Stacking Algorithms

[jeff thomas <mr...@yahoo.com>]

Can anyone please provide details on how to stack algorithms?

From: http://lucene.472066.n3.nabble.com/A-few-questions-regarding-content-based-recommenders-td2168516.html


"Production recommendation systems use several algorythms and combine 
them with weights. This is called 'stacking'. You might wish to write 
a stacking version of Recommender. "

Re: Stacking Algorithms

[Sean Owen <sr...@gmail.com>]

I think this just refers to computing an average or weighted average over
several algorithms, and using that for the basis of recommendation. It's not
supported directly, but fairly easy to roll yourself.

Beware that this only makes sense if the values you are averaging are of the
same scale. The conventional algorithms are based on an estimate of rating
for unknown items, so it makes sense to average those estimates. But
recommenders don't have to rank on this, and if they're ranking on some
other value it may not be valid to average those.

On Mon, Jun 6, 2011 at 9:27 PM, jeff thomas <mr...@yahoo.com> wrote:

> Can anyone please provide details on how to stack algorithms?
>
> From:
> http://lucene.472066.n3.nabble.com/A-few-questions-regarding-content-based-recommenders-td2168516.html
>
>
> "Production recommendation systems use several algorythms and combine
> them with weights. This is called 'stacking'. You might wish to write
> a stacking version of Recommender. "
>

Re: Stacking Algorithms

[Lance Norskog <go...@gmail.com>]

Good god, you're quoting ME?

One application of this is using several statistical recommenders,
weighted. Another use case is that the GroupLens dataset comes with
movie categories; combining a statistical recommender and a genre
database sounds fun.

On Mon, Jun 6, 2011 at 1:53 PM, Steven Bourke <sb...@gmail.com> wrote:
> Try something simple like this
>
> 1 - Identify which recommendation configurations you believe work best with
> whatever data you currently have.
> 2 - Run said configurations across your data, track the recommendation list
> generated by each configuration.
> 3 - Combine the scores for items across your various recommendation lists
> 4 - Present results to whomever
>
> You should probably read the quickstart guide to get familiar with how to
> set the recommendation aspect of mahout up (Check the wiki and examples in
> trunk)
>
> On Mon, Jun 6, 2011 at 9:27 PM, jeff thomas <mr...@yahoo.com> wrote:
>
>> Can anyone please provide details on how to stack algorithms?
>>
>> From:
>> http://lucene.472066.n3.nabble.com/A-few-questions-regarding-content-based-recommenders-td2168516.html
>>
>>
>> "Production recommendation systems use several algorythms and combine
>> them with weights. This is called 'stacking'. You might wish to write
>> a stacking version of Recommender. "
>>
>



-- 
Lance Norskog
goksron@gmail.com

Re: Stacking Algorithms

[Steven Bourke <sb...@gmail.com>]

Try something simple like this

1 - Identify which recommendation configurations you believe work best with
whatever data you currently have.
2 - Run said configurations across your data, track the recommendation list
generated by each configuration.
3 - Combine the scores for items across your various recommendation lists
4 - Present results to whomever

You should probably read the quickstart guide to get familiar with how to
set the recommendation aspect of mahout up (Check the wiki and examples in
trunk)

On Mon, Jun 6, 2011 at 9:27 PM, jeff thomas <mr...@yahoo.com> wrote:

> Can anyone please provide details on how to stack algorithms?
>
> From:
> http://lucene.472066.n3.nabble.com/A-few-questions-regarding-content-based-recommenders-td2168516.html
>
>
> "Production recommendation systems use several algorythms and combine
> them with weights. This is called 'stacking'. You might wish to write
> a stacking version of Recommender. "
>