Hi,

I am new to mahout and looking for ideas to implement "Customers with
Similar Searches Purchased" and "What Do Customers Ultimately Buy After
Viewing This Item?" style recommendations on amazon.com. Is it possible with
the current Taste implementation? Any pointers will be helpful.
Thanks,
Pramit

These are slightly different from conventional collaborative
filtering, but I think solutions are available.

"Customers with Similar Searches Purchased"

To apply user-based CF you need a notion of user-user similarity. You
could think of this as a sub-problem, where users are users and
searches are items, and apply any of the standard UserSimilarity
measures to compute user-user similarity.

Then, yes this becomes user-based collaborative filtering, but without
ratings. You can just use GenericUserBasedRecommender with your
UserSimilarity.

That just gets you started -- I think there's room to optimize and
improve on that basic start, such as implementing a custom
UserNeighborhood.
"What Do Customers Ultimately Buy After Viewing This Item"

This isn't really CF, but association rule mining. You might look at
the "Frequent Pattern Mining" support here instead.
On Sun, Aug 29, 2010 at 6:25 PM, Pramit Vamsi <[EMAIL PROTECTED]> wrote:
> Hi,
>
> I am new to mahout and looking for ideas to implement "Customers with
> Similar Searches Purchased" and "What Do Customers Ultimately Buy After
> Viewing This Item?" style recommendations on amazon.com. Is it possible with
> the current Taste implementation? Any pointers will be helpful.
> Thanks,
> Pramit
>

These are examples of what I call cross-recommendation where you have user x
item1 and user x item2 data and you
want item1 => item2 recommendations.

All of the standard techniques apply (user-based, LLR cooccurrence, SVD,
latent factor models), but you have to rejigger things here
and there.

Sean, can Mahout's recommendation system do this cross recommendation?

On Sun, Aug 29, 2010 at 10:39 AM, Sean Owen <[EMAIL PROTECTED]> wrote:

> "Customers with Similar Searches Purchased"
>
> To apply user-based CF you need a notion of user-user similarity. You
> could think of this as a sub-problem, where users are users and
> searches are items, and apply any of the standard UserSimilarity
> measures to compute user-user similarity.
>
> Then, yes this becomes user-based collaborative filtering, but without
> ratings. You can just use GenericUserBasedRecommender with your
> UserSimilarity.
>
> That just gets you started -- I think there's room to optimize and
> improve on that basic start, such as implementing a custom
> UserNeighborhood.
>
>
> "What Do Customers Ultimately Buy After Viewing This Item"
>
> This isn't really CF, but association rule mining. You might look at
> the "Frequent Pattern Mining" support here instead.

Yes, this is a simpler problem. You just want to find which items are
most similar to a given item, for some definition of 'similar'.
GenericItemBasedRecommender has a mostSimilarItems() method that just
saves you the trouble of computing this by hand, and any
ItemSimiliarity function you like can be used.

On Sun, Aug 29, 2010 at 7:26 PM, Ted Dunning <[EMAIL PROTECTED]> wrote:
> These are examples of what I call cross-recommendation where you have user x
> item1 and user x item2 data and you
> want item1 => item2 recommendations.
>
> All of the standard techniques apply (user-based, LLR cooccurrence, SVD,
> latent factor models), but you have to rejigger things here
> and there.
>
> Sean, can Mahout's recommendation system do this cross recommendation?
>

I have some understanding now. So given 2 matrices user * (page view/search
term) and user * (purchased item), how do you connect these 2 matrices
given that I can define the user or item sim methods?

Also, can the second use case can be solved with CF or association mining is
needed?

Pramit

On Mon, Aug 30, 2010 at 12:07 AM, Sean Owen <[EMAIL PROTECTED]> wrote:

> Yes, this is a simpler problem. You just want to find which items are
> most similar to a given item, for some definition of 'similar'.
> GenericItemBasedRecommender has a mostSimilarItems() method that just
> saves you the trouble of computing this by hand, and any
> ItemSimiliarity function you like can be used.
>
> On Sun, Aug 29, 2010 at 7:26 PM, Ted Dunning <[EMAIL PROTECTED]>
> wrote:
> > These are examples of what I call cross-recommendation where you have
> user x
> > item1 and user x item2 data and you
> > want item1 => item2 recommendations.
> >
> > All of the standard techniques apply (user-based, LLR cooccurrence, SVD,
> > latent factor models), but you have to rejigger things here
> > and there.
> >
> > Sean, can Mahout's recommendation system do this cross recommendation?
> >
>

--
Thanks,
Pramit

I really need to take a look at the Mahout code (I haven't had a chance yet)
so I'm not sure if this type of rec is possible, but what I would do is
something like this:

Enumeration of Search Terms:
1      Bath soap
2      Headphones
3      Computer laptop

Enumeration of Users:
userid 1
userid 2
userid 3
userid 4

Joined Matrix
UserID      TotalCountOfSearchItem      SearchItemID
LocalCountSearchItem
1                               15                                     1
                            3
1                                7                                      2
                              1
2                                15                                    1
                            5
2                                7                                      2
                              4
3                                10                                    3
                            5
3                                15                                    1
                            7
4                                10                                    3
                            5
4                                7                                      2
                              2

