[FEAT] Different clustering algorithms besides Connexted Components

[NickCrews]

[FEAT] Different clustering algorithms besides Connexted Components

[RossKen]

Hey @NickCrews - thanks for raising this issue.

This brings up some interesting questions about how people use Splink and what problems they are solving. We have discussed this internally in the past and concluded that the explainability of connected components is one of its key strengths. Given we, primarily, use Splink in a government setting, that transparency is particularly important to us.

We are open to the idea of using different clustering methodologies, particularly if there is a performance boost, but up to this point there have always been other developments which we have seen as higher priority.

I think this taps into a wider point on how we measure performance more generally. We are aware that we have various metrics and methods for assessing performance at the pairwise comparison level, but not very much at the cluster level. This is something we aim to tackle this year, so how we create clusters may well get looked at in more depth at that stage.