A couple of weeks ago Anthony Rose asked us to imagine the coolest way of interacting online with your friends around television programmes. I wonder if he was thinking about something like this:

As you can see from the screenshot above, I’ve just added Last.fm style friend compatibility to the iPlayer Facebook application. When looking at a user page you now get a compatibility of Low, Medium or High with that user and can see a couple of the programmes you have in common.

I’ve also added a My Friends page that ranks your ten most compatible friends, showing the programmes you both like, alongside recently popular programmes amongst your friends:

How It Works

The recommendation code behind the new features is pretty basic at the moment, with the Euclidean Distance correlation from the book Programming Collective Intelligence used to calculate the similarities. I also tried the Pearson correlation but it seemed to produce some pretty strange results with the ratings data and I’m not quite sure why. If anyone has any experience in this area please do get in touch.

As I’m still scraping the iPlayer site and don’t have the real brand data, the brands shown as common programmes are derived by splitting the episode titles based on the colons. Showing common brands seems to make the most sense, episode level seems to be too granular for this sort of usage.

Problems

There a quite a few problems with the recommendations as they currently stand. As already mentioned the system is pretty naive and not particularly accurate.

1. Compatibility is computed solely based on similar ratings, so if you only have one programme in common with another user but you both gave it the same rating then you have a High compatibility. If, however, you have a number of programmes in common but have rated them slightly differently then you have Medium compatibility.
2. It would also be good to be able to recommend based on simply watching a programme rather than having to rate it, but unfortunately I can’t do that as I don’t know when the player is started. It would certainly be possible for iPlayer to do this.
3. Given that brand seems to be the right level to recommend at, there should should be a brand weighting that takes into account the number of episodes watched in a brand and the average rating given to those episodes.
4. The recommendation is done based on your last 30 ratings, so if you rate a lot of programmes it is easy to get into a situation where it will say you have no programmes in common with someone even when you do.
5. Recently popular ranks by popularity before date, so the programmes may not actually be very recent. For example the current number one on my list is from five months ago!

So there are quite a few problems there, and I’m sure there are more, but I thought it was worth putting it out there anyway. Hopefully it will start some discussion around how best to do user-user recommendation for television and radio programmes.

Future

Going forward I’d like to see the BBC tackling these issues and providing an API, but in the mean time there are a couple of things on the wish list. It would be good to use the compatibility scores to decide which friends get priority on the Home page. APML is also likely to make an appearance at some point.

In the meantime, have a look and let me know what you think.