I love this idea, but how would Cambrian House implement it? The way Cambrian House seems to work is that Management comes up with a list of features that need to be implemented, then the community implements those features for Royalty Points. But winning the NetFlix prize is primarily a research problem; a ton of research would need to be done just to find out what algorithms or AI methods might need to be implemented, and I don't think thats the specialty of Cambrian House management.
.
Like I said, I definitely like the idea of working toward the NetFlix Prize, and if someone organizes a good way for a community to work toward this prize, I will be there. But I just dont see how Cambrian House could do it.

BillWaite: I know what you're saying. However, I've been working on the Netflix Prize myself, and I see what it comes down to is really aggregating a bunch of algorithms to make one that "gets every case covered". The statistics masters have already played their hands, and it doesn't get the problem solved. It takes more than that, such as linking to more detailed information pulled from IMDB via a Web Service, or some clever "user matching" algorithm. I think using a divide-and-conquer strategy for building different aspects of the algorithm would have a better chance of winning than any silver bullet approach.

Put more concisely, CH could do this by first writing a framework, then letting contributors plug in their sub-algorithms. Taken in combination, the final algorithm could be very impressive indeed.

Another aspect that CH has to offer that the average person doesn't have on hand is computing power. CH could acquire or supply some supercomputing time or at least a server farm to test algorithms.

I guess it's worth noting that if CH wants to try this, I have an existing Java codebase for the project that I'd be willing to contribute in its entirity.

I would be all over this. :) Upvote!

From experience I can tell you....

This would utilize a significant amount of CH team time, with little chance of success. Of course, it's possible that there's a hidden genius within CH, but how much are we willing to risk? It's basically like buying 100,000 lottery tickets to try to win $1M.

The current leaders on the board are Ph.Ds with specialization in the field. And I'm guessing the prize will be wone within the next 6 months (it's progressing 10 times faster than NetFlix ever imagined).

I'm not saying CH couldn't do it. But it would seriously derail the current business model. I'm expecting CH will make a heck of a lot more than $1M by following their current plan. ;-)

cbegin: You're right that it would be an unlikely victory, but you're wrong about it wasting any CH time. CH would essentially be able to skip the "market test" stage altogether. And the development would be done almost entirely by the community (a large part of it probably done by me personally).

The assertion that it would "derail" the business plan is unfounded. This would no more derail things than doing a market test that later turns out to fail.

I can tell you I'm crazy busy, and don't have the skills to make a dent in that challenge.

Gordon: Well, okay, but there are plenty of us who do have time to work on it...

Isn't the whole point of "Crowds Build It" to make it not about how busy the CH coders are?

I agree it would be good for CH recognition/publicity , but I would not start on this before we know we have a good chance to win. If we can gather the necessary skills, we could go for it.

Netflix is not the only award of this type. We could also setup a taskforce to compete for similar prices we know we can win.

>> but currently nobody's gotten any better score
>> than about 5% improvement.

Keep in mind that 5% isn't disappointing. In fact, it's absolutely amazing. It's at least 10 - 20 times faster than NetFlix ever expected (they underestimated the power of crowds).

My bet: Someone will take the $1 Million prize before Netflix ever pays out a single $50k progress prize.

Also keep in mind, that some of these teams are *sponsored* teams of university students led by people with a Ph.D. in this field.

And I do believe that it is outside of the CH model.

1) It's not mass market software.
2) Skipping the market test isn't a good thing -- that's our due process.
3) It's directly competetive with only one winner -- there's no second place. CH can make a heck of a lot of money making second place mass market products (kind of like how many "second place" brands are more profitable than "first place" brands in the home electronics space).

I do hope that I don't discourage YOU from trying though. I also have my attempt in progress...best I can do is .979 so far...not even worth posting. :-)

Cheers,
Clinton

cbegin: If CH didn't think this fit their business model, they could have rejected the idea. TrashTube and Robinhood Fund are not "mass market software". CH is a bit more diverse in their goals than it first appears, I think.

There are two benefits to this, btw, beyond just "winning". First of all, it would help CH develop a project organization framework. Second, it would provide some great PR/public visiblity.

Anyway, I understand your concerns. I guess I just feel this is something that could be organized and executed "on the cheap" and therefore would be a very low-risk item for CH. Plus it can be abandoned at ANY time with no substantial loss.

Cool out of the box thinking! It has my full vote!

cbegin is surely right about phds plotting to nail this one.

However, the aggregation of inteligence here is greater than in an average institute, simply because more people frequent here. Maybe we can start from there..

So, please tell me what you think about this approach:

1. Our algoritm would be based on neural nets. We would be training a complex neural network with the qualifying set(s). We would not aim to optimize the network for the testing set (we want a final problem solution - that is, similar error rates for any testing sets).

2. We will harness the power of our community.

- the (enlarged) CH community will be used to serve as enhancement towards the qualifiyng set. This way, you will be in position to simulate suplementary test set(s), which is an enormous help.

- we all understand this is an NP problem. I'd rather play dirty and introduce tags and other tricks and treat, rather than to send the netflix a clean C listing with a miniscule error rate at the end of the month. Because, don't forget those are people rating the movies. I mean, if a random STD would make a viewer downvote Dirty Dancing as an accident, there might be a global phenomenon (say, a plane crash) accidentaly influencing a vote on a movie (say, Top Gun) I would even correlate the theatrical release dates of movies in the qualifying sets with google trends, if that's a good idea and if it would help solve the problem!

- We would have the CPU power (as the author suggested) to do simultaneous retraining of the neural nets, for instance, towards a better subsumption architecture. How many people we have that are into AI and data mining, anyways?

i dont think we have the right community yet...