You'd be right about using Pearson, but a difference metric is only one part of the algorithm. That's whats used to take a vector of [{storyid, rating}, {storyid, rating}] of one user compared to another. At that point a nearest neighbor algorithm is used to find the users most like you (through the Pearson Correlation). If you're really interested in machine learning and recommender algorithms, I'd encourage you to take a gander at the Wikipedia page on Collaborative Filtering, it's the basis of many of the ideas used in such systems.

I'm writing a blog post about this project, from development to institution. My prototype was in Python but the production system leverages already written code from the Apache Mahout project. I've used Mahout at work which made it a pretty small lift to get this project from idea to product in ~2 nights.

At some point I plan to publish all my Data Mining & Machine Learning notes I took from my graduate program. They're where much of my inspiration for these kinds of projects come from and I describe the process by which many common algorithms work. That being said, I managed to fall rather ill during that course so the projects near the end are of poor quality.

If you're at all interested in algorithms & programming, I hope you'll join irc.darklordpotter.net #programming -- it's my hope to build a strong development community and hopefully work on some projects as a community. The new PatronusCharm will be an open source effort.

 

Recommendations now include all WBA forums, unincludes a hidden forum called the "Trash Bin", a source of much error in the algorithm.

Refreshes now occur every 5 minutes.

 

Huh. I don't know which, but some of the changes really did the trick. Nonjon is entirely gone from that list. Now it's starting to look interesting.

In fact, it recommends me my own story. Considering that I write what I'd like to read, you have a winner, Raven.

 

I've been working on a feedback system, e.g. a "yes this was a good rec" vs. "no this was awful" which should solve this issue.

 

Two updates:

The recommender refreshes every minute now.

Also:

There's an alternative (LogNormal-based) recommender https://forums.darklordpotter.net/recommended.php?alg=new I'd love some feedback on it.

 

Isn't that list completely different for you, then?

I get a totally different bunch of recommendations than before. The funny thing is, though, that they, too, fit. There's a lot of WbA threads in there I for some reason or another never rated but read and enjoyed. So I can't say yet which is better, but neither is off-base.

 

Hehe. First recc:

Tagging in For Review: Browsing the Library by pairings
2 ratings. 30 posts. Last post on January 17, 2012 at 04:06 pm