Alexey Stukalov

Thanks for the clarifications. I think we definitely should not do a hard switch from `Matrix` to `SparseMatrixCSC`, because it will kill the performance for most of the use cases....

Thanks for the report! You are most welcome to submit a fix. Otherwise, if you have a small (

@pcjentsch I guess it was a lack of maintainers resources :( Unfortunately, the PR was submitted quite long time ag, and the codebase has changed since. To the level that...

We have started to relax the input matrix requirements for some algorithms (#129, #138). It could be, though, that for optimal performance DBSCAN relies on the input matrix being dense....

This is, sort of, reasonable: to efficiently calculate distance, Distances.jl (note that the error is triggered from within that package) uses BLAS matrix multiplication routines. These routines require that the...

Thanks for the detailed investigation! If Integer indices are working, you may consider using them instead of UInt. Or maybe it could be fixed by adding Uint support to existing...

`sharedinfodist` :+1: I'd rather avoid abbreviation, so what about `sharedinfodistance` or `sharedinfometric`? I was also thinking about `infovariation`, but IMHO "shared info metric/distance" describes this function much better than "variation...

> The disadvantage being that it requires converting the matrix to a floating point one, so effectively copying the data. Why so? The `centers` are calculated in `update_centers!()`, and from...

I think it's worth merging. I'd rename the parameter to `ntries`. The default (10) seems reasonable given that the method is quite fast. Obviously, the PR has to be updated...

I had a brief look, it seems SKLearn has a separate implementation of e.g. Ward algorithm if connectivity matrix is provided. So I'm not sure it's would be a matter...