[scikit-learn] Artificial neural network not learning lower values of the training sample

[chris brew]

Experienced machine learning people usually start by trying to exactly
replicate what the paper did, using
exactly the same data, and exactly the same methods, and if possible, even
exactly the same software. It is very comforting
if you can do this, because you can then go ahead and make changes, secure
in the knowledge that the reasons for any
changes in output are due to the changes you made, and not to small
mistakes in data preparation, tool arguments or something.

So it is a good plan to start by seeing whether you can arrange to have the
same setup as the writers of the original paper. It is
polite to try to do this on your own if you can (and the effort is a useful
learning experience, usually) Maybe the data and/or software is publicly
available. If not, authors are often willing and able to share material
privately, so you could write to them and ask for help. Since you are
showing interest in their work, they will want to help if they can. And if
they do, you have the beginnings of a useful personal connection. The only
pre-condition is that you should (again as matter of politeness) first make
an honest attempt to replicate their work, and make sure you really
understand it. After you have done that, asking for help is appropriate.

Generally speaking, the time to ask questions on a public forum like
scikit-learn is also after you have done all you can to solve your problem.

Best wishes

Chris




On 1 June 2016 at 13:05, muhammad waseem <m.waseem.ahmad at gmail.com> wrote:

> It's not the same data (different locations) but I have tried to use same
> input and output variables.
>
> Thanks
> Waseem
>
> On Wed, Jun 1, 2016 at 12:02 PM, Andrew Holmes <andrewholmes82 at icloud.com>
> wrote:
>
>> A previous commenter asked about the published research you mentioned in
>> which this was working ok. If you’re using the same data as them, could you
>> try to replicate their results first?
>>
>> Best wishes
>> Andrew
>>
>> @andrewholmes82 <http://twitter.com/andrewholmes82>
>>
>>
>>
>>
>>
>>
>>
>>
>> On 31 May 2016, at 20:05, muhammad waseem <m.waseem.ahmad at gmail.com>
>> wrote:
>>
>> I try to balance it out, the dataset is very periodic type (similar
>> behaviour in an year)
>>
>> On Tue, May 31, 2016 at 8:01 PM, Andrew Holmes <andrewholmes82 at icloud.com
>> > wrote:
>>
>>> Is the training set unbalanced between high and low values? Ie, many
>>> more of the high ones?
>>>
>>> Best wishes
>>> Andrew
>>>
>>> @andrewholmes82 <http://twitter.com/andrewholmes82>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>> On 31 May 2016, at 20:00, muhammad waseem <m.waseem.ahmad at gmail.com>
>>> wrote:
>>>
>>> Yes, it has poor performance (higher errors) on lower values.
>>> I have tried random forest but as I mentioned it did not give good
>>> results either, I can try SVR.
>>>
>>> Kindest Regards
>>> Waseem
>>>
>>> On Tue, May 31, 2016 at 6:54 PM, Andrew Holmes <
>>> andrewholmes82 at icloud.com> wrote:
>>>
>>>> When you say it’s not learning ‘lower values’, does that mean the model
>>>> has good predictions on high values in the test set, but poor performance
>>>> on the low ones?
>>>>
>>>> Have you tried simpler models like tree, random forest and svm as a
>>>> benchmark?
>>>>
>>>> Best wishes
>>>> Andrew
>>>>
>>>> @andrewholmes82 <http://twitter.com/andrewholmes82>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>> On 31 May 2016, at 16:59, Andrew Holmes <andrewholmes82 at icloud.com>
>>>> wrote:
>>>>
>>>> If the problem is that it’s confusing day and night, are you including
>>>> time of day as a parameter?
>>>>
>>>> Best wishes
>>>> Andrew
>>>>
>>>> @andrewholmes82 <http://twitter.com/andrewholmes82>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>> On 31 May 2016, at 16:55, muhammad waseem <m.waseem.ahmad at gmail.com>
>>>> wrote:
>>>>
>>>> Hi All,
>>>> I am trying to train an ANN but until now it is not learning the lower
>>>> values of the training sample. I have tried using different python
>>>> libraries to train ANN. The aim is to predict solar radiation from other
>>>> weather parameters (regression problem). I think the ANN is confusing lower
>>>> values (winter/cloudy days) with the night-time values (probably). I have
>>>> tried the following but none of them worked;
>>>>
>>>>  1. Scaling data between different values e.g. [0,1],[-1,1]
>>>>  2. Standardising data to have zero mean and unit variance
>>>>  3. Shuffling the data
>>>>  4. Increasing the training samples (from 3 years to 10 years)
>>>>  5. Using different train functions
>>>>  6. Trying different transfer functions
>>>>  7. Using few input variables
>>>>  8. Varying hidden layers and hidden layers' neurons
>>>>
>>>> Any idea what could be wrong or any directions to try?
>>>>
>>>> Thanks
>>>> Kindest Regards
>>>> Waseem
>>>> _______________________________________________
>>>> scikit-learn mailing list
>>>> scikit-learn at python.org
>>>> https://mail.python.org/mailman/listinfo/scikit-learn
>>>>
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> scikit-learn mailing list
>>>> scikit-learn at python.org
>>>> https://mail.python.org/mailman/listinfo/scikit-learn
>>>>
>>>>
>>> _______________________________________________
>>> scikit-learn mailing list
>>> scikit-learn at python.org
>>> https://mail.python.org/mailman/listinfo/scikit-learn
>>>
>>>
>>>
>>> _______________________________________________
>>> scikit-learn mailing list
>>> scikit-learn at python.org
>>> https://mail.python.org/mailman/listinfo/scikit-learn
>>>
>>>
>> _______________________________________________
>> scikit-learn mailing list
>> scikit-learn at python.org
>> https://mail.python.org/mailman/listinfo/scikit-learn
>>
>>
>>
>> _______________________________________________
>> scikit-learn mailing list
>> scikit-learn at python.org
>> https://mail.python.org/mailman/listinfo/scikit-learn
>>
>>
>
> _______________________________________________
> scikit-learn mailing list
> scikit-learn at python.org
> https://mail.python.org/mailman/listinfo/scikit-learn
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/scikit-learn/attachments/20160601/f7721eb7/attachment-0001.html>