mahout-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Sean Owen <sro...@gmail.com>
Subject Re: Similar User->Item recommenders
Date Sat, 11 Dec 2010 09:47:53 GMT
I think you can probably find something like what you're looking for (I
don't have this off the top of my head) though I might disagree with the
premise.

The only "real" way to test a recommender is for users to list what their
actual next-best preferences are and compare against that. But rather by
definition they don't know that. A next-best proxy is looking at the next
items the user rates or prefers and compare against that. Even this isn't
perfect. The next thing a user clicks could be the 29th-most-preferred item
"in reality" from the universe of unknown items. So that test would be
faulting a perfect recommender for listing 1-10. Still, this next-best test
is what you get in the evaluator code, mostly. It actually picks a random
subset to withhold, ignoring time.

If you had two reasonably different approaches that returned the same
thing... I guess I'd say they're not so different. The reason the different
implementations exist is to give not only different performance
characteristics but different definitions of what estimated preference and
best recs should be. At some level they're supposed to do different things.
I wouldn't say it's broken just because it doesn't have the same answers as
another.


On Sat, Dec 11, 2010 at 3:25 AM, Lance Norskog <goksron@gmail.com> wrote:

> I'm implementing a separate Recommender. I don't see how I can declare
> it working until it strongly (but not completely) correllates with the
> output of an existing trusted Recommender. The only way I can see to
> find a trusted recommender is to find two similar ones, thus this
> request:
>
> Can anyone please give the precise constructors and common dataset
> for: two different User->Item Recommender implementations that return
> similar results for most users for a given datamodel.
>
> It can be any particular datamodel, but GroupLens preferred. It can be
> any two different User->Item Recommenders.
>
> --
> Lance Norskog
> goksron@gmail.com
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message