mahout-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Sean Owen <sro...@gmail.com>
Subject Re: Ease of recommendation for a user
Date Thu, 28 Oct 2010 20:33:14 GMT
I see, you want to give less weight to users in your comparison when they
are hard to recommend for anyway. That's a fine intuition. I don't think
it's a case of over-fitting. I also don't know that there is any
particularly good or bad way to think about this. However I don't know that
the standard deviation of ratings necessarily says something about whether
it's "hard" to recommend. I would more readily use the simple count of
ratings -- fewer means harder. I might suggest, more particularly, using the
log of the count of ratings as a weight. More ratings means more weight but
weight increases less as count goes up, the higher it goes.

On Thu, Oct 28, 2010 at 9:27 PM, Lance Norskog <goksron@gmail.com> wrote:

> Thanks, both of you.
>
> The goal is that I'm writing a recommender and I would like to compare
> its output to what the existing recommenders give. But if a user gives
> a western movie and an Italian horror movie, the recommenders don't
> have much to go on. So, when I create one number (somehow) for the
> performance of the recommender, I would like to downgrade the value of
> the recommendations for that user.
>
> Does this cross the line into over-fitting?
>
> On Thu, Oct 28, 2010 at 4:05 AM, Steven Bourke <sbourke@gmail.com> wrote:
> > It's always worth considering that a model does not reflect the true
> nature
> > of your recommendation. You can't really measure good serendipity or
> > diversity. You could possibly look at the quality of the users
> neighborhood
> > when generating the recommendation. i.e is their a good level of
> similarity
> > between the users or is it quite low. If its quite low, in theory the
> > recommendations are probably not going to be great given the sparsity of
> > overlap.
> >
> > Couple of good starting points here on scholar.
> >
> http://scholar.google.com/scholar?hl=en&client=safari&rls=en&q=recommendation+sparse+data&um=1&ie=UTF-8&sa=N&tab=ws
> >
> > On Thu, Oct 28, 2010 at 9:23 AM, Sean Owen <srowen@gmail.com> wrote:
> >
> >> What's your intuition -- what would you do with this figure? Users with
> >> higher variance are easier or harder to recommend well for? I don't know
> if
> >> that directly affects the quality... probably the diversity of quantity
> of
> >> prefs is more directly relevant.
> >>
> >> On Thu, Oct 28, 2010 at 7:31 AM, Lance Norskog <goksron@gmail.com>
> wrote:
> >>
> >> > The following code is an attempt to decide how easy it to give a
> >> > recommendation to a given user. Does the user give us enough to go on?
> >> >
> >> > The idea is to get recommendations for a user and measure their
> standard
> >> > deviation. As a separate task, get the raw preferences value from the
> >> data
> >> > model for that recommendation, if available, and measure their
> standard
> >> > deviation also. I'm not sure this is the right approach.
> >> >
> >> > Are there standard models for this question?
> >> >
> >> >
> >>
> >
>
>
>
> --
> Lance Norskog
> goksron@gmail.com
>

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