Error in Page Rank Computation in PageRank.scala

[pfontana3w2]

I saw something strange in the Page Rank computation for runUntilConverge() in PageRank.scala. It uses the oldPR instead of the resetProb. Note that the run() Method in PageRank.scala uses resetProb as my correction does here (see Lines 95–96 of PageRank.scala).

Here is the diff that I see (in case it is hidden later):

-valnewPR= oldPR + (1.0- resetProb) * msgSum
+// Equation: resetProb * (1-resetProb)*msgSum+valnewPR= resetProb + (1.0- resetProb) * msgSum

This might not be the right correction, but I brought a pull request to see if I have found an error or not. If it is not correct, just deny the pull request.

Best Wishes

[AmplabJenkins]

Can one of the admins verify this patch?

[pfontana3w2]

Do take a look. I noticed that with this patch, dangling nodes no longer have page ranks equal to their reset probabilities, so there may not be an error, or my solution may not be the right one. Here is a small test that I did.

Edge Table:

Node1	Node2
1	2
1	3
3	2
3	4
5	3
6	7
7	8
8	9
9	7

Node Table:

NodeID	NodeName
a	1
b	2
c	3
d	4
e	5
f	6
g	7
h	8
i	9
j.longaddress.com	10

Page Ranks Before Patch:

(4,0.29503124999999997)
(1,0.15)
(6,0.15)
(3,0.34124999999999994)
(7,1.3299054047985106)
(9,1.2381240056453071)
(8,1.2803346052504254)
(10,0.15)
(5,0.15)
(2,0.35878124999999994)

Page Ranks After Patch:

(4,0.2488125)
(1,0.3)
(6,0.3)
(3,0.5325)
(7,0.23925000000000002)
(9,0.19335000000000002)
(8,0.45600000000000007)
(10,0.3)
(5,0.3)
(2,0.2488125)

Note that node 10 is not in the edge table

[ankurdave]

ok to test

[SparkQA]

QA tests have started for PR 2100 at commit 18eb231.

• This patch merges cleanly.

[SparkQA]

QA tests have finished for PR 2100 at commit 18eb231.

• This patch fails unit tests.
• This patch merges cleanly.
• This patch adds no public classes.

[pfontana3w2]

It looks like I may be mistaken (since the unit tests failed), so I am closing this pull request.

[ankurdave]

To answer your concern, the runUntilConvergence version of PageRank uses delta messages, where the msgSum operates on the delta graph where the rank of each vertex is the change in PageRank from one iteration to the next. You can see that on lines 144 and 149: the sendMessage function uses edge.srcAttr._2 as the source rank, which was set to newPR - oldPR.

As a result, we have to add in oldPR at each step to obtain the PageRank of the input graph rather than the delta graph.

[pfontana3w2]

Thank you for taking the time to explain the code. I did another test and noticed that the two methods seem to give different outputs for a test case. See Issue: https://issues.apache.org/jira/browse/SPARK-3206?jql=project%20%3D%20SPARK%20AND%20resolution%20%3D%20Unresolved%20AND%20priority%20%3D%20Major%20ORDER%20BY%20key%20DESC