[mythtv] Re: Myth recommendation system

[Charlie Brej]

IvanK. wrote:
> 1) For this to work, we'll need a sufficient number of users participating.
> 2) You can't guarantee that a malicious user won't poison the database with 
> misleading recommendations.
> 3) Even if you can assure 1) and prevent 2), how reliable is other people's 
> taste in tv programming when it comes to predicting what *I* might like?

For 1 and 2 I think it is up to volunteers. I think the system will have very 
low bandwidth usage (for the sake of the dial up users as well as for the server 
cost). It will require a sysadmin to look over the health of the database and if 
there is someone sending garbage then they will need to be kicked etc. I would 
be happy to spend a hour a week looking over the status and spotting mistakes. 
Besides it relies on the fact that there are more good people than bad people 
(ala wiki/forums etc.). There will always be a troll but they will not agree 
with the majority of people and they will have a low trust weight.

Secondly the system is designed to run in two modes. You can connect to the 
server and exchange your recommendations with other people (exchange being the 
correct word as the server needs your recommendations to determine your taste).

Or you can do it in a peer2peer mode where you take other people your trust 
recommendation lists and and feed them in. This part is already written. They 
will only pop up as hints and will not record anything unless you explicitly 
make it do so.


> In view of that may I propose a different approach.  I believe we already have 
> fields for genre/director/actors in the xml data we pull from datadirect.  
> Say we record a romantic comedy with Hugh Grant (yeah, like hell we 
> will! :-) ), directed by so and so.  Shouldn't myth be looking for other 
> films of genre comedy, or films in which Hugh Grant appears, or which are 
> directed by a the same director?

Obviously the server code would be a separate application but its algorithms 
could be a shared library which works in your client side version of myth. You 
could even run your own server turning on only the word/actor/genre recognition 
algorithms.

> This way we avoid the whole central server/many people to rely to scenario 
> while at the same type we base our predictions for what we may find likable 
> on what we've already liked.

Certainly the server would not be an essential part of myth as some people don't 
have constant internet access but there is always the option of running your own 
or the methods being ported back into myth.

> These are just a few random thoughts I decided to share just to kickstart a 
> discussion on this very interesting topic.

Thanks for those. I heard your concern from other people and am trying to keep 
the non-centralized structure.


-- 
         Charlie Brej
APT Group, Dept. of Computer Science, University of Manchester
Web: http://brej.org/                 Tel: +44 (0)161 275 6177
Mail: IT303, University of Manchester, Manchester, M13 9PL, UK