Re: Recommender systems (Re: Voting on messages: a way to resolve the mailing list problems)
Thanks for writing to my email address; I'm not subscribed to the list as you
may have realized.
Le December 29, 2008 06:59:30 am MJ Ray, vous avez écrit :
> Filipus Klutiero <chealer@gmail.com> wrote:
> > MJ Ray wrote:
> > > I consider filtered indices, auto-responses, shadow lists of only
> > > "good" messages, highlighting, integration with db.debian.org and some
> > > of the other uses for this data to be recommendation systems.
> >
> > A filtered thread index as proposed is not a recommendation list.
>
> A filtered thread index as proposed could be a recommendation system
> according to both descriptions posted, although it depends how one
> interprets "suggest", "support" and so on, and how much
> personalisation one believes is needed to be a recommendation system.
I don't remember anything in this thread suggesting any level of
personalisation, so I don't understand why you question the efficiency of
message-voting due to concerns with recommender systems.
Even if you considered a filtered thread index as a recommendation list, your
quote does not mean that recommender systems perform badly, it just means
that some of the current systems have suboptimal aspects and proposes
solutions. Furthermore, these aspects do not apply in the case of filtered
thread indices.
>
> One can just as well see many drawbacks by looking at more general
> "collaborative filtering" research - or even out into more general
> population clustering work to see the reasons for the drawbacks - but
> it's a bit older, so less of it is online, so I didn't refer to it.
> I'm pretty sure that someone would react to the obvious problems in
> using an unpersonalised filtered thread index (which is a
> collaborative filter, isn't it?) by personalising it to make some sort
> of simple recommendation system, wouldn't they?
An unpersonalised filtered thread index wouldn't be an application of
collaborative filtering.
According to Wikipedia:
> Collaborative filtering systems usually take two steps:
>
> 1. Look for users who share the same rating patterns with the active
> user (the user whom the prediction is for).
> 2. Use the ratings from those
> like-minded users found in step 1 to calculate a prediction for the active
> user
It is possible that readers using an unpersonalised filtered thread index
would want something personalized. I don't know, that's why I think data on
the efficiency of CF in discussion systems would be interesting.
Since you seem to have misunderstood the meaning of CF, I assume the previous
paragraph may be confused, but if you think it stands, please specify your
concerns.
Reply to: