAPageRank were moved to the flinkgellyexamples jar in the org.apache.flink.graph.examples
package. A library algorithm was added that supports both source and sink vertices. This limitation
of the old algorithms was noted in the class documentation and I understand to be an effect
of delta iterations. The new implementation is also significantly faster (https://github.com/apache/flink/pull/2733#issuecomment278789830).
PageRank can be run using the examples jar from the command line, for example (don’t wildcard
the jar file as in the documentation until we get the javadoc jar removed from the next release).
$ mv opt/flinkgelly* lib/
$ ./bin/flink run examples/gelly/flinkgellyexamples_2.111.3.1.jar \
algorithm PageRank \
input CSV type integer simplify directed input_filename <filename> input_field_delimiter
$'\t' \
output print
The output can also be written to CSV in similar fashion to the input.
The code to call the library PageRank from the examples driver is as with any GraphAlgorithm
(https://github.com/apache/flink/blob/release1.3/flinklibraries/flinkgellyexamples/src/main/java/org/apache/flink/graph/drivers/PageRank.java):
graph.run(new PageRank<K, VV, EV>(dampingFactor, iterations, convergenceThreshold));
Please let us know of any issues or additional questions!
Greg
On Jul 22, 2017, at 4:33 PM, Kaepke, Marc <marc.kaepke@hawhamburg.de<mailto:marc.kaepke@hawhamburg.de>>
wrote:
Hi there,
why was the PageRank version (which implements the GraphAlgorithm interface) removed in 1.3?
How can I use the new PageRank implementation in 1.3.x?
Why PageRank doesn’t use the graph processing models (vertexcentric, sg or gsa) anymore?
Thanks!
Bests,
marc
