You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@mahout.apache.org by Matthew Runo <ma...@gmail.com> on 2011/01/12 18:53:40 UTC

Some basic questions

Hello folks -

I've been reading this user list for many many months now - and it's
time that I go from being interested to actually getting started on
some mahout stuff! I'd like to say that I understand most of what you
talk about - but that'd be a lie. Forgive me if this question is
trivially obvious...

I've been starting with the collaborative filtering stuff in Taste,
using the ItemSimilarity logic. I think I have that working for the
most part, but now I'd like to do some "more advanced" things with my
product recommendations.

-- The Question --
Say I have several product types - books, shoes, kitchen stuff, and
automotive. What would be the best way to force a recommendation from
each product type? Would you build a BookRecommender, a
ShoeRecommender - and so on - and then pull n recommendations from
each recommender and combine?

If I did that - I'd split my preference data so that each product type
had a different set and go that way right? I assume that I'd loose the
benefit of "this book correlates strongly to this tire" by splitting
it out and so the cross-product-type recommendations would be a bit
worse... Is there a way to do this without loosing the link between
that book+tire?

Which recommender would you use as the base for this- KNN, Slope One,
SVD? I still need to read the papers on the algorithms, so feel free
to ignore this question if it'll become obvious!

Thank you for all your work on this library!

-Matthew Runo

Re: Some basic questions

Posted by Sean Owen <sr...@gmail.com>.
It does get called a lot. I'd cache the list of eligible IDs in
memory. A hundred thousand longs isn't bad. No I would not SELECT from
the database each time.

On Wed, Jan 12, 2011 at 7:05 PM, Matthew Runo <ma...@gmail.com> wrote:
> Ok, that makes sense. I like the user-based approach. One question though -
>
> Assuming that my product corpus is on the other of several hundred
> thousand products - wouldn't a database hit in a IDRescorer slow the
> system down terribly? The JavaDoc for Rescorer<T> leads me to believe
> that this would get called /a lot/.
>
> Would you take the hit and load a map itemId,prodcutTypeId or similar
> into memory and use that? Is it a non isssue?
>
> I appreciate the help as I architect this in my head..
>
> --Matthew
>

Re: Some basic questions

Posted by Matthew Runo <ma...@gmail.com>.
Ok, that makes sense. I like the user-based approach. One question though -

Assuming that my product corpus is on the other of several hundred
thousand products - wouldn't a database hit in a IDRescorer slow the
system down terribly? The JavaDoc for Rescorer<T> leads me to believe
that this would get called /a lot/.

Would you take the hit and load a map itemId,prodcutTypeId or similar
into memory and use that? Is it a non isssue?

I appreciate the help as I architect this in my head..

--Matthew

On Wed, Jan 12, 2011 at 10:17 AM, Sean Owen <sr...@gmail.com> wrote:
> You can simply partition your input into the book data, shoes data,
> etc. and recommend separately, and then interlace the results.
>
> Slightly smarter might be to use a user-based recommender based on all
> the data. Then, use an IDRescorer to filter out everything but books
> and recommend books. Repeat for other types as above, and interlace
> the results. This at least lets you leverage what you know about books
> to recommend shoes.
>
> On Wed, Jan 12, 2011 at 5:53 PM, Matthew Runo <ma...@gmail.com> wrote:
>> Hello folks -
>>
>> I've been reading this user list for many many months now - and it's
>> time that I go from being interested to actually getting started on
>> some mahout stuff! I'd like to say that I understand most of what you
>> talk about - but that'd be a lie. Forgive me if this question is
>> trivially obvious...
>>
>> I've been starting with the collaborative filtering stuff in Taste,
>> using the ItemSimilarity logic. I think I have that working for the
>> most part, but now I'd like to do some "more advanced" things with my
>> product recommendations.
>>
>> -- The Question --
>> Say I have several product types - books, shoes, kitchen stuff, and
>> automotive. What would be the best way to force a recommendation from
>> each product type? Would you build a BookRecommender, a
>> ShoeRecommender - and so on - and then pull n recommendations from
>> each recommender and combine?
>>
>> If I did that - I'd split my preference data so that each product type
>> had a different set and go that way right? I assume that I'd loose the
>> benefit of "this book correlates strongly to this tire" by splitting
>> it out and so the cross-product-type recommendations would be a bit
>> worse... Is there a way to do this without loosing the link between
>> that book+tire?
>>
>> Which recommender would you use as the base for this- KNN, Slope One,
>> SVD? I still need to read the papers on the algorithms, so feel free
>> to ignore this question if it'll become obvious!
>>
>> Thank you for all your work on this library!
>>
>> -Matthew Runo
>>
>

Re: Some basic questions

Posted by Sean Owen <sr...@gmail.com>.
You can simply partition your input into the book data, shoes data,
etc. and recommend separately, and then interlace the results.

Slightly smarter might be to use a user-based recommender based on all
the data. Then, use an IDRescorer to filter out everything but books
and recommend books. Repeat for other types as above, and interlace
the results. This at least lets you leverage what you know about books
to recommend shoes.

On Wed, Jan 12, 2011 at 5:53 PM, Matthew Runo <ma...@gmail.com> wrote:
> Hello folks -
>
> I've been reading this user list for many many months now - and it's
> time that I go from being interested to actually getting started on
> some mahout stuff! I'd like to say that I understand most of what you
> talk about - but that'd be a lie. Forgive me if this question is
> trivially obvious...
>
> I've been starting with the collaborative filtering stuff in Taste,
> using the ItemSimilarity logic. I think I have that working for the
> most part, but now I'd like to do some "more advanced" things with my
> product recommendations.
>
> -- The Question --
> Say I have several product types - books, shoes, kitchen stuff, and
> automotive. What would be the best way to force a recommendation from
> each product type? Would you build a BookRecommender, a
> ShoeRecommender - and so on - and then pull n recommendations from
> each recommender and combine?
>
> If I did that - I'd split my preference data so that each product type
> had a different set and go that way right? I assume that I'd loose the
> benefit of "this book correlates strongly to this tire" by splitting
> it out and so the cross-product-type recommendations would be a bit
> worse... Is there a way to do this without loosing the link between
> that book+tire?
>
> Which recommender would you use as the base for this- KNN, Slope One,
> SVD? I still need to read the papers on the algorithms, so feel free
> to ignore this question if it'll become obvious!
>
> Thank you for all your work on this library!
>
> -Matthew Runo
>