Score domination of frequent labels in modified adsorption?

[pvenkat...@gmail.com]

Hello.

I'm working on a multi-label classification problem with this tool-kit under 2 different labels.

The labeled data is pretty much around 1% of the whole data. But, even that label distribution is skewed. the first label overpowers the other in number by around 11 times.

I ran the algorithm without any real problems, only to find out that the minority labels have very less influence than the majority ones. Like, if I should assign labels to the unlabeled data, every unlabeled node will be labeled as the majority label.

What is the problem here? Are there any parameters I can modify to make the algorithm more efficient to my problem? 

Thanks in advance.

[Partha Talukdar]

On Sat, Jun 18, 2016 at 12:02 PM, <pvenkat...@gmail.com> wrote:

The labeled data is pretty much around 1% of the whole data. But, even that label distribution is skewed. the first label overpowers the other in number by around 11 times.

I ran the algorithm without any real problems, only to find out that the minority labels have very less influence than the majority ones. Like, if I should assign labels to the unlabeled data, every unlabeled node will be labeled as the majority label.

What is the problem here? Are there any parameters I can modify to make the algorithm more efficient to my problem? 

There are variety of things you can do: give equal number of seeds from both categories, class mass normalization (sec 11.5 in http://www.iro.umontreal.ca/~lisa/pointeurs/bengio_ssl.pdf), etc.

[pvenkat...@gmail.com]

Thanks for the reference sir, we are looking into mass normalization. 

By the way, sir, all the edges we are dealing with have weights of the order of 1e-4. The scores from the labels seem to be dependent on the edge weights, so they too are of the same order. But whatever the edge-weights may be, the dummy label always has a score to the order of 1e-1. They are over-powering. So, after going through this google group, I found a solution and changed the mu3 parameter in steps till I reached 1e-6. Now, they are all of same order. 

My question is, does decreasing a parameter by that much have any effect on the accuracy of the algorithm?

Thankyou.

[Partha Talukdar]

On Sat, Jun 18, 2016 at 8:47 PM, <pvenkat...@gmail.com> wrote:

My question is, does decreasing a parameter by that much have any effect on the accuracy of the algorithm?

Yes it may. mu3 controls regularization.