Doubts about Hubs & Authorities implementation

[dkigel]

Hi everybody,

First of all let me give you my congratulations for this excelent piece of software you are developing. I've started using Gephi a few weeks ago and I'am really impressed with its practical user interface, high visualtization capabilities and varaiety of network analysis algorithms included.

At this moment I am working with eigenvector centralities for information retrival, particularly I am writing you to ask about the Hits & Authorities implementation.

My doubts are about some diferences I've found between the source code Hits.java in github and the algorithm's description done in the reference article http://www.cs.cornell.edu/home/kleinber/auth.pdf.

More precisely my questions have to do with nodes' authority and hub value update in each iteration:

(1) At each iteration you init new nodes' authority and hub values "temp_authorities" and "temp_hubs" with values obtained at the previous iteration and then you add to this original values neighbors' hubs or authorities values respectively. Shouldn't zero initial value for "temp_authorities" and "temp_hubs" be used ?

(2) I understand at a node's hub value update you should sum authority values of nodes connected by outcoming edges. Why are you iterating over incoming edges as you do for node's authority value update ?

I've done some tests comparing Hubs and Authorities values obtained from Gephi with those calculated with Networkx, a python library which also implements HITS, and I've found they are definitely different.

Probably my doubt aren't a problem and I am missing some important points here!

Can help me to understand these diferences?

Thanks in advance!

David

[admin]

Hi David,

Thank you for pointing this. The guy who coded the algorithm is not part of Gephi anymore (it was 4 years ago), so it's hard to answer now...

Anyway, if you find ground truth that our implementation is wrong and networkx is right, please open an issue on github and we'll fix it.

[ignacio.morer]

Hi everyone,

I've spotted a problem with this algorithm too, i'm not sure if it is the same one. The 'authority' ranking is only based on nodes' in-degree value, which doesn't match with the procedure explained in the reference article. I checked on github and I've seen there is an open issue about HITS resolution, so I guess we'll have to wait

I'd like to congratulate you as well for the application and its constant improvement. I'm looking forward to being able to export data from graph's statistics (Because it cannot be done yet, right?). I think I read something about this being done this summer, awesome!

Thanks

Ignacio

[juliancab]

Hi there,

I would like to know if this issue with the calculation of hub authority has been fixed as of September 2013.

I checked the status of issues reported in GitHub (https://github.com/gephi/gephi/pull/756; https://github.com/gephi/gephi/issues/620) and it appears that the issue was resolved in March 2013, but I want to double check. I get some different results when comparing the values obtained with Pajek in a directed graph.

Thanks!

Julián