I want to solve an svdd quadratic problme witch is a one classe classification
probleme but with constrainte [ Sum (alph i)==1]
so i have try to use one_class_svm witch i have download from libsvm(version 2.8)
with those paraimetrs
param->svm_type=2
param->kernel_type=2
param->degree=0
param->gamma=0.5
param->cache_size=100
param->eps=0.001
param->probability=0
param->shrinking=1
but i have a probleme because i haven't Sum(alph i)==1
how can i solve this problem please
how cnn i fixed the paramaitre delta in Generalised SMO algorithme

min 0.5(\alpha^T Q \alpha) + b^T \alpha
//
// y^T \alpha = \delta
// y_i = +1 or -1
// 0 <= alpha_i <= Cp for y_i = 1
// 0 <= alpha_i <= Cn for y_i = -1
 
I need olso the Smo generalised code source on C++.

Re /m3:
Lingyun i guess you are looking at the paper to be presented next monday.

There are tricks to let SVM do multi-class classification, like binary coding, or as you suggested, one_class-vs-other binary classification.
Yes, the length requirement of NIPS prevents them to put more experimental details. I hope you get their idea of constructing kernels over automata.

In the introduction, it says "Investment performance depends upon sequences of interdependent decisions, and is thus path-dependent". What does "path-dependent" mean? I guess decisions made in early stages will have effect on not only how to decide but also what to decide later.

Aside from being dense & cryptic, I was disappointed by the "results" section as well. Part of me thinks the
problem is that they tried to squeeze a 30-page paper
into an 8-page NIPS submission.

In 4.3) they mention the task is to classify the utterance into 1 of 38 classes. Since SVM's typically do binary classification, what was their setup? If you train 38 SVM's in parallel, you can't guarantee unique classification, which I assume they need.

I know I have chosen a dense and cryptic paper. I am glad to answer your request for clarifications.