Sunday, January 3, 2010

What Are We Going to Need to Get a Whuffie System to Work?

After messing around with the idea of whuffie and various extensions of it for quite a while, long before this previous post, I have ended up with lots of vague ideas, many of them fairly far detached from immediate reality. In an attempt to bring things back to a more practical realm, I'd like to do a bit of organizing.

Looking at what is out there now, there are a number of recommendation systems that already function fairly well, Amazon, Last.fm, Digg, Reddit, Slashdot, etc. Unfortunately, all are quite limited to their specific domains with their own specific ways of functioning. With that as a starting point, I should also ask where it is that I want to be going. How would a person actually interact with this system I am trying to imagine?

At the simplest level, one could use it as we use existing recommendation/reputation systems: suggestions as to what product to buy, what song to listen to, what comment or news story to read. The recommendations would simply be better. More specifically, how do I want to be able to use such a system that I can't already?

I want to be able ask a system questions such as, "What books should I read?", "What's the best product x in price range y for people with my tastes?", "What is the best discussion on this issue?", "Can I see these candidates ranked using this mix of metrics?", "Based on my tastes what are the best TV shows for me?", "Who is the best person in my social circle in terms of x?", "Who is the most respected writer out there on issue y?"

A lot of that is already done to some extent via search engines, or would require more advanced natural language processing. But I can't help but feel that the various subjective ranking stuff out there should be used better. I suppose that's really the issue. We have ranking systems, but they're all completely disconnected. Rankings should be fungible and aren't.

Reputation is decided not simply by people's explicit declarations, but by things such as what people buy, what people link to, where people go, etc., and it would be very complicated to reduce all that into one universal system. Back to the original idea of keeping things simpler without going into all those complications... aside from the fact that I should learn more about machine learning algorithms, I suppose my initial impulse for wanting to find a foundation for all this comes from how completely useless most of the explicit rating systems I've seen are.

We see systems where people can rank things 1 to 5 or 1 to 10 on eBay, Yelp, Google Maps, Amazon.com, etc. But, I've never been able to get anything close to an accurate result from these systems. There is no correlation weighted between my opinions and the opinions of the reviews as I would like, obviously a review should be given higher weight for me if the reviewer has similar tastes to mine. Some people rate everything a five out of five or rank it zero, with nothing in between. Different systems have different metrics. Because of problems such as this, these systems seem to only be vaguely used, or have their use specific to one particular site. The Amazon.com recommendation systems recommends items to you based upon what other people have bought, not on what they've ranked, that ranking system seems largely superfluous and tacked on. On Digg and Reddit, systems which do use explicit declarations, there is no range of ranking, only up or down votes.

I would like to see a system that that can see people's reviews across multiple platforms, weight them based on how relevant they are to the user, and provide some regularity to their usage.

Then, once we get that system together, then I suppose we could move forward. I'd love to see things such as plugging together Amazon.com's and Last.fm's recommendation systems together with those ranking systems. Then you can start using things such as tags, location data, social contacts, etc, to get some really accurate recommendation systems. And then maybe we can begin to think about a Whuffie system.

No comments:

Post a Comment