Hi all,

After a daunting struggle, I was able to corner the issue with the poor
accuracy for the specific leaf dataset. The dataset has classes from 1 to
36. However, there are no classes from 16th - 22nd. i.e. Classes go as
1,2,..,14,15,23,24,...,35,36

Then, while converting these class labels to enums in H-2-O (combined with
the fact that there's very little data for each class) confuses H-2-O and
causes it to *assign different enum values for the same classes in
different datasets*. Which manifest itself as a poor accuracy.

I suspect that there's a mismatch between the labels provided by JavaRDD
and enums produced by H-2-O as well. I'm looking into this issue right now.

Thank you

On Mon, Aug 10, 2015 at 11:16 AM, Thushan Ganegedara <[email protected]>
wrote:

> Hi all,
>
> I've been testing the new Deeplearning component with few different
> datasets (mainly leaf dataset) and the leaf dataset seems to be not working
> as expected for an unknown reason.
>
> However, I tested the Deeplearning component extensively with the leaf
> dataset and identified several potential problems that might be causing the
> poor accuracy.
>
> 1. Need to have higher number of epochs (compared to other datasets) to
> produce a reasonable accuracy.
>
> 2. Too many neurons causing overfitting thereby causing poor accuracy.
>
> 3. Some classes have quite closely related features (Especially the latter
> classes are misclassified often)
>
> I was able to get an accuracy of 86% with Logistic Regression L-BFGS.
> Which is quite reasonable. But I'm having trouble reaching that accuracy
> with Deeplearning (which should be quite easy). Highest accuracy I reached
> so far is 71.xx%
>
> So I'm still looking for any definite issues causing the poor accuracy.
>
> Thank you.
>
>
> --
> Regards,
>
> Thushan Ganegedara
> School of IT
> University of Sydney, Australia
>



-- 
Regards,

Thushan Ganegedara
School of IT
University of Sydney, Australia
_______________________________________________
Dev mailing list
[email protected]
http://wso2.org/cgi-bin/mailman/listinfo/dev

Reply via email to