Replies to This Discussion

Permalink Reply by Stephen Cronin on June 21, 2010 at 10:08am

Paul what are you using a TSVM for? I've been looking for a reason to code this up in MATLAB and add it to my companies libraries (actually a TSMO for the speedup). I think you will have some trouble finding this open source at least until one of them get it on the priority list.

Permalink Reply by Paul Wilson on June 21, 2010 at 12:42pm

Hi Stephen,

I'd like to use TSVM in reject inferencing situations.

Say for example a company sends mail to a group of customers and wants to build a model to predict future customer's likelihood to respond.

Well, companies usually don't contact all of their customers in one campaign, so some customers would not have a chance to respond to begin with due to them not being selected in the selection process.

I'd like to make use of their data as well and have TSVM first estimate what the customers that were not contacted would have done, assign "responder" or "nonresponder" label to them, and then estimate my model using entire dataset.

This can be used in many other situations like, say a bank wants to build "good vs bad" loan models, but some customers never got their loan applications approved to begin with etc.

Yeah, I'm having trouble finding this algorithm implemented in some sort of GUI, open source or not unfortunately :(
All I found was some C or C++ code, but I'm not familiar with the code itself so I'm not able to run it
:(

Permalink Reply by Stephen Cronin on June 21, 2010 at 1:32pm

Very interesting. I would love to collaborate on getting a MATLAB version of this up and running to be compiled into C. only problem I see is that SMO scales as an O(n^2) or with some clever tricks a little less then that so it would depend on the size of your data set. if you have say tens of millions of vectors the data crunching cost may be astronomical.

Permalink Reply by Steffen Springer on June 21, 2010 at 3:23pm

I am not that sure, but I think that the commercial version of rapidminer includes a implementation of transductive svm (in java).

I dream of the day when http://mahout.apache.org/ includes an implementation... (https://cwiki.apache.org/MAHOUT/algorithms.html) ...

Permalink Reply by Stephen Cronin on June 21, 2010 at 3:37pm

confirmed, and my friend who uses the enterprise rapid miner at work says it does a decent job. I wonder what the cost tag is associated with it.

Permalink Reply by Paul Wilson on June 22, 2010 at 10:15am

Cool, nice to know that Rapid Miner has it.
I was thinking of spending some time in near future to learn their application.

Too bad TSVM is not in their community edition. That would have been a really good incentive to start learning and using Rapid Miner :)