mahout-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ted Dunning <ted.dunn...@gmail.com>
Subject Re: What's the best practice to specify df for TPrior and alphaByLambda for ElasticBandPrior
Date Mon, 06 Jun 2011 11:39:42 GMT
Here is the link:

http://www.stat.columbia.edu/~gelman/research/unpublished/*prior*s7.pdf


On Mon, Jun 6, 2011 at 4:39 AM, Ted Dunning <ted.dunning@gmail.com> wrote:

> This paper has good guidance.
>
> I think that their recommendation of df=1 and scale=2.5 is very reasonable.
>  Their suggestion of mean subtraction would be disastrous, but much of the
> rest is very good.
>
> df=0.5 is a very (VERY) long tailed distribution.  Even for df=1, there is
> no mean or variance.
>
> I would actually consider using df=2 as well.
>
> In practice, I have found that the L1 prior encourages sparsity more
> strongly than the T prior.  It does, however, entangle the regularization
> factor with the scale of the prior.
>
>
> On Mon, Jun 6, 2011 at 1:06 AM, XiaoboGu <guxiaobo1982@gmail.com> wrote:
>
>> Is 0.5 a good point to start?
>>
>>
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message