Social devaluation sounds interesting as well. I can't say much for Hootie though :)

Daerax: The main algorithm that is being used is an additive distance metric developed by my supervisor in the 80s. The original paper he published on it is in Romanian, but it's called "A metrical method for multicriteria decision making", or something along those lines. Over the past year we've modified it here and there though to better reflect the problem (missing value dimensions in people's belief system and for incomplete ratings for entities), and added the concept of importance via weights to the algorithm as well.

Then there are similarity algorithms that are based on a paper called 'on the similarity and distance metrics' (or something, can't remember the exact name, but that should work in a search). The similarity and distance algos of course use the belief systems as inputs, or item ratings when it comes to item similarity. And then the rest of the algorithms in use are simple aggregation.

Traction is another concern I guess. I've read a lot about how wikipedia started out, and even now how the core number of people who actually contribute regularly is just about 1000 people. The rest have actually made a contribution maybe as little as once. I really don't have an answer about gaining users, obviously something like this would require a lot of user input to work. I'm imagining it would start in niches, like only for videogames at first, then for restaurants in country X only, then something else, and this would carry on till it's quite massive and things could be linked up.

I believe the Netflix people used an approach that was quite different than what was normally employed for recommendations, matrix factorization as opposed to nearest-beighbour models. And it gave an improvement of over 10%. I know too much information may be not good

Quote: And if desire is fleeting, then the ratings are fleeting, and the system is broken.

I woud say this depends on how many bad apples there are in the crate. I do not believe this is a case of one bad apple will ruin them all. It's not actually, if the number of ratings are low, then this will be a problem, and is a problem with all recomender systems...

That's interesting about the 5 stars thing though. I'll add that into my questionaire to see what people say :)

Awesome. So it sounds like it places the patterns in an n-dimensional real space and tries to either auto generate clusters or find close neighbors cant tell which based on description (maybe neighbors based on wiki requirement). And your advisers have found a performant similarity metric.

The netflix people actually started with the standard neighbor method but found it failed to find hidden associations which NMF could extract. They ended using an ensemble of different methods including Matrix Factorization and Nearest Neigbor to get that final performance boost.

And yes they found as well that a binary or ternary system has faar less noise than a 5 point system for much the same reason as Heath's insightful post.

FWIW I stated that too much features might be a problem due to the increased chance of silly associations and not useful generalizations and performing poorly at the long tail of personalities where many reside.

As well I do not think even focusing will be enough for wider traction, for example - only a few people write walkthroughs. But I do think there is definite niche potential and that this could definitely get plenty of traction if it powered something like say a dating site. I think the potential there is large.

With all that said, you should check out hunch for competitive reasons - learning from their mistakes say. They do almost exactly what you state but have neither been validated as possessing a viable business model.

When you ask if it's a neighborhood method or cluster, are you talking about the ranking algorithm? Or...?

The ranking algorithm basically determines which items are the closest to your belief system. Items are rated over multiple dimensions, so if an item is high on price, but your belief is that "prices should be low", then this item is father away from your belief.

When it comes to recommendations, we find the items that are similar (alteration of the original distance algorithm) to the item you are looking at, and give it a rating based on your belief system. This way, you don't have to have rated anything in the system at all to get recommendations. You will have to set your belief system though, which brings me to ....

Holy crap Hunch is quite similar, I'm a wee bit annoyed I must say :| I suppose the questions can be considered as value dimensions of a sort, and once you input your answers, tada. I'm going to hope that the belief system I use is better though, because we use ordinal value dimensions that can be ranked and ordered, but Hunch uses nominal ones which can't. And we have an addition of weights as well. So I guess the concept it quite similar. My goal is different though: while recommendations play a huge part, I'm more after a rating-wiki end game. Sort of like wikipedia, but for ratings instead of facts. And the belief system part allows you to see exactly why something is rated as it is because it adds a lot of transparency.

Yes, the niche thing is what I meant when I said traction. So if it turned into a dating raking, then you got a celebrity ranking based on teh same system, they could be linked up, then you get a food ranking, then a company ranking (all niche markets with a few people who contribute) and then you link them up as well.