Open-Source "Ratings & Recommendations" Software?

The Llama King asks: "Our group has an interesting idea for being able to rate different items, then receiving preferences for similar items, a feature found at sites such as NetFlix and Amazon. Unfortunately, we have big ideas and a small budget. I've searched high and low for an open-source version of this kind of algorithm, with no success. Are there any out there worth compiling?" Update: 05/16 10:30PM EDT by C :As it turns out, Jamie has some words on the subject, click below for more.

In an email from co-editor, Jamie:

"I researched this stuff for a possible project some years back. Not much has changed.

There isn't any open-source code out there that I know of, but, people have been writing masters' theses and dissertations about it for several years now. They can go search the literature if they're really interested. But there isn't just a perl module you can install to get this stuff...yet.

You should probably try these search terms:

• 'recommender system'
• 'recommendation system'

'FireFly' is another one -- that was the name of some (fairly successful) recommendation software which was purchased by our favorite innovator, Microsoft, three years ago and repackaged as (surprise!) 'Passport.'

[And for those interested]...here's a promising link .

No.

[Anonymous Coward]

Sorry.

Re:Hire someone

[PD]

Sounds like the /. moderation system. Is Taco a statistician?

Re:No.

[Phork]

+1 funny?
no, sorry, try -1 offtopic.

Collaborative filtering?

[dave-fu]

Heard the buzwords, tried the systems... at the end of the day, I'm consistently unimpressed. The best that I've seen is All Music [allmusic.com] which finds stylisticaly similar albums and does a more or less good job.
P.S.: the age of mana from VCs is long since over. Here's where I'd recommend you come up with something smarter than "we're running headlong up against Amazon with a recommendation system that we found somewhere on the internet" for your business plan if you're at all serious about this.

Don't know about "OpenSource"

[Viking Coder]

I don't know about the OpenSource part of it, but you should check out GroupLens [umn.edu] at the University of Minnesota. They've published some papers that are available online.

Hire someone

[einhverfr]

It can't be that hard. I am sure that I could design a basic scheme pretty fast. Backend code would probably take me 4 to 8 hrs at most using PHP and MySQL...

Re:Hire someone

[einhverfr]

Well, I am not so sure that it needs to be that statistical (except for reporting purposes to the porject managers for monitoring purposes). How about listing the 10 most common recommendations, or the the top 10%?

To be useful and valuable, the information does not have to be statistically valuable-- who needs to know that 97% of users thought that Slackware was a good distrobution with a stadard deviation of +/- 1% or that 15% of people recommeded PostgreSQL +/- 10%?

Rather the importance should be in delivering useful results to be incorporated in a web page.

Re:Hire someone

[raju1kabir]

Well, I am not so sure that it needs to be that statistical (except for reporting purposes to the porject managers for monitoring purposes). How about listing the 10 most common recommendations, or the the top 10%?

Yes, that would be trivially easy, but that's not what was being asked for here.

To my reading, the poster was looking for a system that would provide the most desirable items for a given user, based on that user's past ratings of other items.

So you need to model what that person likes and figure out how that correlates to what other people like.

If I like cheese and tomatoes, and you like cheese, tomatoes, and spinach, then presumably I might like spinach too. The system's confidence that I would like spinach grows in line with the following:

• The number of people who have also expressed an affinity for cheese, tomatoes, and spinach
• The number of people who have expressed that affinity but have not expressed a strong dislike for something else I do like
• The strength of my indicated affinity for cheese and tomatoes (as expressed through multiple ratings or a particular strong rating)
• My trustworthiness and predictability as a rater
• Your trustworthiness and predictability as a rater
• Alignment between your and my rating bias (do both of us tend to rate everything fairly high or low)
• ... and so on ...

Not an easy problem at all. Especially when you want to optimize it for rapid matching.

Re:Hire someone

[raju1kabir]

It can't be that hard. I am sure that I could design a basic scheme pretty fast. Backend code would probably take me 4 to 8 hrs at most using PHP and MySQL...

I don't know about that...

Last time I got involved in a project like this I was knee-deep in stats books before I knew it. Brought in some statisticians who were scratching their heads tuning formulas more than a page long. Finally the funding came through for something else and it all got shelved.

In order to have a workable system, you have to account for all sorts of scoring biases, results significance that changes with sample size, variations in attribute preference, and of course troublemakers who want to fuck with the ratings for one reason or another.