How do 'stacked' recommenders (like the Netflix winners) work?

On Wed, Jul 20, 2011 at 9:22 PM, Jamey Wood <[email protected]> wrote:
> Great.  Thanks, Ted!
>
> --Jamey
>
> On Wed, Jul 20, 2011 at 9:57 PM, Ted Dunning <[email protected]> wrote:
>
>> Oh... you do have to be careful with this a bit because some of these side
>> factors can have disastrously non-sparse characteristics.  For instance, a
>> large fraction of the people in the world are in each age range.  Likewise,
>> there are entirely too many romance novels in the world.  These issues of
>> prevalence can seriously impact your algorithm run-time (adversely).  You
>> can compensate for this by sampling or just recognizing that such pervasive
>> features inherently cannot be very useful since too many things would be
>> recommended.
>>
>> On Wed, Jul 20, 2011 at 8:51 PM, Ted Dunning <[email protected]>
>> wrote:
>>
>> > Yes.  This can work.  And it can go both ways since you might do
>> something
>> > like combine recommendations for a specific book with more general
>> > recommendations for a specific author or genre.  You can also have
>> > recommendations for, say, an author or genre based on demographic
>> quantities
>> > such as geo-location or age range.
>> >
>> > It can be a bit tricky to combine all of these features.  One principled
>> > way would be to extend the log-linear latent factor approach to include
>> > these multiple cross terms.  A less principled, but pretty effective
>> method
>> > is to score all kinds of recommendations independently and then
>> recalibrate
>> > based on percentiles (if you can make sense of that, often not possible)
>> or
>> > by some declining function of rank.
>> >
>> >
>> > On Wed, Jul 20, 2011 at 7:18 PM, Jamey Wood <[email protected]>
>> wrote:
>> >
>> >> Is there any precedent for treating users' demographic characteristics
>> as
>> >> items (particularly for item-based recommendation)?  For example, if one
>> >> were performing item-based recommendation within a bookselling site,
>> it'd
>> >> be
>> >> natural to include the user:item purchases as boolean preferences.  But
>> >> could it also make sense to include certain user:demographic pairs as
>> >> boolean preferences (e.g. user123:age40-to-50)?  Of course, these items
>> >> would need to be filtered (by a Rescorer) in the recommendation outputs,
>> >> but
>> >> I'm curious whether including them as inputs is potentially helpful.
>> >>
>> >> Thanks,
>> >> Jamey
>> >>
>> >
>> >
>>
>



-- 
Lance Norskog
[email protected]

Reply via email to