A Different Way to Discover Podcasts

[Hittman]

I don't know the name for this technology, but if it could be implemented on a podcast directory, it would be one kick-*** feature.

Both Amazon and Netflix have it. Netfilx for intstance will show you five movies and ask you to rate them on a scale of one to five, with a "haven't seen it" and "sounds horrible" selection also available. After you've rated 15 or so, it starts making recommendations, which you also rate, further fine tuning your selection.

The recommendations work by comparing your opnions to those of other people in the system. Because people who like A and B tend to like C as well, they'll recommended C to you if you rate A and B highly. After enough picks, it starts working so well it's eerie.

I don't think I'm explaining this too well, so let me try with a self-centered example.

I give high ratings to The Sounds In My Head, Bandtrax, and Old Wave Radio. Someone else comes along, takes the survey, and gives high ratings to Bandtrax and OWR. The system says "Hey, based on other people's opinions, you'll probably like TSIMH as well." If someone picked TSIMH and OWR, it might recommend Bandtrax.

I don't imagine this would be simple to program, but wouldn't be surprised if there's some open source stuff out there that require little or no modification.

Does anyone know what this method is called?

[jeffoest]

Collaborative Filtering

It's a cool technology that has been around (at least in theory) for at least 15 years now. I studied it in grad school as there are a lot of statistics involved in it's implementation. There were companies out there in the early days of net commerce that oversold the technology and couldn't ride out he wave.

Bottom line - to be of use at all it requires a LOT of data - ie. a LOT of customer / listener behaviour statistics.

Amazon is probably one of the few retailers with enough data to do it BUT you know it always ends up recommending me? Stuff by artists that I bought already.... but occasionally it comes up with good recommendations I"ll give it that. I can see iTunes having a potential of using this technology.

Bottom line - I agree that it is and can be a great tool for selecting content among a land that may have 20,000+ 'channels"... but it will only really work effectively if there is one big 'directory' service and not a number of ones that fragment the market. There lies the rub as well - I can't say I want to argue for a monopoly either as that also carries it's disadvantages....

[jeffoest]

OH I should add - now that you've got me thinking of this. Besides the need for heavy amounts of data, this technology is expensive. It not only requires pretty hefty CPU (although that's admittedly much cheaper these days), but it requires some pretty heavy duty math and IT folks to set it up, collabrate it, and manage it. Remember neural nets? Very similar technologies.

So... there is the ROI issue involved with implementing a very expensive technology. Helping folks find podcasts among 20,000+ probably isn't going to be a real revenue generator at least in the short term....

My hunch is that Amazon combines a light-weight use of these models with products that they want to 'push' i.e. that's why I always get recommendations for artists I have bought before who have a new offering.... just a hunch....

[kickasspodcast]

I know exactly what you mean Hittman and I think its a great idea. I think that it really only requires an advanced algorithm or set of algorithms. It really wouldn't be that hard to make it yourself in Scheme.

http://www.scheme.org/

But it would take you quite a while.

Best o luck

Jack b.

[SFEley]

Amazon is probably one of the few retailers with enough data to do it BUT you know it always ends up recommending me? Stuff by artists that I bought already.... but occasionally it comes up with good recommendations I"ll give it that. I can see iTunes having a potential of using this technology.

Absolutely. This could, in fact, be one of the major advantages of podcasting in iTunes, if anybody bothers to develop it. I can hit the center button twice on my iPod and give any track a 1 to 5 star rating. Takes maybe two seconds. I always do this for songs; haven't bothered yet with podcasts, since I don't generally listen to them twice.

But if there were a highly visible "matching" service, such that my star ratings on podcasts were matched with others' and new podcasts were recommended to me, I'd have a motivation for rating them. If this weren't built into iTunes, somebody else could build it using the iTunes playlist sharing features.

[Hittman]

Collaborative Filtering

Thanks.

I knew it wasn't easy, but also knew it's been around a long time, and was wondering of someone has put together a faster, cheaper way to do it.

Is there some kind of formula to determine how much data is needed before it becomes useful?

[jeffoest]

Excellent question, Hittman.... I haven't kept up with it so I could not tell you... if you don't beat me to it, I will try to Google a bit and see if there is anything interesting on it...

[Hittman]

Wikipedia has a good article on it, and links to various places where you can get software, including a commercial company that went belly up and made their software public domain. (Alexlit.com)

http://en.wikipedia.org/wiki/Collaborative_filtering

Here's an overly wordy article about it by Malcom Gladwell.

http://www.gladwell.com/1999/1999_10_04_a_sleeper.htm