Multiclass transductive SVM software package? Anyone know of one? I was looking at SVM-light but it seems like you can only use it in the binary classification setting. There is also SVM-multiclass but that doesn't have tSVM support. Any help appreciated. Thanks!
 A: You may want to look at SVMlin. In its homepage, it is mentioned: 

Semi-supervised [can use unlabeled data as well]
  
  
*
  
*Multi-switch linear Transductive L2-SVMs
  
*Deterministic Annealing (DA) for Semi-supervised Linear L2-SVMs
  

A: no traditional TSVMs like SVMLin only work well for binary classification (for text problems)
you will need to build a multiclass TSVM from a collection of well designed binary TSVMs using the 1 vs 1 approach
I have a very detailed blog post on this
https://charlesmartin14.wordpress.com/2014/09/23/machine-learning-with-missing-labels-transductive-svms/
this explained in detail in the book:
http://olivier.chapelle.cc/ssl-book/
in the section where they review the benchmarks on all methods. Notice this is not in the tables but is hidden in the discussion section when reviewing the results of the TSVM and other methods as applied to text.  
There are several alternatives to SvmLin , such as UniverSVM and QN-S3VM, but they have the same limitations.
to my knowledge, no one has solved this problem for multiclass TSVM text problems and actually released production quality code , although there are many matlab versions of ideas out there.  There are several proposed methods for multiclass label propagation / graph based semi supervised learning algos but these are not applicable to text problems, and I dont know how well they work on other problems
