I've figured out that I can use the graph model to extract neighbourhoods, and then use those neighbourhoods to label the umap clustering and tune the hyperparameters to make sense with respect to those neighbourhoods.
This seems like the best way to arrange the umap embeddings so the global structure reflects the neighbourhoods we've found.
I've figured out that I can use the graph model to extract neighbourhoods, and then use those neighbourhoods to label the umap clustering and tune the hyperparameters to make sense with respect to those neighbourhoods.
This seems like the best way to arrange the umap embeddings so the global structure reflects the neighbourhoods we've found.