Parameter C impact on SSVM in pystruct

[Hana Jarraya]

Hi,

I know they are stupid questions maybe but I could not find explanation for this.

I am using the same input data for SSVM with EdgeFeatureGraphCRF with C very high value for two cases with inference method QPBO:

1/ I use only three features for the nodes and for the edges. It does not work. It brings me error regularization problem

2/I use more features for my data and SSVM works correctly.

The second problem for the same value of C, not all the inference methods brings solution. For example QPBO offers results for High value of C but max product and ad3 are not able to work with the same value of C.

the third problem, I used the same configuration the parameters of SSVM with same data. I got an accuracy value and after 2 weeks I run it again I got a different value.