[
https://issues.apache.org/jira/browse/MADLIB-1084?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16410107#comment-16410107
]
Jingyi Mei commented on MADLIB-1084:
------------------------------------
[~hpandey] I personally like the idea of dealing with special nodes by
per-subgraph when grouping:
1. For the current pagerank implementation, when specifying group, the graph
will be split into several sub-graphs based on different groups, and pagerank
score is calculated separately for each sub-graph. From this perspective,
special nodes can also be different in each sub-graph and hence calculate
different personalized pagerank scores, which looks more consistent.
2. A random walker explanation: in a specific network we have multiple users,
if we group by userid, we will get several sub-graph of the big network, and
for each user, they may have different start point (special nodes). Our
calculation is based on per user and their start point.
3. If the dataset is big with lots of groups, it might be hard for every group
has exactly the same special nodes, and if we always error out for this, it may
not be convenient.
I agree that we can start with a simplest approach, but I feel eventually a
localized one by group makes more sense.
> Graph - Personalized PageRank
> -----------------------------
>
> Key: MADLIB-1084
> URL: https://issues.apache.org/jira/browse/MADLIB-1084
> Project: Apache MADlib
> Issue Type: New Feature
> Components: Module: Graph
> Reporter: Frank McQuillan
> Assignee: Himanshu Pandey
> Priority: Major
> Fix For: v1.14
>
>
> Personalized PageRank which is a variant of regular PageRank.
> Please refer to
> [http://madlib.apache.org/docs/latest/group__grp__pagerank.html] as a
> starting point.
> Reference:
> Neighborhood Formation and Anomaly Detection in Bipartite Graphs
> [http://www.cs.cmu.edu/~deepay/mywww/papers/icdm05.pdf]
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
