Github user srowen commented on the issue:
https://github.com/apache/spark/pull/15450
@sethah I should say I am not trying to handle cases where clusters start
separate and converge to nearly the same point. I don't that's something we
should even try to do.
To elaborate, here are the relevant cases, I think. I'm ignoring cases
where the data size is >> k because it all behaves as desired with high
probability already.
*Case 1*. Data = [A, B], k = 3, init = random
Now, you'll get 3 centroids and definitely have a duplicate. After this
change, you'd get 2.
*Case 2*. Data = [A, A, B, B], k = 3, init = random
Now, you'll get 3 centroids and definitely have a duplicate. After this
change, same thing.
*Case 3*. Data = [A, B, C, D], k = 3, init = random
Now, you'll get 3 centroids and occasionally have a duplicate. After, this
change, you'd get 3 distinct centroids always.
*Case 4*. Data = [A, B], k = 3, init = parallel
Now, you'll get 3 centroids and definitely have a duplicate. After this
change, you should get 2 if I read the code right (that it won't pick a
centroid at distance 0 from another).
*Case 5*. Data = [A, A, B, B], k = 3, init = parallel
Now, you'll get 3 centroids and definitely have a duplicate. After this
change, you should get 2 for the same reason above.
*Case 6*. Data = [A, B, C, D], k = 3, init = parallel
Now, you'll get 3 centroids and occasionally have a duplicate. After this
change, same thing.
Case 1/4/5 are, I think, positive changes. Case 3 seems like a clear win.
Case 2 could be made consistent with 1/4/5 easily, I suppose, by calling
`.distinct`. Case 6 is the interesting one -- can we make it manage to return 3
distinct centroids always because that's clearly possible? like case 3? I
punted on that though.
I think Derrick's case is like 2 or 5. There are duplicates in the data and
not many unique points. Filtering the duplicates early is a performance win. At
least -- that is the case I am trying to solve. If that's not what is described
in the JIRA then I think it's a different issue or question.
I think there's a benefit to speed and meaningfulness of the model in case
1/2/4/5.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
