motiwari

7 ms·

1.
by motiwari 3y ago
Hmmmm.... not sure exactly what you mean. I believe the setting you're describing is: we have a dataset we're trying to fit with a GMM, we know the number of components k, and we're trying to determine the parameters of the G
2.
by motiwari 3y ago
Definitely possible, but it would require some extensions to the algorithm. More specifically, as new datapoints enter the stream, they could be compared with the existing medoids to see if swapping them would lower the clustering loss. Thi
3.
by motiwari 3y ago
Where is this benchmark from? We'd be happy to run BanditPAM on these datasets and report the results
4.
by motiwari 3y ago
A strength if you're strictly looking to minimize squared L2 loss from each point to its closest mean -- but for a lot of other applications, it's a weakness! As the other poster mentioned, with KMedoids you can use arbitrary loss
5.
by motiwari 3y ago
You got me! As @dang mentioned, once I got feedback from him I was allowed to repost and enter the second-chance pool (My real first post was submitted too hastily without receiving @dang's feedback)
6.
by motiwari 3y ago
Hey, thanks! Shout out to Daniel @ HN who gave me a lot of great feedback on how to make this post better.
7.
by motiwari 3y ago
Thanks for bug report and repro steps! I've filed this issue at https://github.com/motiwari/BanditPAM/issues/244 on our repo. I suspect that this is because the scikit-learn implementation of KMeans subs
8.
by motiwari 3y ago
Right! Though to clarify a few nits: we use successive elimination instead of successive halving. And we talk about the Maximum Inner Product Search problem (very similar to NN problem) in our followup work: https://ar5iv.org&#x2
9.
by motiwari 3y ago
We talk exactly about clustering pictures in our blog post! https://ai.stanford.edu/blog/banditpam/
10.
by motiwari 3y ago
One thing that's important to note is that k-medoids supports arbitrary distance metrics -- in fact, your dissimilarity measure need not even be a metric (it can be negative, asymmetric, not satisfy the triangle inequality, etc.) An im
11.
by motiwari 3y ago
Interesting, thanks for the references! I'm not too familiar with this line of work; let me read up on it and get back to you
12.
by motiwari 3y ago
The elbow method is pretty common! https://en.wikipedia.org/wiki/Elbow_method_(clustering) You can also use some regularization criterion (AIC, BIC, or other)
13.
by motiwari 3y ago
Oh cool, neat trick, thanks!
14.
by motiwari 3y ago
Thank you for the positive feedback! Will definitely take you up on that coffee next time I visit Ilan :) To answer your questions: 1. When looking at k-medoids algorithms, we realized that PAM hadn't been improved since the 80s! Other
15.
by motiwari 3y ago
Really funny that you mention that! Some of our more recent work focuses on using adaptive sampling techniques in approximate-nearest-neighbor search (actually, the related problem of maximum inner product search: https://ar5iv.o
16.
Show HN: Want something better than k-means? Try BanditPAM(github.com)
281 points by motiwari 3y ago|41 comments
17.
by motiwari 3y ago
Sorry for the late response. And actually, despite the plethora of clustering algorithms, k-means is still a very commonly-used out-of-the-box technique! (Probably for its simplicity and existing popularity)
18.
by motiwari 3y ago
Sorry, I don't understand your argument. Could you clarify what you mean by "everything"? Is there another clustering algorithm you're thinking of?
19.
by motiwari 3y ago
Want something better than k-means? I'm happy to announce our SOTA k-medoids algorithm from NeurIPS 2020, BanditPAM, is now publicly available! `pip install banditpam` or `install.packages("banditpam")` and you're good t
20.
Show HN: Want something better than k-means? Try BanditPAM(github.com)
7 points by motiwari 3y ago|5 comments