This is just a quick comment on a recent debate in Philosophy & Technology between David Watson and Tom Sterkenburg. They discuss if clustering methods, nowadays classed as a type of machine learning, can find natural kinds.
The point of contention is an epistemic claim by Watson:
Some clustering algorithms can identify some natural kinds.1
Sterkenburg basically replies: No they don’t. (Because no general guarantees can be given)
To which Watson answers that his original intention was just to show “that clustering algorithms (or something very much like them) are an essential component of any effort to identify natural kinds.”
I sure hope that the clustering algorithms I use are the ones that identify natural kinds!
Maybe we should not overburden innocent clustering methods with epistemic demands. Maybe we should discuss a constructive claim instead:
Some clustering algorithms can construct some clusters.
Identifying the constructed clusters with natural kinds has then to be done by someone who is not a clustering algorithm. The epistemic burden is shifted. Such a shift is actually acknowledged by Watson who says that the results of a cluster analysis have to be “corroborated (or not) with other methods (e.g., testing for divergent survival trajectories among subtypes).” Corroborations are therefore not an essential part of the clustering method.
Links to the papers:
Watson’s paper on clustering
Sterkenburg’s reply
Watson’s reply to the reply
- To be fair this is the claim they seem to agree upon in the last iteration of the discussion. ↩︎