[mythtv-users] [OT] Commercial Nosiness

Matthew M Murphy mmmurf at umich.edu
Thu Feb 5 20:06:30 EST 2004


On Thu, 5 Feb 2004, Patrick Reynolds wrote:

> On Thu, 5 Feb 2004, Andrew Dodd wrote:
>
> > Given that TiVo has repeatedly stated that their data is no more fine-grained
> > than the zipcode level (which people can most likely verify with packet sniffs
> > of the TiVo's network traffic when it "phones home"), I don't see the problem here.
>
> Sniffing buys you nothing.  You have to authenticate when you dial up,
> right?  (Else, how would they enforce the $13 fee?)  So know exactly who's
> at the other end of the connection when your set-top uploads your viewing
> habits.  They may (and indeed they claim to) strip identifying information
> out at their end, but you probably can't verify it from your end.
>
> > You realize that if we ever want a reccommendations engine similar to TiVo's,
> > we'll have to accept our Myth boxes "phoning home" to a server somewhere too.
>
> Granted.  But there are two very big differences.  First, there's (almost)
> no reason to ask Myth users to authenticate, so your info can be tied only
> to your current IP address, not to your billing info.  And second, a
> recommendation server should only care which shows I tape and/or watch,
> not every daggum press of the "pause," "skip," or "rewind" buttons.
>
> And, of course, there's the detail that a Myth recommendation server would
> be opt-in, while Tivo is opt-out.  (Or do they even give you a choice at
> all?)
>
> --Patrick
>


A recommendation engine is a fantastic idea!  I routinely use Movielens,
http://movielens.umn.edu, for movie predictions, and since I started the
quality of movies that I watch has been greatly improved.

It seems to me that the best way to create such a recommendation engine
would be to do the following on an opt-in basis:

1) Track which shows are recorded by each user
2) Store voluntary rating information for any show, regardless of whether
or not it was recorded.
3) Use an approach such as Collaborative Filtering
(http://www.sims.berkeley.edu/resources/collab/)
to generate a recommendation list for other users.

I would love to be a part of the development and testing of this.

I am currently working on using CART (Classification and Regression Trees)
to predict preference data over a set of data points that include a lot of
missing data.  So far the research is in an early phase, but it shows
promise and could be implemented very efficiently.

I think it's particularly important to have the ratings be rich enough to
generate good predictions.

If there is a parallel thread being started on myth-dev, please let me
know.

-Matt


More information about the mythtv-users mailing list