The library is great (as usual with your work). I wonder whether there are plans to augment the implementation with DensMAP support to preserve local data density. This is critical if we are to use the reduced results for clusterization - with UMAP the same embedding (for example from ada-002) is placed a tad differently in the reduced dimensionality every time:
numberOfNeighbors = 15, K-Means++ clustering
This can be, to an extent mitigated with hyperparameter fine-tuning:
numberOfNeighbors = 80, DBSCAN clustering
DensMAP is an extension and if I'm not missing something, seems to be implemented only in two files:
The library is great (as usual with your work). I wonder whether there are plans to augment the implementation with DensMAP support to preserve local data density. This is critical if we are to use the reduced results for clusterization - with UMAP the same embedding (for example from ada-002) is placed a tad differently in the reduced dimensionality every time:
This can be, to an extent mitigated with hyperparameter fine-tuning:
numberOfNeighbors = 80, DBSCAN clustering
DensMAP is an extension and if I'm not missing something, seems to be implemented only in two files: