> >I don't know about you guys, but something that highlighted > programs I > >might be interested in would certainly be of benefit to me, saving me > >trawling through the listings. > > > > Recommender systems are common now (Amazon for example) and have a > decent history (e.g. RINGO from MIT). > > Is the training data rich enough? > > I watch two to seven programmes a week, apart from random channel > switching, for example, and I tend to find them on BBC3 and BBC4 > (first, before they head off to BBC2). > > Gordo
The general (unproven, not particularly closely thought through) idea I had was to build a feature vector based on polygram analysis of item entries within the XML, with different weighting based on location within the elements - for example words in a Genre element would have more weight than words within a Synopsis. Really this is glorified keyword searching, but with the keywords being dynamically generated based on a user's perceived preferences, and with a lot fuzziness being built in. The only possible problem is that the terseness of the programme synopsis could lead to a situation where the language implies something that isnt explicitly stated. Can't think of an example of this off the top of my head, but its why I'd consider bi- or poly-gram analysis as opposed to unigram. As far as the same programmes coming round again, that would be down to individual implementation style I guess, your options are to ignore it, to not suggest it again, or to try to track whether it ought to be suggestable based on prompting the user to complete a set of what he/she watched previously (assuming that the CRID for a programme doesn't change when it switches channel). I think really its something that you can debate the ins and outs of til the cows come home, but until someone is able to sit down and try it, you can't judge how well it will/won't work. - Sent via the backstage.bbc.co.uk discussion group. To unsubscribe, please visit http://backstage.bbc.co.uk/archives/2005/01/mailing_list.html. Unofficial list archive: http://www.mail-archive.com/backstage@lists.bbc.co.uk/