Choosing the optimal K value

[yassineS]

Hello,

Thanks for the great software. In the multithreaded runs, does the software output anything that allows the choice of the optimal K values? Something similar to the CV error in admixture.

Also, Would you recommend running the software multiple times with different seeds and then averaging the Q matrices?

[AlbertDominguez]

Hello,

Thanks for your interest in the software :)

As of now, there is no K-choosing method for the multi-head version of Neural ADMIXTURE. However, we are looking into some approaches and we will add it to the software if we find a conclusive approach.

Regarding averaging different Q matrices, I would not recommend it doing in a vanilla way as clusters might be permuted across runs, and averaging would destroy all the information.

Depending on the application, it might be sensible to do some kind of cluster matching and then average the runs. However, I am not sure if there would be an advantage in doing that as we expect the network not to be sensitive to the RNG given the initialization should drive the network to very similar solutions.

[yassineS]

Thanks @AlbertDominguez.