change the way arbitrary similarities are handled

pneuvial / adjclust

Adjacency-constrained hierarchical clustering of a similarity matrix

https://pneuvial.github.io/adjclust/

16 stars 8 forks source link

Closed tuxette closed 1 year ago

tuxette commented 6 years ago

make the clustering without modifying the similarity increase the gains to ensure positivity

pneuvial commented 6 years ago

Duplicate of #19

tuxette commented 6 years ago

When a similarity is not a kernel, we add a positive number to the diagonal to make it a kernel. Doing so:

clustering is not changed
height are translated from this number but:
model selection is modified
the different types of height in the dendrogram are also modified (probably increase with the number of merges) However, model selection has no sense when the similarity is not a kernel so this maybe sounds reasonable.

tuxette commented 2 years ago

Add this information when plotting a message on matrix transformation ~ "model selection may not be correct"