Open sa-lee opened 8 years ago
@gtonkinhill i think this might be reasonable for testing different label switching algs. should look into Stephens, 2000 as well.
As discussed the other day - these are the underlying problems here:
The suggestion by Melanie to solve 2. is to pass a confusion matrix as input for a sample being mislabelled. i.e. say we have a Q-matrix with correct labels then in each replicate we add noise to the Q and swap labels with some probability given by the confusion matrix.
Alternatively
The label.switching package has a function called permute.mcmc
that I think does what we want.
see http://www.cs.cmu.edu/~epxing/Class/10701-08s/recitation/dirichlet.pdf and https://en.wikipedia.org/wiki/Dirichlet_distribution
Will have to fix labelling at some point but it is a start.