Support pre-calculated affinity matrix in ht.cluster.Spectral

[ClaudiaComito]

Feature functionality
Allow users to provide a pre-calculated affinity matrix when computing spectral clustering, as in scikit-learn SpectralClustering.

Also generally adapt the ht.Spectral API to scikit-learn.

Additional context
Related to parallelization efforts within the SCIMES project with @dcolombo

[ClaudiaComito]

We will work on this issue.

Keep in mind that the eigensolver is based on Lanczos at the moment (ht.linalg.lanczos), on small datasets it will be inefficient. A more efficient distributed eigensolver is in the making by @mrfh92

[github-actions]

Branch features/1740-Support_pre-calculated_affinity_matrix_in_ht_cluster_Spectral created!

[ClaudiaComito]

Hi @dcolombo,

I've updated Spectral on the branch features/1740-etc. After a git pull you should have:

• An API that's a bit more similar to scikit-learn (SpectralClustering instead of Spectral, affinity instead of metric)
• the option to set affinity = 'precomputed'

Keep in mind that the matrix decomposition on Heat is currently done with the Lanczos method, so you'll definitely get different results wrt scikit-learn.

[ClaudiaComito]

Still to do:

• if affinity == precomputed, sanitize input matrix (must be square and symmetric)
• TBD

[github-actions]

This issue is stale because it has been open for 60 days with no activity.