Implement other classification methods in the workflow

[rogerkuou]

For the current version of the workflow, the classification is based on the average of the similarity scores. There are potentially better ways.

Talked together with Sonja about the classification methods:

1. K-nearest:

   • Compute similarity scores based on all samples used to train the Siamese
   • Take nearest neighbors
   • Take the majority of as the classification

2. SVM

   • Use sklearn, see 1.4.6.2. "Custom Kernels"; This allows one define a kernel as a Python function
   • use Gaussian kernel. See this paper for the kernel function;

     kernel(x,y) = exp(-gamma*dist(x,y)) # kernel function
     dist(x,y) = 1 - sim_score(x,y) # distance function, based on the trained Siamese

   • Use 1-sim_score as dist function
   • Still use all samples involved in the Siamese training to train the SVM
   • Or use the precomputed kernel, with each element similarity score

[rogerkuou]

Ask Sonja: Average sim score and K-nearest, any literature?