run clustering on KG embeddings

[kdutia]

• [x] try GPU-accelerated DBSCAN using cuml (link)
• [x] make notebook to explore clusters generated from above
• [x] understand parameters and run better parameter tuning

• [x] add to visualisation

• [ ] refine parameters
• [ ] re-add to visualisation
• [ ] ask for feedback

[kdutia]

Closing this. An analysis of cluster assignments vs curators' category assignments would be interesting later on, if there's time and it made an interesting write-up

[kdutia]

After further tweaking the min_cluster_size and min_samples parameters (see docs), there seems to be no 'better' version of clustering. I think that this highlights we need a more use case driven approach to clustering.

In the future it may be useful to try this again with a separate set of embeddings where collection categories have been removed as a property, as at the moment the layout roughly represents the structure of the collections.