Consider these three questions (in chronological order):

  1. What exactly is a hyperparameter?
  2. What do we mean by hyperparameters?
  3. What's in a name: hyperparameters

I think it's pretty clear that they are exact duplicates (currently only #2 is closed as a dup of #1). However, all three contain excellent answers. I would therefore suggest to merge all three questions into one of them.

I think the best candidate for the "master thread" is #3, because it has most answers, most upvotes, and most views, despite being the newest.

I flagged these threads for mod attention, but as per @Glen_b's suggestion I post this on meta too so that everybody could take a look and vote. Please upvote the question iff you agree that merging is a good idea.


I think one potential difficulty is that the one that references Pattern Recognition and Machine Learning* seems to be dealing with a different intent than the other two questions.

*(presumably Bishop's book, though it's not the only one by that title)

Specifically, the two questions that are already duplicates seem to be specifically dealing with what I'd have called a hyperparameter, but Bishop is using it for what I'd call a regularization parameter or a tuning parameter and for which I usually wouldn't say "hyperparameter" unless it was cast as a parameter in a hierarchy (e.g. if it was cast in a Bayesian framework, you might be able to make it a parameter of a distribution from which observation parameters were drawn).

It is this difference that made me hesitate to merge them; I think the differences in the usage makes good answers to the other questions poor answers to that question on Bishop's use of the term.

If we rewrote the question the answers were merged into, or edited some of the answers, we might be able to make it work tolerably well. An alternative (and perhaps a better one) would be to link the questions (perhaps noting the difference).

  • 1
    $\begingroup$ Okay, I can see this reasoning. I appreciate that there seem to be two rather distinct meanings of "hyperparameter" (HP): roughly speaking, Bayesian HP and ML HP, and that Qs #1 and #2 ask about Bayesian HPs, whereas #3 seems to be asking about ML ones. However, the answers are all over the place. DikranMarsupial's answer (which has max number of upvotes in all these threads) to #2 talks mostly about ML HPs, even though it mentions both meanings. And the most upvoted answer to #3 seems to talk about Bayesian HPs. My suggestion was driven by this confusion. $\endgroup$ – amoeba says Reinstate Monica May 18 '16 at 16:11
  • 1
    $\begingroup$ If we want to keep them separate, then we should edit the titles of the Qs to make it clear that they are asking about different things. And live with the fact that some of the answers are assuming different meaning. This can also work "tolerably well", but it still seems to me that there will be less confusion if everything is merged together and then perhaps some of the answers edited to make the distinction very clear. $\endgroup$ – amoeba says Reinstate Monica May 18 '16 at 16:13
  • $\begingroup$ @amoeba If we keep the questions separate, I would not lean toward changing the titles (since the titles people chose may be similar to the titles people use to search) but would make the distinction clear when linking the posts. Specifically, people searching won't know about the distinction in usages. $\endgroup$ – Glen_b -Reinstate Monica May 18 '16 at 22:02
  • 1
    $\begingroup$ If we combined them, some editing would seem to be necessary to avoid confusing the reader with the answers relating to different meanings without highlighting that there are different meanings. If done well, the combine-and-edit would probably end up with a result that is more useful to later readers, since explanation of the existence and meaning of the different interpretations would all be explicit and in one place. $\endgroup$ – Glen_b -Reinstate Monica May 18 '16 at 22:06

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .