You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@mahout.apache.org by Loic Descotte <lo...@kelkoo.com> on 2011/09/12 15:52:05 UTC
SGD/SVM classification : minimum dataset size for training
Hi all,
My classification problem is very similar to the "20 newsgroups"
example. But I don't have the possibility to use a large quantity of
data for training.
I'd like to know what would be the "minimum" size of training data for
SGD or SVM algorithms to have reasonable results.
My datas are the same kind of the 20 newsgroups example but they have
fewer lines.
The body of each entry is about beetween 40 and 60 words.
I'd like to try with 10 examples by category (with 2 or 3 category),
choosing good examples with the more frequent keywords to be sure that
the learning phase will be efficient.
Can it be relevant with so little data ?
Thanks,
Loic
Re: SGD/SVM classification : minimum dataset size for training
Posted by Loic Descotte <lo...@kelkoo.com>.
OK Thanks
I'll try both and I'll tell you how it works :)
Loic
Le 12.09.2011 22:24, Ted Dunning a écrit :
> SVM is reasonable.
>
> SGD with hand-tuning of the learning parameters may work.
>
> With so little training data, you will have a difficult assessing whether
> your system is working.
>
> Sometimes, you can rephrase your problem so that all of your training data
> across many situations can be pooled together. There is a nice paper on
> google priority mail about just such an example where Google used
> meta-features so that they could train a few models across all users
>
> On Mon, Sep 12, 2011 at 6:52 AM, Loic Descotte<lo...@kelkoo.com>wrote:
>
>> My classification problem is very similar to the "20 newsgroups" example.
>> But I don't have the possibility to use a large quantity of data for
>> training.
>> ...
>> I'd like to try with 10 examples by category (with 2 or 3 category),
>> choosing good examples with the more frequent keywords to be sure that the
>> learning phase will be efficient.
>>
>> Can it be relevant with so little data ?
>>
>>
Re: SGD/SVM classification : minimum dataset size for training
Posted by Ted Dunning <te...@gmail.com>.
SVM is reasonable.
SGD with hand-tuning of the learning parameters may work.
With so little training data, you will have a difficult assessing whether
your system is working.
Sometimes, you can rephrase your problem so that all of your training data
across many situations can be pooled together. There is a nice paper on
google priority mail about just such an example where Google used
meta-features so that they could train a few models across all users
On Mon, Sep 12, 2011 at 6:52 AM, Loic Descotte <lo...@kelkoo.com>wrote:
>
> My classification problem is very similar to the "20 newsgroups" example.
> But I don't have the possibility to use a large quantity of data for
> training.
> ...
> I'd like to try with 10 examples by category (with 2 or 3 category),
> choosing good examples with the more frequent keywords to be sure that the
> learning phase will be efficient.
>
> Can it be relevant with so little data ?
>
>