CODEWITHSUNDEEP
matlabsvmlibsvm
MUH
MUH

Reputation: 123

LIBSVM output regarding model.SVs

I am new to libsvm. I am having trouble understanding the output of libsvm. I just want to know how do you find the alpha_i of all the support vectors? Does the parameter model.sv_coef give you the alpha_i or alpha_i*y_i? and the other main question that I wanted to ask is what does model.SVs give? I am using libsvm in matlab and the output I am getting when I look for model.SVs is:

my training set is: 

-1 1:0.747253 2:0.894737     
+1 1:0.692308 2:-0.824561     
-1 1:0.362637 2:0.789474     
+1 1:0.769231 2:-0.321637

my test set:

-1 1:-0.351648 2:-0.602339

+1 1:-0.21978 2:-0.263158



var1=model.sv_indices

>>var1 =

     2
     4
     3



model.SVs

>>ans =



(1,1)       2.0000    
(2,1)       4.0000    
(3,1)       3.0000   
(1,2)      -0.2204    
(2,2)       0.2870    
(3,2)       0.9774   
(1,3)       1.1592   
(2,3)       0.7978    
(3,3)      -0.3999   
(1,4)      -0.3999    
(2,4)       0.0250   
(3,4)       0.7548   
(1,5)       0.7978    
(2,5)       0.6952    
(3,5)       0.0250

what are these (1,1),(2,1),(3,1)...? I mean I have only 4 training set so how can this second index go to 5 and only two test set so how can the first index go to 3.

Upvotes: 2

Views: 979

Answers (1)

lennon310
lennon310

Reputation: 12699

model.sv_coef contains all the alpha_i * y_i, model.SVs are all the support vectors x_i. The weights can be written as

w = model.SVs' * model.sv_coef;
b = -model.rho;

You may not need to care much about the labeling of (1,2), (1,3)...m.nSV will give you the number of support vectors for each class. Since you have only two classes, the answer should be a 2*1 vector with each row representing the support vector number for the corresponding class. Check this document for more details.

Upvotes: 2

Related Questions