hivemall-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Shadi Mari <shadim...@gmail.com>
Subject Re: Early Stopping / FFMs performance
Date Tue, 10 Oct 2017 16:10:18 GMT
Hi Makoto,

Thanks for your valuable feedback.

I am using Criteo 2014 dataset for CTR prediction, which is 45M examples in
total. Do you think 8 hours is still resonable training duration given than
I am using your  EMR configurations? I never assumed this can be such time
consuming.

I already built a version from the master branch. As per your feedback, i
assume FFM implementation can not yet be used in production! correct?

Regards,


On Tue, Oct 10, 2017 at 7:03 PM, Makoto Yui <myui@apache.org> wrote:

> Hi Shadi,
>
> 2017-10-10 23:38 GMT+09:00 Shadi Mari <shadimari@gmail.com>:
> > My first question, Do you support Early stopping? If not, Does L1/L2
> really
> > helps with FFMs to prevent overfitting?
>
> Early stopping is not yet supported.
> Libffm need to provide validation dataset in addition to training
> dataset in training and it's not good at usability.
>
> Instead of early stopping, Hivemall's FFM uses FTRL-proximal (L1/L2
> regularization + adagrad-like adaptive learning rate) [1] to avoid
> overfitting.
> [1] https://static.googleusercontent.com/media/
> research.google.com/ja//pubs/archive/41159.pdf
>
> FTRL is known to work well for CTR prediction tasks with (FM-like)
> feature interactions.
> https://www.kaggle.com/sudalairajkumar/ftrl-starter-with-leakage-vars/code
>
> > The second question is, I am trying to teach an FFM model using Hive over
> > Tez distributed over two nodes with total 56 cores and 165 GB memory,
> > however the process is taking 8 hours on 10 iterations and FTRL
> optimizer.
> > Does this sound reasonable ? or there is something has to be optimized to
> > shorten the duration.
>
> What's the size of your training dataset?
>
> I'm using the following settings for FFM evaluation on EMR.
> https://gist.github.com/myui/aaeef548a17eb90c4e88f824c3ca1bcd
>
> Last but not least, FFM implementation is merged to master but still
> beta and WIP.
> Thus not documented and stealth mode yet.
>
> I need to revise V initialization scheme of FFM, default
> hyperparameters, regularization scheme, and so on.
>
> Thanks,
> Makoto
>
> --
> Makoto YUI <myui AT apache.org>
> Research Engineer, Treasure Data, Inc.
> http://myui.github.io/
>

Mime
View raw message