Item-based no neighbourhood

[Evgeny Karataev <ka...@gmail.com>]

Hello,

Are there any particular reasons, why the implementation of item-based
recommendations does not support item neighborhood formation, e.g. based on
similarity threshold or by providing a number for the most similar items to
be used?

I know that there is linear interpolation item-based recommendation, but it
seems a little bit different than just classical item-based recommendation.
And anyway it does not support threshold based neighborhood.

Thank you.

-- 
Best Regards,
Evgeny Karataev

Re: Item-based no neighbourhood

[Sebastian Schelter <ss...@apache.org>]

I think it would definitely be a useful addition. The
org.apache.mahout.cf.taste.impl.recommender.BiasedItemBasedRecommender
already has this functionality e.g.

/s


On 27.11.2012 03:59, Evgeny Karataev wrote:
> Hello,
> 
> Are there any particular reasons, why the implementation of item-based
> recommendations does not support item neighborhood formation, e.g. based on
> similarity threshold or by providing a number for the most similar items to
> be used?
> 
> I know that there is linear interpolation item-based recommendation, but it
> seems a little bit different than just classical item-based recommendation.
> And anyway it does not support threshold based neighborhood.
> 
> Thank you.
> 

Re: Item-based no neighbourhood

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

Zhipeng's first answer is probably closest to a reason. You could do
so, but it is of less use. You are already beginning from a
constrained set of items, and neighborhood*s* around each of them may
be too sparse to be useful. Not to mention that there are many of
them.

But the neighborhood process is really just considering things with
low similarity to have no weight (similarity). And something like that
exists via CandidateItemStrategy.

On Tue, Nov 27, 2012 at 2:59 AM, Evgeny Karataev
<ka...@gmail.com> wrote:
> Hello,
>
> Are there any particular reasons, why the implementation of item-based
> recommendations does not support item neighborhood formation, e.g. based on
> similarity threshold or by providing a number for the most similar items to
> be used?
>
> I know that there is linear interpolation item-based recommendation, but it
> seems a little bit different than just classical item-based recommendation.
> And anyway it does not support threshold based neighborhood.
>
> Thank you.
>
> --
> Best Regards,
> Evgeny Karataev

Re: Item-based no neighbourhood

[zhipeng zhao <zh...@yahoo.com>]

Hi,

I just begin to learn Mahout and have not got the chance to read the code. But according to the book "Mahout in Action", first paragraph on page 59, also the psedo-code on page 57 for the algorithm, the user already has a list of itmes of which he has preference, thus implicitly define the "neighbor" for the target item. (Another way to state it is, for user-based recommendation, since usually the number of user is very large, we need to have a limited number of similar user to model the target user. However, for item based recommendation, for a perticular user, the items he shows references is already limited, thus we will use all of them as the neighbor for the target item.)

Just my 2 cents
-zhipeng 
--- On Mon, 11/26/12, Evgeny Karataev <ka...@gmail.com> wrote:

> From: Evgeny Karataev <ka...@gmail.com>
> Subject: Item-based no neighbourhood
> To: "user" <us...@mahout.apache.org>
> Date: Monday, November 26, 2012, 9:59 PM
> Hello,
> 
> Are there any particular reasons, why the implementation of
> item-based
> recommendations does not support item neighborhood
> formation, e.g. based on
> similarity threshold or by providing a number for the most
> similar items to
> be used?
> 
> I know that there is linear interpolation item-based
> recommendation, but it
> seems a little bit different than just classical item-based
> recommendation.
> And anyway it does not support threshold based
> neighborhood.
> 
> Thank you.
> 
> -- 
> Best Regards,
> Evgeny Karataev
>