05:39 < jenkins-mlpack> Project mlpack - nightly matrix build build #471: STILL UNSTABLE in 1 hr 38 min: http://big.cc.gt.atl.ga.us:8080/job/mlpack%20-%20nightly%20matrix%20build/471/

05:39 < jenkins-mlpack> * andrewmw94: added code for accessing immediate child nodes (need to think of a way to rename this to be less confusing). Some more quasi-code to split nodes and insert points.

05:39 < jenkins-mlpack> * andrewmw94: R tree stuff

05:39 < jenkins-mlpack> * andrewmw94: more R Tree stuff

11:58 < Anand> Hi Marcus!

11:58 < Anand> I was thinking over the class conversion problem

12:01 < Anand> Since I already have the confusion matrix, irrespective of the number of classes, I think we can easily use the one vs all approach as I have already done for basic metrics like accuracy, precision

12:02 < Anand> I was just wondering if it is correct to employ the same approach to all metrics

12:02 < Anand> Actually, multi class conversion is still an area of active research

12:02 < Anand> I am exploring more but, I need your inputs.

12:03 < Anand> I think the one vs all approach should work for all metrics. But before doing that I need to confirm things

12:03 < Anand> What do you say?

12:08 < marcus_zoq> Anand: Hello!

12:08 < marcus_zoq> Anand: I think the conversion problem is the tricky part. You are right maybe the one vs all approach works for all metrics. I have to think about it a little bit more, but right now I don't see any reason why this shouldn't work.

12:08 < Anand> Same for me.

12:08 < Anand> I want to do it for all metrics but I am not sure

12:09 < Anand> I did some readings but didn't get anything specific

12:13 < marcus_zoq> I'll have a look and get back to you tonight or tomorrow. If this is okay?

12:14 < Anand> Yeah fine. I am also doing some more research. Will finalize tomorrow after having your opinion

12:14 < marcus_zoq> Yeah, sounds good.

12:44 < oldbeardo> naywhayare: I came across a problem while testing the code

12:45 < oldbeardo> if I pass the transpose of the GroupLens100k dataset, it behaves quite erratically, errorwise

12:47 < oldbeardo> Theorem 1 of the paper states that it is an unbiased estimate of the projection

12:47 < oldbeardo> which may be true theoretically but depends a lot on the number of samples that I use to estimate it

12:49 < oldbeardo> I tried using '1000 * log(num_points)' samples and it gives a good result

12:49 < oldbeardo> but that defeats the whole purpose, as it becomes too slow

15:50 < naywhayare> oldbeardo: I spent a little time thinking about this, but with my paper deadline coming up I don't really have time to dig into it too much

15:50 < naywhayare> so, let's do this --

15:50 < naywhayare> for now, ignore the cosine tree tests, and go ahead and implement quic-svd using the cosine tree you've written

15:51 < naywhayare> writing a test for quic-svd shouldn't be hard... just check the residual of the matrix as reconstructed from the SVD (|| A - A' ||_F)

15:51 < naywhayare> and make sure that quantity is small

15:51 < naywhayare> then we can worry about how to test the cosine tree later, after June 6th

15:51 < naywhayare> does that seem reasonable?

15:52 < oldbeardo> naywhayare: sure, I will do that

15:56 < naywhayare> okay. I'm sorry that I don't have more time to devote to this right now

15:56 < naywhayare> literally all I do now is wake up, eat, write paper, eat, write paper, eat, write paper, sleep

16:00 < oldbeardo> naywhayare: sorry about that

16:01 < oldbeardo> I read your messages on the chat history

16:01 < oldbeardo> it's okay, you have a deadline coming up

16:02 < oldbeardo> when's the deadline by the way?

16:08 < naywhayare> June 6th, so, this upcoming friday

20:26 < sumedhghaisas> naywhayare: Hey ryan, you free??

22:21 < naywhayare> sumedh_: I'm sorry, I didn't get your message until now, and I'm about to leave for dinner now

22:21 < sumedh_> Its okay... msg me if you are free after dinner :)