I wrote a blog post about Naive Bayes for classification tasks that
describes this type of layout here:

http://www.thedatascientist.com/2010/05/22/how-i-would-use-the-google-prediction-api/

But this type of data layout is algorithm agnostic, so you can use it for
whatever you need to do.  Its just a matter of feeding the data into a form
that Mahout will recognize (my guess)

Chris
On Mon, Aug 30, 2010 at 10:47 AM, Pramit Vamsi <[EMAIL PROTECTED]>wrote:

> I have some understanding now. So given 2 matrices user * (page view/search
> term) and user * (purchased item), how do you connect these 2 matrices
> given that I can define the user or item sim methods?
>
> Also, can the second use case can be solved with CF or association mining
> is
> needed?
>
> Pramit
>
> On Mon, Aug 30, 2010 at 12:07 AM, Sean Owen <[EMAIL PROTECTED]> wrote:
>
> > Yes, this is a simpler problem. You just want to find which items are
> > most similar to a given item, for some definition of 'similar'.
> > GenericItemBasedRecommender has a mostSimilarItems() method that just
> > saves you the trouble of computing this by hand, and any
> > ItemSimiliarity function you like can be used.
> >
> > On Sun, Aug 29, 2010 at 7:26 PM, Ted Dunning <[EMAIL PROTECTED]>
> > wrote:
> > > These are examples of what I call cross-recommendation where you have
> > user x
> > > item1 and user x item2 data and you
> > > want item1 => item2 recommendations.
> > >
> > > All of the standard techniques apply (user-based, LLR cooccurrence,
> SVD,
> > > latent factor models), but you have to rejigger things here
> > > and there.
> > >
> > > Sean, can Mahout's recommendation system do this cross recommendation?
> > >
> >
>
>
>
> --
> Thanks,
> Pramit
>

I would take the user-search "matrix" and create one DataModel from
it. Define a LogLikelihoodSimilarity object on top of that. That's
your user-user similarity measure.

Then the user-purchase "matrix" forms the basis of another DataModel
that is actually plugged into a GenericUserBasedRecommender with the
similarity measure above (which actually drives off different data).

Off the top of my head, that ought to work out.
The second situation might just be really simple. If you have view and
purchase data, simply count up and find which purchase was most
frequent among all purchases that followed from a view of the current
item. That's simple, perhaps oversimplified for your context.

I can think of ways to construe this as a CF problem but I think it
just adds complication with no value. It's not really CF.

On Mon, Aug 30, 2010 at 3:47 PM, Pramit Vamsi <[EMAIL PROTECTED]> wrote:
> I have some understanding now. So given 2 matrices user * (page view/search
> term) and user * (purchased item), how do you connect these 2 matrices
> given that I can define the user or item sim methods?
>
> Also, can the second use case can be solved with CF or association mining is
> needed?
>
> Pramit
>
> On Mon, Aug 30, 2010 at 12:07 AM, Sean Owen <[EMAIL PROTECTED]> wrote:
>
>> Yes, this is a simpler problem. You just want to find which items are
>> most similar to a given item, for some definition of 'similar'.
>> GenericItemBasedRecommender has a mostSimilarItems() method that just
>> saves you the trouble of computing this by hand, and any
>> ItemSimiliarity function you like can be used.
>>
>> On Sun, Aug 29, 2010 at 7:26 PM, Ted Dunning <[EMAIL PROTECTED]>
>> wrote:
>> > These are examples of what I call cross-recommendation where you have
>> user x
>> > item1 and user x item2 data and you
>> > want item1 => item2 recommendations.
>> >
>> > All of the standard techniques apply (user-based, LLR cooccurrence, SVD,
>> > latent factor models), but you have to rejigger things here
>> > and there.
>> >
>> > Sean, can Mahout's recommendation system do this cross recommendation?
>> >
>>
>
>
>
> --
> Thanks,
> Pramit
>

Metaphorically speaking if user x search term is A and user x item is B,
then transpose(B) * B is item x item, transpose(A) * B) is search term x
search term and transpose(B)*A is item x search-term.

Depending on what kind of recommendation system you are using, the actual
mechanics will be different, but the shape of the computation will still be
fairly similar to the matrix multiplication.

In data-base terms, you are joining on user id, but I find the database view
of this less helpful for generating good insight into what is happening.

I gave a talk the last half of which is on this topic.  This talk is
recorded here:

    http://fora.tv/2009/10/14/ACM_Data_Mining_SIG_Ted_Dunning

On Mon, Aug 30, 2010 at 7:47 AM, Pramit Vamsi <[EMAIL PROTECTED]>wrote:

> I have some understanding now. So given 2 matrices user * (page view/search
> term) and user * (purchased item), how do you connect these 2 matrices
> given that I can define the user or item sim methods?
>

On Mon, Aug 30, 2010 at 9:44 AM, Ted Dunning <[EMAIL PROTECTED]> wrote:

> Metaphorically speaking if user x search term is A and user x item is B,
> then transpose(B) * B is item x item, transpose(A) * B) is search term x
> search term and transpose(B)*A is item x search-term.
>

And in fact, this is one map-reduce job in Mahout, as A.times(B) for
DistributedRowMatrices A, B performs exactly transpose(A) * B.

  -jake