Bi-Factorization vs Tri-Factorization for recommender systems

[Parimi Rohit <ro...@gmail.com>]

Hi All,

Are there any (dis)advantages of using tri-factorization (||X - USV'||) as
opposed to bi-factorization ((||X - UV'||)) for recommender systems? I have
been reading a lot about tri-factorization and how they can be seen as
co-clustering of rows and columns and was wondering if such as technique is
implemented in Mahout?

Also, I am particularly interested in implicit-feedback datasets and the
only MF approach I am aware of is the ALS-WR for implicit feedback data
implemented in mahout. Are there any other MF techniques? If not, is it
possible (and useful) to extend some tri-factorization to handle
implicit-feedback along the lines of "Collaborative Filtering for Implicit
Feedback Datasets" (the approach implemented in Mahout).

I apologize for any inconvenience as this question is very general and
might not be relevant to Mahout and I would really appreciate any
thoughts/feedback.

Thanks,
Rohit

Re: Bi-Factorization vs Tri-Factorization for recommender systems

[Parimi Rohit <ro...@gmail.com>]

Thanks for your reply Ted!!

To summarize, if I am interested in constraining (normalizing) both user
and item vectors, then tri-factorization is a better approach. But if I
want only one, either user or item features, then bi-factorization works.

Please correct me if I understood it incorrectly.

Also, is there any implementation in Mahout that allows me to tri-factorize
user-item preferences (implicit feedback)? If not, are there any other
alternatives?

Thanks,
Rohit






On Mon, Nov 24, 2014 at 11:28 AM, Ted Dunning <te...@gmail.com> wrote:

> There is no inherent mathematical difference, but there may be some pretty
> significant practical differences.
>
> Using the three matrix form (X = USV') puts the normalization constants
> into a place where you can control them a bit easier.  This can be useful
> if you want *both* user and item vectors that are normalized.
>
> If you only want item vectors, then it really doesn't matter since you can
> incorporate as much of S as you like into the item vectors as you like and
> the rest winds up in the factor that you aren't looking at anyway.
>
>
>
> On Thu, Nov 20, 2014 at 1:34 AM, Parimi Rohit <ro...@gmail.com>
> wrote:
>
> > Hi All,
> >
> > Are there any (dis)advantages of using tri-factorization (||X - USV'||)
> as
> > opposed to bi-factorization ((||X - UV'||)) for recommender systems? I
> have
> > been reading a lot about tri-factorization and how they can be seen as
> > co-clustering of rows and columns and was wondering if such as technique
> is
> > implemented in Mahout?
> >
> > Also, I am particularly interested in implicit-feedback datasets and the
> > only MF approach I am aware of is the ALS-WR for implicit feedback data
> > implemented in mahout. Are there any other MF techniques? If not, is it
> > possible (and useful) to extend some tri-factorization to handle
> > implicit-feedback along the lines of "Collaborative Filtering for
> Implicit
> > Feedback Datasets" (the approach implemented in Mahout).
> >
> > I apologize for any inconvenience as this question is very general and
> > might not be relevant to Mahout and I would really appreciate any
> > thoughts/feedback.
> >
> > Thanks,
> > Rohit
> >
>

Re: Bi-Factorization vs Tri-Factorization for recommender systems

[Ted Dunning <te...@gmail.com>]

There is no inherent mathematical difference, but there may be some pretty
significant practical differences.

Using the three matrix form (X = USV') puts the normalization constants
into a place where you can control them a bit easier.  This can be useful
if you want *both* user and item vectors that are normalized.

If you only want item vectors, then it really doesn't matter since you can
incorporate as much of S as you like into the item vectors as you like and
the rest winds up in the factor that you aren't looking at anyway.



On Thu, Nov 20, 2014 at 1:34 AM, Parimi Rohit <ro...@gmail.com>
wrote:

> Hi All,
>
> Are there any (dis)advantages of using tri-factorization (||X - USV'||) as
> opposed to bi-factorization ((||X - UV'||)) for recommender systems? I have
> been reading a lot about tri-factorization and how they can be seen as
> co-clustering of rows and columns and was wondering if such as technique is
> implemented in Mahout?
>
> Also, I am particularly interested in implicit-feedback datasets and the
> only MF approach I am aware of is the ALS-WR for implicit feedback data
> implemented in mahout. Are there any other MF techniques? If not, is it
> possible (and useful) to extend some tri-factorization to handle
> implicit-feedback along the lines of "Collaborative Filtering for Implicit
> Feedback Datasets" (the approach implemented in Mahout).
>
> I apologize for any inconvenience as this question is very general and
> might not be relevant to Mahout and I would really appreciate any
> thoughts/feedback.
>
> Thanks,
> Rohit
>